Telemedicine adoption in cardiology: Determinants and predictors identified using Bayesian Model Averaging and Machine Learning.

Telemedicine adoption in cardiology: Determinants and predictors identified using Bayesian Model Averaging and Machine Learning - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice PLOS Digit Health . 2026 Apr 20;5(4):e0001359. doi: 10.1371/journal.pdig.0001359 Search in PMC Search in PubMed View in NLM Catalog Add to search Telemedicine adoption in cardiology: Determinants and predictors identified using Bayesian Model Averaging and Machine Learning Pascal Petit Pascal Petit 1 Institute of Engineering, Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, SANGRIA, Grenoble, France Conceptualization, Data curation, Formal analysis, Validation, Visualization, Methodology, Writing – original draft, Writing – review & editing Find articles by Pascal Petit 1, ☯ , Jonathan Nübel Jonathan Nübel 2 Department of Cardiology, University Hospital Heart Centre Brandenburg, Brandenburg Medical School Theodor Fontane, Bernau/Neuruppin, Germany 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany Conceptualization, Investigation, Methodology, Writing – original draft, Writing – review & editing Find articles by Jonathan Nübel 2, 3, ☯ , Marie Josephine Walter Marie Josephine Walter 2 Department of Cardiology, University Hospital Heart Centre Brandenburg, Brandenburg Medical School Theodor Fontane, Bernau/Neuruppin, Germany 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany Investigation, Writing – original draft, Writing – review & editing Find articles by Marie Josephine Walter 2, 3 , Christian Butter Christian Butter 2 Department of Cardiology, University Hospital Heart Centre Brandenburg, Brandenburg Medical School Theodor Fontane, Bernau/Neuruppin, Germany 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany Writing – original draft, Writing – review & editing Find articles by Christian Butter 2, 3 , Martin Heinze Martin Heinze 4 Department for Psychiatry and Psychotherapy, Center for Mental Health, Immanuel Hospital Rüdersdorf, Brandenburg Medical School Theodor Fontane, Rüdersdorf, Germany Writing – original draft, Writing – review & editing Find articles by Martin Heinze 4 , Yuriy Ignatyev Yuriy Ignatyev 4 Department for Psychiatry and Psychotherapy, Center for Mental Health, Immanuel Hospital Rüdersdorf, Brandenburg Medical School Theodor Fontane, Rüdersdorf, Germany Writing – original draft, Writing – review & editing Find articles by Yuriy Ignatyev 4 , Anja Haase-Fielitz Anja Haase-Fielitz 2 Department of Cardiology, University Hospital Heart Centre Brandenburg, Brandenburg Medical School Theodor Fontane, Bernau/Neuruppin, Germany 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany 5 Institute of Social Medicine and Health System Research, Otto von Guericke University Magdeburg, Magdeburg, Germany Conceptualization, Investigation Find articles by Anja Haase-Fielitz 2, 3, 5 , Nicolas Vuillerme Nicolas Vuillerme 1 Institute of Engineering, Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, SANGRIA, Grenoble, France 6 Institut Universitaire de France, Paris, France Conceptualization, Funding acquisition, Methodology, Supervision, Writing – original draft, Writing – review & editing Find articles by Nicolas Vuillerme 1, 6, ‡ , Felix Muehlensiepen Felix Muehlensiepen 1 Institute of Engineering, Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, SANGRIA, Grenoble, France 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany 7 Center for Health Services Research, Faculty for Health Sciences, Brandenburg Medical School Theodor Fontane, Rüdersdorf, Germany Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing Find articles by Felix Muehlensiepen 1, 3, 7, *, ‡ Editor: Mathew V Kiang 8 Author information Article notes Copyright and License information 1 Institute of Engineering, Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, SANGRIA, Grenoble, France 2 Department of Cardiology, University Hospital Heart Centre Brandenburg, Brandenburg Medical School Theodor Fontane, Bernau/Neuruppin, Germany 3 Faculty of Health Sciences, Joint Faculty of the University of Potsdam, the Brandenburg Medical School Theodor Fontane, and the Brandenburg Technical University Cottbus-Senftenberg, Cottbus, Germany 4 Department for Psychiatry and Psychotherapy, Center for Mental Health, Immanuel Hospital Rüdersdorf, Brandenburg Medical School Theodor Fontane, Rüdersdorf, Germany 5 Institute of Social Medicine and Health System Research, Otto von Guericke University Magdeburg, Magdeburg, Germany 6 Institut Universitaire de France, Paris, France 7 Center for Health Services Research, Faculty for Health Sciences, Brandenburg Medical School Theodor Fontane, Rüdersdorf, Germany 8 Stanford University School of Medicine, UNITED STATES OF AMERICA The authors have declared that no competing interests exist. ☯ ☯ These authors contributed equally and share the first authorship on this work. ‡ NV and FM also contributed equally and share the last authorship on this work. ✉ * E-mail: [email protected] Roles Pascal Petit : Conceptualization, Data curation, Formal analysis, Validation, Visualization, Methodology, Writing – original draft, Writing – review & editing Jonathan Nübel : Conceptualization, Investigation, Methodology, Writing – original draft, Writing – review & editing Marie Josephine Walter : Investigation, Writing – original draft, Writing – review & editing Christian Butter : Writing – original draft, Writing – review & editing Martin Heinze : Writing – original draft, Writing – review & editing Yuriy Ignatyev : Writing – original draft, Writing – review & editing Anja Haase-Fielitz : Conceptualization, Investigation Nicolas Vuillerme : Conceptualization, Funding acquisition, Methodology, Supervision, Writing – original draft, Writing – review & editing Felix Muehlensiepen : Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing Mathew V Kiang : Editor Received 2025 Jul 10; Accepted 2026 Mar 29; Collection date 2026 Apr. © 2026 Petit et al This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice PMCID: PMC13095100  PMID: 42008527 Abstract In this secondary analysis of a German cross-sectional survey data, we investigated key determinants and predictors of telemedicine (TM) use among healthcare professionals (HCPs) treating cardiology patients. We applied Bayesian Model Averaging (BMA) for explanatory analysis and Machine Learning (ML) for predictive modeling. BMA identified TM determinants after excluding collinear variables and selecting variables based on LASSO regression. The extreme gradient boosting (XGBoost) ML algorithm predicted TM use and identified key predictors, using nested cross-validation to prevent overfitting. ML model performance was assessed via area under the receiver operating characteristic curve (AUROC), while predictor importance was evaluated using Shapley additive explanations. Among 112 HCPs, 64 (57%) used TM. BMA identified 12 determinants, including positive associations with TM knowledge, being a cardiologist, female gender, and perceiving TM as suitable for heart failure and for monitoring events. Negative associations included concerns about insufficient patient benefits, perceptions that TM is less suitable for acute events, and skepticism regarding its relevance for extending aftercare intervals. The XGBoost model showed strong predictive performance (AUROC: 0.88 [95% CI: 0.75; 1.00], accuracy: 0.79) for TM use. Key promoting factors included TM knowledge, being a cardiologist, female gender, number of average patients per quarter, and perceiving TM as suitable for arrhythmias, device follow-up, and heart failure. Limiting factors included older age, personal use of TM for one’s own health, and skepticism about TM’s relevance in acute situations. These findings emphasize the importance of knowledge and attitudes in shaping TM adoption and show that ML can accurately identify healthcare professionals most likely to use TM, supporting targeted interventions and safer implementation in cardiology. Author summary In our study, we wanted to better understand why some healthcare professionals in Germany use telemedicine for cardiology patients, while others do not. We surveyed doctors about their experiences and opinions on using digital tools in heart care. To understand better what drives or hindered participants we used two types of analysis: one that finds important influencing factors, and one that predicts which professionals are most likely to use telemedicine. We found that healthcare professionals who had strong knowledge of telemedicine and who perceived it as relevant for patient care were much more likely to adopt it. Being a cardiologist and self-identifying as female were also consistently associated with higher use. In contrast, those who expressed doubts about its patient benefits, considered TM unsuitable for acute events, or reported insufficient knowledge were less likely to engage with it. Our predictive model achieved high accuracy in identifying those most likely to adopt telemedicine. While these results may appear intuitive, our use of advanced statistical and machine learning methods provides robust, data-driven evidence for patterns that often seem self-evident in practice. By quantifying these associations, our work offers a stronger foundation for healthcare policy, such as promoting education on telemedicine and addressing barriers for those who feel less confident using digital tools. Introduction The digital transformation of healthcare is advancing rapidly, with cardiology historically serving as a key driver of medical innovation [ 1 ]. From the early adoption of the electrocardiogram (ECG) as one of the first digital diagnostic tools [ 2 ] to the widespread implementation of remote monitoring [ 3 ] and AI-assisted diagnostics [ 4 ], cardiology has consistently embraced and shaped technological advancements to improve patient care [ 5 ]. With an aging population and a rising prevalence of cardiovascular disease [ 6 ], the demand for cardiology services is increasing even as the shortage of healthcare professionals intensifies [ 7 ]. Digital health technologies, particularly telemedicine (TM), offer a promising solution to address this growing healthcare challenge [ 8 – 13 ]. TM encompasses a wide range of applications, from remote consultations to continuous telemonitoring, offering diverse solutions for patient care and clinical decision-making. The American Heart Association has underscored these advantages, yet significant challenges, such as insufficient technological infrastructure, reimbursement issues, and limited digital literacy, continue to impede its widespread adoption [ 14 ]. This issue is particularly relevant in rural regions [ 15 ], such as Brandenburg in Germany, where long travel distances to specialized care create barriers to timely treatment [ 16 ]. Despite the technological potential of TM, a critical question remains: who is both willing and able to integrate TM effectively into clinical practice? Previous studies suggest that healthcare professionals (HCPs) act as key gatekeepers in the adoption of digital health interventions, with their acceptance or reluctance ultimately determining the extent to which TM is implemented in real-world healthcare settings [ 17 ]. Historically, theoretical models such as Rogers’ Diffusion of Innovation Theory, the Technology Acceptance Model (TAM), and the Unified Theory of Acceptance and Use of Technology (UTAUT-II) have been employed to explain the adoption of digital health interventions [ 18 ]. While these frameworks offer valuable insights, they remain largely descriptive and are limited in their ability to precisely predict TM utilization. With the rise of artificial intelligence (AI), new opportunities are emerging to systematically identify key predictors of TM adoption. Predictive modeling enables a precise analysis of the factors influencing HCPs’ willingness to integrate TM, helping to tailor technologies to clinical needs. Understanding adoption barriers is crucial for developing inclusive cardiology care models and optimizing healthcare structures while mitigating risks associated with suboptimal TM implementation. Different statistical analysis approaches have been used in the past few decades in order to understand the factors associated with TM use. Most previous investigations of TM adoption among HCPs in cardiology-related fields have employed either qualitative methods [ 19 – 21 ], descriptive statistics [ 22 , 23 ] clustering techniques [ 24 , 25 ], or logistic regression models [ 26 – 28 ]. While these approaches have provided valuable insights, they also carry limitations, particularly in terms of addressing model uncertainty, achieving robust predictive performance, and capturing complex nonlinear relationships between predictors and outcomes. In this study, we aimed to address these methodological gaps by applying Bayesian model averaging (BMA) alongside modern explainable machine learning (ML) techniques. BMA is designed to identify determinants that are statistically associated with an outcome (e.g., TM use) and provide insights into potential mechanisms and drivers of behavior. In contrast, ML methods such as extreme gradient boosting (XGBoost), prioritize predictive accuracy and are able to capture complex, non-linear relationships that might remain hidden in conventional regression-based approaches. Unlike conventional regression analyses, which generally identify a single “best-fitting” model and disregard the uncertainty surrounding model specification, BMA formally accounts for this uncertainty by averaging across multiple plausible models weighted by their posterior probabilities [ 29 ]. This approach yields parameter estimates that are more robust, less sensitive to sampling variability, and therefore more reliable in contexts such as TM adoption, where multiple interdependent factors are at play [ 17 ]. In addition, we applied a machine learning approach using XGBoost, which is widely recognized for its strong predictive accuracy, computational efficiency, and capacity to capture nonlinear and high-dimensional dependencies [ 30 – 32 ]. By combining BMA and ML in the present study, we sought to distinguish between explanatory relevance (factors associated with TM use) and predictive relevance (factors that improve classification of TM users vs. non-users). This dual perspective not only provides complementary insights into the determinants of TM adoption among HCPs but also allows us to assess the robustness of findings across distinct analytical frameworks. Methods Overview This study is a secondary analysis of previously published data from a web-based survey assessing knowledge, acceptance, and utilization of TM among cardiologists, internists, and general practitioners [ 33 ]. The pre-validated 34-item questionnaire covered TM-related topics, including perceived benefits, barriers, and clinical applications. The full list of questionnaire items is provided in Supplementary S1 Table . The primary study identified physician subgroups based on TM attitudes and usage patterns. Studied population Participants were eligible if they met the following inclusion criteria: 1) working in inpatient or outpatient cardiology care, including cardiologists, internists, and general practitioners, 2) practicing in Brandenburg or Berlin, and 3) providing informed consent. Physicians who did not meet these criteria were excluded, along with questionnaires in which less than 50% of the items were completed. Recruitment was conducted through professional email invitations, a QR code published in a regional health journal, and additional outreach via platform X. Data collection took place from May 28, 2021, to February 28, 2022, using the web-based survey tool LimeSurvey [ 34 ]. Statistical analysis All statistical analyses were performed using R software 4.5.0® (R Core Team, Vienna, Austria) for Windows 11©. The identification of TM use determinants followed the same methodology as in a previous work [ 33 ]. Data management Independent variables were defined based on all columns except the outcome variable TM use. Ordered variables were retained as ordered factors, and nominal categorical variables with more than two levels were one-hot encoded, with one reference category omitted per variable to prevent multicollinearity [ 24 ]. Missing values were handled as a separate category, under the assumption that nonresponse may reflect systematic differences in TM use. After these steps, the dataset included 53 independent covariates. Bayesian model averaging (BMA) analysis We implemented a structured three-step pipeline to identify determinants of TM use. First, collinearity screening was conducted by examining pairwise correlations and calculating the variance inflation factor (VIF), with variables exhibiting VIF > 2.5 excluded to mitigate multicollinearity [ 24 ]. Next, LASSO regression was applied to reduce dimensionality and retain the most informative variables. Optimal penalization parameters were identified using five-fold cross-validation (CV), and variables with non-zero coefficients at the optimal lambda were retained. These variables were subsequently entered into a BMA framework with logistic regression. BMA averages across all candidate models, weighting by posterior probabilities, thus accounting for model uncertainty (Depaoli et al. 2021). This process provided posterior inclusion probabilities (PIPs) for each feature, indicating the likelihood that the variable has a non-zero effect on TM use. Posterior mean coefficients and standard deviations were calculated across all models (full BMA estimates), and we also summarized results from the single highest posterior probability model and the five highest posterior probability models for comparison. Variables with non-zero PIPs were considered determinants of TM use, with higher PIPs reflecting stronger evidence of association. Visual summaries of group-level differences were produced using radar plots, where continuous and ordinal variables were normalized to a 0–1 range using min–max scaling and summarized by group medians, whereas dichotomous variables were summarized by group proportions. BMA was performed using the BMA package (version 3.18.17) [ 27 ]. Regarding priors for BMA, we assumed that all candidate models were equally likely a priori (same prior weight). Machine learning analysis To complement the BMA analysis and leverage the predictive capacity of ML, we applied a binary XGBoost model. Unlike regression models, XGBoost handles multicollinearity through its tree-based structure, so no pre-filtering for collinearity was performed [ 30 – 32 ] To avoid biased evaluation, a nested CV framework was applied. The dataset was first split into a 70% training set and a 30% holdout test set using stratified sampling to preserve the proportion of TM users and non-users. Within the training set, a 10-fold outer CV was performed, with each fold sequentially serving as validation while the remaining folds formed the training subset. Within each outer training fold, a 10-fold inner CV was conducted to tune hyperparameters. Due to computational constraints, 30 random hyperparameter combinations were tested for each fold, with learning rate ([0.05-1]), maximum depth ([3–7]), minimum sum of instance weights (hessian) needed in a child node ([1–10]), subsample ([0.5-1]), and subsample ratio of columns when constructing each tree ([0.5-1]) varied within pre-specified ranges. The scale_pos_weight parameter was set to the ratio of majority to minority class to account for class imbalance. Hyperparameters were optimized by maximizing the mean area under the receiver operating characteristic curve (AUROC) across inner folds. The best hyperparameter configuration for each outer fold was used to train an XGBoost model, which was then evaluated on the corresponding outer validation set. Following nested CV, the hyperparameter configuration yielding the highest AUROC across outer folds was selected to train a final XGBoost model on the full training set. This model was evaluated on the independent 30% holdout test set. Evaluation metrics included AUROC with 95% confidence intervals, accuracy with 95% confidence intervals, F1-score, and additional metrics such as precision, recall, and specificity. The F1 score is a measure of predictive performance, combining precision and recall into a single value, ranging from 0 (poor performance) to 1 (perfect performance). The XGBoost algorithm was computed using the xgboost package version 1.7.8.1 [ 35 ]. The caret package version 7.0-1 was used to calculate the confusion matrix [ 36 ]. The pROC package version 1.18.5 [ 37 ] was used to calculate the AUROC, while the MLmetrics package version 1.1.3 [ 38 ] was used to compute evaluation performance metrics not provided by the caret package. The stratified sampling was performed with the splitTools package version 1.0.1 [ 39 ]. Model training, cross-validation, and hyperparameter tuning were carried out with the mlexperiments package [ 40 ] in combination with the mllrnrs package [ 41 ], which provide standardized machine learning pipelines and learner interfaces. To enhance the interpretability of the XGBoost model, SHapley Additive exPlanations (SHAP) values were calculated using the holdout test set to quantify the contribution of each predictor to the model’s output. SHAP values reflect the individual contribution of each predictor by evaluating its marginal impact on the model’s output [ 42 , 43 ].Predictors that directed the model toward TM use were considered promoting factors, whereas predictors directing the model toward non-use were considered limiting factors. Directionality was inferred using a hybrid heuristic. For binary and sparse features, mean SHAP values were compared across groups, while for continuous features, generalized additive models (GAMs) were fitted to SHAP values, and the average derivative of the fitted curve was used to determine association direction. When GAMs failed to converge, Spearman’s rank correlation between the feature and its SHAP values served as a fallback. Sensitivity analysis A sensitivity analysis was performed by excluding individuals with missing values to evaluate the robustness of our findings. Ethics statement The study was approved by the local ethics committee of the Brandenburg Medical School (E-01–20210304). Data processing was based on the informed consent of the participants in the study. Participation in the study was not remunerated. Personal data were only collected from the participants to be able to process any requests in accordance with current law. These data were deleted after the end of the study. No other personal data were collected [ 44 ]. Results Healthcare professional characteristics A total of 112 healthcare professionals (HCPs) were considered, including 64 (57%) TM users. Compared to our previous work, we included all 112 HCPs to prevent the introduction of a selection bias. There were 54 males (48%), 28 females (25%), 2 non-binaries (1.8%), and 28 persons (25%) who did not declare their gender. Most HCPs were 50–59 years (29%), followed by HCPs aged 40–49 (21%), 30–39 (14%), ≥ 60 (9%), and 20–29 (4%). There were 63 internal medicine specialists (56%), 47 cardiologists (42%), and 12 general practitioners (11%). On average, HCPs had 800 patients per quarter. For sample characteristics we refer to the recently published study [ 19 ]. BMA analysis Of the 53 independent covariates initially considered, 32 (60%) were retained after excluding collinear variables. Among these 32, 12 (38%) were selected using Least Absolute Shrinkage and Selection Operator (LASSO) regression and subsequently included in the BMA analysis. Examination of posterior model probabilities revealed model uncertainty. No single model dominated, with the highest posterior model accounting for only 11.0% of the total weight, and the top five models together representing 41.3%. Estimates based on only the top model, or approximated across the top five models, differed from the fully model-averaged estimates ( Table 1 and Fig 1 ). In contrast, posterior means from the full BMA, which incorporate information from the entire model space, provided more stable estimates that were often shrunk toward zero. Table 1. Factors identified through BMA and their association with TM use. Determinants Top model Top 5 models BMA Posterior inclusion probability (%) Direction mean mean mean sd TM knowledge 1.45 1.00 1.45 0.36 100 Positive association TM barrier: no remuneration possible 0 0 0.13 0.48 8.80 Positive association TM barrier: data security 0 0 0.02 0.21 2.20 Positive association TM barrier: insufficient data for the benefit of patients 0 0.13 -0.27 0.72 16.5 Negative association TM use suitable for heart failure 0 0 0.17 0.46 14.1 Positive association TM use suitable for monitoring an event 0 0.55 0.53 0.71 43.0 Positive association TM use relevant for basic monitoring of ECG 0.23 0.27 0.23 0.50 21.4 Positive association TM use relevant for extension of the aftercare interval 0 0 0.03 0.17 3.80 Positive association TM use less suitable in acute events 0 0.18 -0.13 0.41 11.8 Negative association TM use less suitable for extension of the aftercare interval 0 0 -0.004 0.07 0.90 Negative association Self-identified as female 1.42 1.00 1.42 0.93 80.6 Positive association Being a cardiologist 1.26 1.00 1.26 0.77 83.5 Positive association Open in a new tab Fig 1. Coefficient estimates for each predictor variable under three modeling perspectives: the highest posterior probability single model (“Top model”), an approximation averaged over the top five models (“Top 5 models”), and the fully Bayesian model-averaged estimates (“BMA”). Open in a new tab The points represent the estimated coefficient effect sizes, while the error bars show the posterior standard deviations derived from the full BMA analysis, reflecting uncertainty in these averaged estimates. The “Top model” and “Top 5 models” estimates do not include uncertainty intervals because these approximations do not capture model-averaged uncertainty. The full “BMA” estimates incorporate all plausible models weighted by their posterior probabilities, shrinking coefficients toward zero when evidence for inclusion is weak and properly reflecting parameter uncertainty. A total of 12 determinants of TM use were identified ( Table 1 and Fig 2 ). Positive associations were found for TM knowledge, being a cardiologist, self-identifying as female, perceiving TM as suitable for heart failure and for monitoring an event, considering absence of remuneration and data security as barriers to TM, and perceiving TM as relevant for extending the aftercare interval and for basic ECG monitoring. Negative associations were observed for considering insufficient data on patient benefit as a TM barrier, and for perceiving TM as less suitable in acute events or for extending the aftercare interval ( Table 1 ). Fig 2. Profile of healthcare professionals (HCPs) in cardiology using telemedicine (TM) versus non-users. Open in a new tab The radar plot compares key determinants, with TM users (green) and non-users (red). The sensitivity analysis produced similar findings, identifying eight determinants, all of which overlapped with the main analysis and showed consistent directions of association ( Figs 2 and S1 and S1 Table ). However, in this sensitivity analysis, data security, the relevance of TM for ECG monitoring and aftercare extension, and the unsuitability of TM in acute events were not retained as determinants. The determinants identified through BMA were used to describe the profiles of HCPs in cardiology who use and do not use TM ( Fig 2 ). Profiles are presented for both the main and sensitivity analyses. The percentages correspond to the proportion of HCPs endorsing each specified response. For example, 100% (the outer line of the radar chart) indicates that all HCPs selected the given response (e.g., self-identifying as female), whereas 0% (the inner line) indicates that none selected it. Machine learning analysis The XGBoost model demonstrated strong performance, with an AUROC of 0.88 (95% CI: 0.75–1.00), an accuracy of 0.79 (95% CI: 0.62–0.91), a sensitivity of 79%, a specificity of 80%, and an F1 score of 0.81 ( Table 2 ). Table 2. Performance of the final XGBoost model. Performance metric Main analysis Sensitivity analysis Accuracy (95% CI) 0.79 (0.62; 0.91) 0.84 (0.68; 1.00) AUC (95% CI) 0.88 (0.75; 1.00) 0.76 (0.55; 0.90) Kappa 0.59 0.50 No Information Rate (NIR) 0.56 0.60 P value [accuracy > NIR] 0.004 0.07 Mcnemar’s test p value 1.00 1.00 Sensitivity/ Recall 0.79 0.80 Specificity 0.80 0.70 Precision/ positive predictive value 0.83 0.80 Negative predictive value 0.75 0.70 F1 score 0.81 0.80 Detection rate 0.44 0.48 Balanced accuracy 0.79 0.75 Open in a new tab Note: The purpose of the McNemar’s test is to test whether the number of false positives and false negatives are statistically equal. Several of the most influential predictors identified by SHAP analysis (e.g., TM knowledge, being a cardiologist, and female sex) were also among the key determinants highlighted in the BMA analysis ( Fig 3 and Table 3 ). The ML framework identified additional predictors that did not emerge as determinants in the BMA models. For example, perceptions of TM suitability for arrhythmias, pacemaker and defibrillator therapy, and for monitoring blood pressure ranked highly in predictive importance, reflecting the model’s ability to capture more complex or non-linear associations. Conversely, factors such as concerns about insufficient patient benefit, which were negatively associated with TM use in BMA, did not contribute substantially to the predictive performance of the ML model. Fig 3. Feature importance according to SHAP values for the prediction of TM use. Open in a new tab Table 3. Top predictors of TM Use identified with XGBoost. Feature Mean |SHAP| (95%CI) Direction TM knowledge 0.99 (0.27; 2) Promoting Being a cardiologist 0.33 (0.11; 0.75) Promoting TM use less suitable in acute events 0.33 (0.17; 0.69) Limiting Self-identified as female 0.32 (0.13; 0.92) Promoting TM use suitable for cardiac arrhythmias 0.27 (0.09; 0.44) Promoting TM use suitable for treatment with pacemaker, defibrillator, cardiac resynchronization systems 0.26 (0.07; 0.70) Promoting TM use less suitable for extension of the aftercare interval 0.23 (0.09; 0.81) Limiting Number of average patients per quarter 0.22 (0.02; 0.74) Promoting TM use relevant in acute events 0.22 (0.08; 0.35) Limiting TM use relevant for handling/complications of drugs/medical devices 0.21 (0.07; 0.59) Promoting TM barrier: current remuneration structure insufficient 0.19 (0.02; 0.78) Promoting Assessment of relevant TM training among colleagues 0.18 (3.5e-03; 0.46) Promoting Age 0.15 (0.06; 0.33) Limiting TM use relevant for basic monitoring of blood pressure 0.14 (0.05; 0.25) Promoting TM barrier: acquisition of technology 0.13 (0.04; 0.43) Limiting Assessment of the willingness of colleagues to train on TM topic 0.12 (0.03; 0.32) Promoting TM use suitable for heart failure 0.12 (0.04; 0.32) Promoting Being an internal medicine specialist 0.10 (0.05; 0.23) Promoting TM use for own health 0.08 (6.2e-03; 0.24) Limiting TM use suitable for monitoring an event 0.06 (0.02; 0.13) Limiting Willingness to invest for TM 0.06 (0.02; 0.12) Limiting Open in a new tab In terms of directionality, SHAP values indicated that several factors promoted TM use, including TM knowledge, being a cardiologist, female sex, perceiving TM as suitable for arrhythmias and heart failure, higher patient volume per quarter, and viewing TM as relevant for managing complications of drugs and medical devices. In contrast, limiting factors included perceiving TM as unsuitable for acute events or aftercare extension, older age, using TM for own health, and concerns about implementation effort ( Table 3 and Fig 3 ). The sensitivity analysis yielded broadly consistent findings, although some predictors identified in the main model were not retained, such as perceptions of insufficient remuneration as a barrier or suitability for hypertension ( Tables 2 , S2 and S3 and S2 Fig ). At the same time, new predictors emerged, such as the perception that TM is less suitable for basic ECG monitoring or for consulting colleagues. Certain predictors shifted directionality. For instance, the number of patients per quarter and willingness to invest in TM showed opposite associations compared with the main model. These discrepancies emphasize that while the overall predictive structure of the ML model was stable, the importance of individual features may vary depending on the population considered. SHAP value (x-axis) for each HCP and feature are represented with a point. Positive SHAP values imply an impact to the model toward wanting TM use, while negative values impact the model toward not using TM. For each feature, the mean and 95% CI of the absolute SHAP values are reported on the right of the graph. High SHAP value (in absolute value) indicate a high impact on the model output. Red colors indicate that an HCP either answered yes to the considered question (y-axis) regarding categorical feature or tend toward high values for continuous features (e.g., age), while blue colors refer either to “no” answers for categorical features or low values for continuous features, and grey colors refer to missing answers. BMA and XGBoost comparison The explanatory and predictive frameworks converged on several key factors while also highlighting different aspects of TM adoption ( Table 4 ). Both BMA and XGBoost identified TM knowledge, female sex, and being a cardiologist as robust determinants of TM use. These variables demonstrated both a consistent statistical association and a strong contribution to predictive performance, suggesting that they represent core drivers of adoption in cardiology practice. Similarly, perceptions that TM is unsuitable for acute events or for extending aftercare intervals were negatively associated with use in BMA and also reduced predictive probability in XGBoost, reinforcing their importance as barriers. Table 4. Comparison of the feature impact on TM use depending on the statistical method use. Domain Feature BMA (association) XGBoost (prediction) Personal TM knowledge promoting promoting Personal Self-identified as female promoting promoting Personal Age limiting Personal TM use for own health limiting Personal Willingness to invest for TM limiting Occupational Being a cardiologist promoting promoting Occupational Being an internal medicine specialist promoting Occupational Number of average patients per quarter promoting Occupational Assessment of the willingness of colleagues to train on TM topic promoting Occupational Assessment of relevant TM training among colleagues promoting TM use unsuitability Acute events limiting limiting TM use unsuitability Extension of the aftercare interval limiting limiting TM use suitability Cardiac arrhythmias promoting TM use suitability Monitoring an event promoting limiting TM use suitability Heart failure promoting promoting TM use suitability Treatment with pacemaker, defibrillator, cardiac resynchronization systems promoting TM use relevance Acute events limiting TM use relevance Handling/complications of drugs/medical devices promoting TM use relevance Basic monitoring of blood pressure promoting TM use relevance Basic monitoring of ECG promoting TM use relevance Extension of the aftercare interval promoting TM barrier Acquisition of technology limiting TM barrier No remuneration possible promoting promoting TM barrier Data security promoting TM barrier Insufficient data for the benefit of patients limiting TM barrier Current remuneration structure insufficient promoting Open in a new tab Note: empty cells refer to features that were not selected in the models. Beyond these points of convergence, the two approaches diverged in meaningful ways. BMA, by design, emphasized parsimonious associations with stable posterior inclusion probabilities, highlighting concerns about insufficient patient benefit as a limiting factor, an effect that did not substantially improve the predictive capacity of the XGBoost model. Conversely, XGBoost identified perceptions of TM suitability for arrhythmias, device monitoring, and blood pressure management as strong predictors, even though these factors did not consistently emerge as explanatory determinants in BMA. These differences underscore the complementary nature of the two methods. BMA provides insights into determinants with stable associations, while ML captures a broader set of variables that enhance classification accuracy, including those with weaker or more context-dependent effects. Importantly, one feature (perceptions of TM suitability for event monitoring) exhibited incongruent behavior across approaches, showing a positive association in one framework (BMA) but a limiting effect in the other (XGBoost). Such discrepancies highlight the value of applying both explanatory and predictive methodologies, as they draw attention to variables whose influence on TM adoption may depend on context, measurement, or modeling assumptions. Discussion Main findings This study applied BMA and XGBoost to move beyond confirming general associations by estimating the relative importance of predictors, producing more robust effect estimates, and exploring potential non-linear relationships to provide a more nuanced picture of TM adoption among HCPs. The findings can be summarized as follows. First, across both frameworks, TM knowledge, female sex, and being a cardiologist consistently emerged as core drivers of TM use. Second, the lack of remuneration was identified as a main barrier for TM use in both models. Third, the two approaches also highlighted distinct aspects of TM adoption. BMA emphasized concerns about insufficient patient benefit as a limiting factor, whereas XGBoost identified perceptions of TM suitability for arrhythmias, device monitoring, and blood pressure management as strong predictors. This complementarity underscores the value of combining explanatory and predictive perspectives, explanatory models reveal stable associations that clarify underlying mechanisms, while predictive models capture additional and sometimes non-linear factors that improve classification. Comparison with previous work Most previous investigations of TM adoption among HCPs in cardiology-related fields have relied on qualitative methods [ 19 – 21 ], descriptive statistics [ 22 , 23 ], clustering techniques, or classical logistic regression models [ 26 – 28 ]. While these approaches provided valuable insights into general associations, they are limited in addressing model uncertainty, assessing the relative importance of predictors, and capturing complex non-linear relationships. More recently, advanced modelling approaches have been applied to cardiology-related telemedicine data. Koos et al. (2022) used hierarchical logistic regression with cross-validation to disentangle the contributions of patient-, clinician-, and visit-level characteristics to telemedicine use in a large academic cardiovascular center, showing that the individual clinician seen was by far the strongest predictor (AUC = 0.83) [ 45 ]. In contrast, Tisdale et al. (2024) analyzed more than 220,000 patients in the U.S. Veterans Health Administration using multilevel logistic regression and found that 40.5% of the variability in telehealth use was attributable to the patient level, 30.8% to the clinician level, and 7% to the facility level [ 46 ]. These studies highlight the hierarchical structure of TM adoption but modelled mainly linear effects and did not quantify the relative contribution of individual predictors or explore potential non-linear patterns. Although our survey included only HCPs and could therefore capture only clinician-centered domains, both models consistently identified TM knowledge as the most important factor for TM use. This aligns with prior evidence showing that clinicians’ self-reported technological proficiency, comfort with TM platforms, and positive attitudes toward TM are among the strongest predictors of both current and future TM use, independent of demographic or professional background [ 47 – 51 ]. To the best of our knowledge, no prior work has applied approaches capable of assessing the relative importance of predictors and non-linear relationships to provide a more nuanced understanding of TM adoption dynamics among healthcare professionals in cardiology. Nevertheless, our findings can be robustly situated within the existing literature, as they align with consistently reported determinants of TM use. Both models in our analysis consistently identified heart failure as a condition perceived to be particularly suitable for TM. This aligns with evidence from clinical trials such as TIM-HF II, which demonstrated the effectiveness of structured telemedical care in heart failure management [ 45 , 46 ]. It is plausible that our respondents considered TM most appropriate in scenarios they were already familiar with, reflecting existing clinical experience rather than openness to entirely new use cases. Interestingly our BMA framework highlighted concerns about insufficient patient benefit as a limiting factor, this likely reflects uncertainty or limited awareness rather than an actual lack of evidence. TM in cardiology has demonstrated significant clinical benefits. A systematic review by Mohammadzadeh et al. found that telecardiology interventions facilitate early diagnosis and treatment, ultimately reducing mortality in patients with cardiovascular disease. Additionally, TM lowers healthcare costs and improves patient quality of life and satisfaction [ 51 ]. Similarly, studies by Raes et al. and Farabi et al. confirmed that telemonitoring for patients with cardiac implantable electronic devices (CIEDs), as well as telemedicine integrated with usual care, improves clinical outcomes while reducing healthcare costs [ 52 , 53 ]. Robust evidence further shows that telemedicine enhances clinical outcomes, patient satisfaction, and healthcare utilization in cardiology, particularly in chronic conditions such as heart failure, hypertension, and arrhythmias [ 14 , 52 , 53 ]. One possible explanation for the mixed perceptions of telemedicine in acute care contexts is that clinicians may differentiate between specific types of acute scenarios. Telemedicine may be perceived as appropriate for certain situations, such as diagnostic consultation or initial triage, while being considered less suitable for scenarios requiring immediate physical intervention. As the questionnaire did not explicitly distinguish between different subtypes of acute cardiovascular situations, our findings reflect general perceptions rather than evaluations of specific clinical contexts. Future research should therefore investigate in greater detail how clinicians assess the suitability of telemedicine across different acute care scenarios. Taken together, these findings suggest that the negative association observed in our model may reflect knowledge gaps or misconceptions rather than a lack of supporting evidence, underscoring the need for targeted education and awareness initiatives to strengthen confidence in telemedical approaches among healthcare professionals. Both of our models identified lack of remuneration as the strongest barrier to TM use. This aligns with prior studies showing that unclear or insufficient reimbursement is a key obstacle to TM adoption among clinicians in Germany. Gehrmann et al. (2025) reported that more than half of cardiologists, internists, and general practitioners cited inadequate reimbursement as a major barrier [ 53 ] and similar findings were reported by Muehlensiepen et al. (2021) in Rheumatologist [ 54 ] and Brauns and Loos (2015) [ 55 ]. Despite recent reforms, reimbursement for telemedical services remains fragmented and limited [ 56 ] and physicians continue to call for remuneration models equivalent to in-person care [ 57 ]. Our findings thus reinforce the view that structural financial barriers remain a major bottleneck for sustainable TM implementation for German HCPs. Strengths and limitations A key strength of our study is the application of established sophisticated statistical methodologies to real-world survey data collected from actual healthcare professionals. Those real-world data allow us to identify insights grounded in actual clinical perspectives of HCP. Given the complexity of the data and the many potential predictor variables, uncertainty about the best predictive model is inevitable. Taken together, the combination of BMA and ML enhances both the inferential rigor and predictive capacity of our analysis. Whereas prior studies have largely confirmed general associations between HCP characteristics and TM adoption, our study adds value by providing stronger evidence regarding the relative importance of predictors, generating more reliable effect estimates, and uncovering nonlinear relationships that yield a more nuanced understanding of adoption dynamics. By moving beyond descriptive or exploratory approaches, our work contributes a more robust and actionable analytical framework for advancing clinical and policy strategies in TM. In our BMA analysis, the five models with the highest posterior probabilities together accounted for only 41.3% of the total model weight, indicating considerable model uncertainty. Many variables displayed larger effect sizes in the “top model” or across the “top five models” compared with the full BMA estimates, illustrating how reliance on a small subset of models can exaggerate associations by neglecting the broader model space. Because the posterior probabilities of the leading models were relatively low, coefficients derived from them, essentially maximum likelihood estimates conditional on a single specification, risk overstating evidence. In contrast, the BMA posterior means integrate information across the entire space of plausible models, shrinking coefficients toward zero when support for a predictor is weak and thus yielding more stable and reliable estimates. These findings underscore that restricting inference to a single “best” model, or even to a handful of high-probability models, would disregard most of the available information. By averaging over the complete model space, BMA provides more robust inference and guards against spurious conclusions driven by arbitrary model selection. We further explore variable relevance by additionally applying an explainable ML approach, combining XGBoost, a powerful ML algorithm widely applied in predictive modeling tasks [ 58 , 59 ] with SHAP to enhance interpretability. The use of nested cross-validation represents a methodological strength, as it separates hyperparameter tuning from model evaluation and thereby reduces the risk of overfitting. Nevertheless, despite this safeguard, the relatively small sample size (n = 112) and the high number of initial variables (n = 53) could still allow the model to capture dataset-specific patterns, potentially inflating variable importance. The model’s strong predictive performance highlights the relevance of the identified factors and suggests that targeted interventions, such as structured training and administrative reforms, could effectively address adoption barriers and support broader TM integration in clinical practice. Although we did not validate our model using an independent dataset or multiple resampling iterations, the nested CV framework provides a reliable estimate of predictive performance and helps ensure that the identified predictors are supported within the training data. Our dataset includes a heterogeneous group of HCPs from diverse clinical settings, supporting the relevance of the findings. However, generalizability remains limited by the relatively small sample size, geographic restriction to two regions in Germany, and absence of external validation. Future studies should aim to replicate and extend this predictive framework using larger, independent datasets to assess its applicability across broader healthcare contexts. A key limitation of our study is the cross-sectional nature of the survey data, which limits causal interpretation. The aim of this study was not to establish predictive causality, but rather to explore and characterize distinct profiles of TM users. Some of the variables included as predictors, such as TM knowledge, perceptions of TM suitability for heart failure or arrhythmias, experience with cardiac devices, and considerations of remuneration, reflect current practices and attitudes rather than factors that necessarily preceded TM adoption. As such, the predictive models in this study estimate associations with existing TM use, rather than truly forecasting future adoption. A longitudinal design, in which TM use is measured prospectively, would allow for a more rigorous evaluation of causal relationships and the development of predictive models that identify determinants prior to adoption. Future research should validate our findings in independent and temporally structured datasets, ideally with repeated measures of TM use and predictors over time. Such studies would enable more accurate prediction of TM adoption and help distinguish factors that actively drive uptake from those that are concomitant with current usage. Because both BMA and XGBoost involve stochastic processes, the specific random seed used may influence the results. As a consequence, using a different seed could yield slightly different model estimates, variable importance rankings, or performance metrics, introducing potential variability in the findings. A potential limitation of our web-based survey is the likelihood of a positive selection bias, as physicians already interested in digitalization, telemedicine, or telecardiology may have been more inclined to participate. While recruitment included digital invitations, an additional outreach via the print magazine KV-intern (Kassenärztliche Vereinigung Brandenburg) aimed to reach all physicians registered with the Medical Association Brandenburg, mitigating but not eliminating this bias. Implications Our dual statistical perspective strengthens both the interpretability and practical utility of the findings, providing insights into why TM is adopted as well as who is most likely to adopt it. By revealing which factors most strongly shape TM adoption among HCPs, our models can help target interventions more effectively. Based on our findings structured training to increase TM knowledge should be prioritized. Notably, lack of remuneration emerged as a strong barrier in both models, underscoring the need to develop reimbursement frameworks that directly address this structural obstacle. This aligns with position statements from the American College of Cardiology, the European Heart Rhythm Association, and the American Heart Association, which emphasize that resolving reimbursement and regulatory issues is critical for sustainable TM adoption [ 3 ]. More broadly, our approach offers a scalable analytical framework for health system planners and policymakers to identify leverage points for promoting telemedicine uptake in other specialties and settings. Conclusion This study applied both Bayesian model averaging and machine learning to identify and predict key drivers of telemedicine use among healthcare professionals in cardiology. Across both analytical frameworks, telemedicine knowledge, being a cardiologist, and female gender emerged as consistent and robust predictors of use. While BMA emphasized stable statistical associations, such as concerns about insufficient patient benefit, machine learning captured additional non-linear and context-sensitive predictors, including perceptions of telemedicine suitability for specific cardiac conditions like arrhythmias and device monitoring. Importantly, both models confirmed that concerns around insufficient remuneration remain a major barrier to adoption. Together, these findings offer a more nuanced and data-driven understanding of telemedicine adoption dynamics, informing future implementation strategies. By highlighting both explanatory determinants and predictive indicators, this study supports the development of targeted interventions, such as structured training and reimbursement reform, to promote more equitable and sustainable integration of telemedicine into cardiology practice. Supporting information S1 Table. Full list of questionnaire items. Complete list of all 34 items included in the pre-validated web-based survey assessing knowledge, acceptance, and utilization of telemedicine among healthcare professionals in cardiology care. (DOCX) pdig.0001359.s001.docx (23.1KB, docx) S2 Table. BMA results from sensitivity analysis. Determinants of telemedicine use identified through Bayesian model averaging after exclusion of participants with missing values, including posterior inclusion probabilities and coefficient estimates. (DOCX) pdig.0001359.s002.docx (16.2KB, docx) S3 Table. XGBoost predictor importance from sensitivity analysis. Top predictors of telemedicine use identified by the XGBoost model in the sensitivity analysis, including mean absolute SHAP values and directionality. (DOCX) pdig.0001359.s003.docx (16.8KB, docx) S1 Fig. BMA radar plot from sensitivity analysis. Profile comparison of telemedicine users versus non-users based on determinants identified through Bayesian model averaging after exclusion of participants with missing values. (TIF) pdig.0001359.s004.tif (110.8KB, tif) S2 Fig. SHAP feature importance plot from sensitivity analysis. Feature importance according to SHAP values for the prediction of telemedicine use in the sensitivity analysis excluding participants with missing values. (TIF) pdig.0001359.s005.tif (53.8KB, tif) S1 Data. Anonymized Survey Data Set. De-identified individual-level response data from the web-based survey among healthcare professionals in cardiology care.(CSV) pdig.0001359.s006.csv (9.7KB, csv) Data Availability The anonymized survey dataset underlying the findings of this study, including the participant data and data dictionary, is provided in the Supporting information files. The codes used in the analysis of this study is publicly available at: https://doi.org/10.5281/zenodo.17047856 . Funding Statement The publication fee was covered by the Brandenburg Medical School publication fund supported by the Ministry of Science, Research and Cultural Affairs of the State of Brandenburg. The work of PP and NV was has been supported by funding from the French government, managed by the French National Research Agency (ANR) under the France 2030 program (MIAI Cluster) [ANR-23-IACL-0006] and under the “Investissements d’avenir” program [ANR-10-AIRT-0005; ANR-15-IDEX-0002]. References 1. Yeung AWK, Kulnik ST, Parvanov ED, Fassl A, Eibensteiner F, Völkl-Kernstock S, et al. Research on digital technology use in cardiology: bibliometric analysis. J Med Internet Res. 2022;24(5):e36086. doi: 10.2196/36086 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Fisch C. Centennial of the string galvanometer and the electrocardiogram. J Am Coll Cardiol. 2000;36(6):1737–45. doi: 10.1016/s0735-1097(00)00976-1 [ DOI ] [ PubMed ] [ Google Scholar ] 3. Varma N, Marrouche NF, Aguinaga L, Albert CM, Arbelo E, Choi JI, et al. HRS/EHRA/APHRS/LAHRS/ACC/AHA worldwide practice update for telehealth and arrhythmia monitoring during and after a pandemic. J Am Coll Cardiol. 2020;76(11):1363–74. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Nagarajan VD, Lee S-L, Robertus J-L, Nienaber CA, Trayanova NA, Ernst S. Artificial intelligence in the diagnosis and management of arrhythmias. Eur Heart J. 2021;42(38):3904–16. doi: 10.1093/eurheartj/ehab544 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Khera R, Oikonomou EK, Nadkarni GN, Morley JR, Wiens J, Butte AJ, et al. Transforming cardiovascular care with artificial intelligence: from discovery to practice. J Am Coll Cardiol. 2024;84(1):97–114. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Roth GA, Forouzanfar MH, Moran AE, Barber R, Nguyen G, Feigin VL. Demographic and epidemiologic drivers of global cardiovascular mortality. N Engl J Med. 2015;372(14):1333–41. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Joynt Maddox KE, Elkind MSV, Aparicio HJ, Commodore-Mensah Y, De Ferranti SD, Dowd WN. Forecasting the burden of cardiovascular disease and stroke in the United States through 2050—prevalence of risk factors and disease: a presidential advisory from the American Heart Association. Circulation [Internet]. 2024. [cited 2025 Sept 5];150(4). Available from: https://www.ahajournals.org/doi/10.1161/CIR.0000000000001256 [ DOI ] [ PubMed ] [ Google Scholar ] 8. Borges do Nascimento IJ, Abdulazeem H, Vasanthan LT, Martinez EZ, Zucoloto ML, Østengaard L, et al. Barriers and facilitators to utilizing digital health technologies by healthcare professionals. NPJ Digit Med. 2023;6(1):161. doi: 10.1038/s41746-023-00899-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Huerne K, Eisenberg MJ. Advancing telemedicine in cardiology: a comprehensive review of evolving practices and outcomes in a postpandemic context. Cardiovasc Digit Health J. 2024;5(2):96–110. doi: 10.1016/j.cvdhj.2024.02.001 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Ghilencea LN, Chiru MR, Stolcova M, Spiridon G, Manea LM, Stănescu AMA. Telemedicine: benefits for cardiovascular patients in the COVID-19 era. Front Cardiovasc Med. 2022;9:868635. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Kędzierski K, Radziejewska J, Sławuta A, Wawrzyńska M, Arkowski J. Telemedicine in cardiology: modern technologies to improve cardiovascular patients’ outcomes—A narrative review. Medicina. 2022;58(2):210. doi: 10.3390/medicina58020210 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Pegoraro V, Bidoli C, Dal Mas F, Bert F, Cobianchi L, Zantedeschi M. Cardiology in a digital age: opportunities and challenges for e-Health: a literature review. JCM. 2023;12(13):4278. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Iliuță L, Andronesi AG, Rac-Albu M, Furtunescu FL, Rac-Albu M-E, Scafa-Udriște A, et al. Challenges in caring for people with cardiovascular disease through and beyond the COVID-19 pandemic: the advantages of universal access to home telemonitoring. Healthcare (Basel). 2023;11(12):1727. doi: 10.3390/healthcare11121727 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Takahashi EA, Schwamm LH, Adeoye OM, Alabi O, Jahangir E, Misra S, et al. An overview of telehealth in the management of cardiovascular disease: a scientific statement from the American Heart Association. Circulation [Internet]. 2022. [cited 2025 Sept 5];146(25). Available from: https://www.ahajournals.org/doi/10.1161/CIR.0000000000001107 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Thompson SC, Nedkoff L, Katzenellenbogen J, Hussain MA, Sanfilippo F. Challenges in managing acute cardiovascular diseases and follow up care in rural areas: a narrative review. Int J Environ Res Public Health. 2019;16(24):5126. doi: 10.3390/ijerph16245126 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Loccoh EC, Joynt Maddox KE, Wang Y, Kazi DS, Yeh RW, Wadhera RK. Rural-urban disparities in outcomes of myocardial infarction, heart failure, and stroke in the United States. J Am Coll Cardiol. 2022. Jan;79(3):267–79. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Oudbier SJ, Souget-Ruff SP, Chen BSJ, Ziesemer KA, Meij HJ, Smets EMA. Implementation barriers and facilitators of remote monitoring, remote consultation and digital care platforms through the eyes of healthcare professionals: a review of reviews. BMJ Open. 2024;14(6):e075833. doi: 10.1136/bmjopen-2023-075833 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Harst L, Lantzsch H, Scheibe M. Theories predicting end-user acceptance of telemedicine use: systematic review. J Med Internet Res. 2019;21(5):e13117. doi: 10.2196/13117 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Alarabyat IA, Al-Nsair N, Alrimawi I, Al-Yateem N, Shudifat RM, Saifan AR. Perceived barriers to effective use of telehealth in managing the care of patients with cardiovascular diseases: a qualitative study exploring healthcare professionals’ views in Jordan. BMC Health Serv Res. 2023;23(1):452. doi: 10.1186/s12913-023-09368-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Maria ARJ, Serra H, Heleno B. Teleconsultations and their implications for health care: a qualitative study on patients’ and physicians’ perceptions. Int J Med Inform. 2022;162:104751. doi: 10.1016/j.ijmedinf.2022.104751 [ DOI ] [ PubMed ] [ Google Scholar ] 21. Cohen Rodrigues TR, de Buisonjé DR, Keesman M, Reijnders T, van der Geer JE, Janssen VR, et al. Facilitators of and barriers to lifestyle support and eHealth solutions: interview study among health care professionals working in cardiac care. J Med Internet Res. 2021;23(10):e25646. doi: 10.2196/25646 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Aamodt IT, Lycholip E, Celutkiene J, Strömberg A, Atar D, Falk RS, et al. Health care professionals’ perceptions of home telemonitoring in heart failure care: cross-sectional survey. J Med Internet Res. 2019;21(2):e10362. doi: 10.2196/10362 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Queiroz C, Guerreiro C, Oliveira-Santos M, Ferreira D, Fontes-Carvalho R, Ladeiras-Lopes R. Digital health and cardiovascular healthcare professionals in Portugal: current status, expectations and barriers to implementation. Rev Port Cardiol. 2024;43(8):459–67. doi: 10.1016/j.repc.2023.10.014 [ DOI ] [ PubMed ] [ Google Scholar ] 24. Muehlensiepen F, Petit P, Knitza J, Welcker M, Vuillerme N. Prediction of the acceptance of telemedicine among rheumatic patients: a machine learning-powered secondary analysis of German survey data. Rheumatol Int. 2024;44(3):523–34. doi: 10.1007/s00296-023-05518-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Gehrmann J, Hahn F, Stephan J, Steger A, Rattka M, Rudolph I, et al. Current use, challenges, barriers, and chances of telemedicine in the ambulatory sector in Germany—a survey study among practicing cardiologists, internists, and general practitioners. Telemed e-Health. 2025;31(6):779–92. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Brown KJ, Mathenge N, Crousillat D, Pagliaro J, Grady C, Katz N, et al. Social determinants of telemedicine utilization in ambulatory cardiovascular patients during the COVID-19 pandemic. Eur Heart J Digit Health. 2021;2(2):244–53. doi: 10.1093/ehjdh/ztab039 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Knörr V, Dini L, Gunkel S, Hoffmann J, Mause L, Ohnhäuser T, et al. Use of telemedicine in the outpatient sector during the COVID-19 pandemic: a cross-sectional survey of German physicians. BMC Prim Care. 2022;23(1):92. doi: 10.1186/s12875-022-01699-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Piskorz D, Díaz-Barreiro LA, López Santi R, Múnera A, Molina DI, Barroso WS, et al. Blood pressure telemonitoring and telemedicine for hypertension management-positions, expectations and feasibility of Latin-American practitioners. SURVEY carried out by several cardiology and hypertension societies of the Americas. Blood Press. 2022;31(1):236–44. doi: 10.1080/08037051.2022.2123781 [ DOI ] [ PubMed ] [ Google Scholar ] 29. Bahrami S, Hajian-Tilaki K, Bayani M, Chehrazi M, Mohamadi-Pirouz Z, Amoozadeh A. Bayesian model averaging for predicting factors associated with length of COVID-19 hospitalization. BMC Med Res Methodol. 2023;23(1):163. doi: 10.1186/s12874-023-01981-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet]. San Francisco, (CA): ACM; 2016. [cited 2025 Sept 5]. p. 785–94. Available from: https://dl.acm.org/doi/10.1145/2939672.2939785 [ Google Scholar ] 31. Grinsztajn L, Oyallon E, Varoquaux G. Why do tree-based models still outperform deep learning on tabular data? [Internet]. arXiv; 2022. [cited 2025 Sept 5]. Available from: https://arxiv.org/abs/2207.08815 [ Google Scholar ] 32. Moore A, Bell M. XGBoost, a novel explainable AI technique, in the prediction of myocardial infarction: a UK biobank cohort study. Clin Med Insights Cardiol. 2022;16:11795468221133611. doi: 10.1177/11795468221133611 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Muehlensiepen F, Petit P, Knitza J, Welcker M, Vuillerme N. Factors associated with telemedicine use among German general practitioners and rheumatologists: secondary analysis of data from a nationwide survey. J Med Internet Res. 2022;24(11):e40304. doi: 10.2196/40304 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. LimeSurvey [Internet]. LimeSurvey GmbH; 2024. Available from: https://www.limesurvey.org/ 35. Chen T, He T, Benesty M, Khotilovich V, Tang Y, Cho H, et al. xgboost: Extreme Gradient Boosting [Internet]; 2014. [cited 2025 Sept 5]. 1.7.11.1 p. Available from: https://CRAN.R-project.org/package=xgboost [ Google Scholar ] 36. Kuhn M. caret: Classification and Regression Training [Internet]; 2007. [cited 2025 Sept 5]. 7.0-1 p. Available from: https://CRAN.R-project.org/package=caret [ Google Scholar ] 37. Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez J-C, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinform. 2011;12:77. doi: 10.1186/1471-2105-12-77 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Yan Y. Machine Learning Evaluation Metrics [Internet]. 23:50:05 UTC. Available from: https://github.com/yanyachen/MLmetrics [ Google Scholar ] 39. Mayer M. splitTools: Tools for Data Splitting [Internet]; 2020. [cited 2025 Sept 5]. 1.0.1 p. Available from: https://CRAN.R-project.org/package=splitTools [ Google Scholar ] 40. Kapsner LA. mlexperiments: Machine Learning Experiments [Internet]; 2023. [cited 2025 Sept 5]. 0.0.5 p. Available from: https://CRAN.R-project.org/package=mlexperiments [ Google Scholar ] 41. Kapsner LA. mllrnrs: R6-Based ML Learners for ‘mlexperiments’ [Internet]; 2023. [cited 2025 Sept 5]. 0.0.5 p. Available from: https://CRAN.R-project.org/package=mllrnrs [ Google Scholar ] 42. Deshmukh F, Merchant SS. Explainable machine learning model for predicting GI bleed mortality in the intensive care unit. Am J Gastroenterol. 2020;115(10):1657–68. doi: 10.14309/ajg.0000000000000632 [ DOI ] [ PubMed ] [ Google Scholar ] 43. Shapley LS. A value for n -person games. In: Roth AE, editor. The Shapley Value [Internet]. 1st ed. Cambridge University Press; 1988. [cited 2025 Sept 5]. p. 31–40. Available from: https://www.cambridge.org/core/product/identifier/CBO9780511528446A008/type/book_part [ Google Scholar ] 44. Muehlensiepen F, Hoffmann MJ, Nübel J, Ignatyev Y, Heinze M, Butter C, et al. Acceptance of telemedicine by specialists and general practitioners in cardiology care: cross-sectional survey study. JMIR Form Res. 2024;8:e49526. doi: 10.2196/49526 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Koos H, Parameswaran V, Claire S, Chen C, Kalwani N, Osmanlliu E, et al. Drivers of variation in telemedicine use during the COVID-19 pandemic: the experience of a large academic cardiovascular practice. J Telemed Telecare. 2024;30(7):1163–72. doi: 10.1177/1357633X221130288 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Tisdale RL, Ferguson JM, Van Campen J, Greene L, Wray CM, Zulman DM. Patient-, provider-, and facility-level contributors to the use of cardiology telehealth care in the Veterans Health Administration: retrospective cohort study. J Med Internet Res. 2024;26:e53298. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Rossano A, Crijns T, Ring D, Reichenberg J. Clinician preferences for current and planned future use of telemedicine. Telemed e-Health. 2022;28(9):1293–9. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Sharma AE, Khoong EC, Nijagal MA, Lyles CR, Su G, DeFries T, et al. Clinician experience with telemedicine at a safety-net hospital network during COVID-19: a cross-sectional survey. J Health Care Poor Underserved. 2021;32(2 Suppl):220–40. doi: 10.1353/hpu.2021.0060 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Miner H, Fatehi A, Ring D, Reichenberg JS. Clinician telemedicine perceptions during the COVID-19 pandemic. Telemed e-Health. 2021;27(5):508–12. [ DOI ] [ PubMed ] [ Google Scholar ] 50. Connolly SL, Miller CJ, Gifford AL, Charness ME. Perceptions and use of telehealth among mental health, primary, and specialty care clinicians during the COVID-19 pandemic. JAMA Netw Open. 2022;5(6):e2216401. doi: 10.1001/jamanetworkopen.2022.16401 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Vinadé Chagas ME, Sigal Linhares T, De Oliveira Laguna Silva G, Ricardo Fernandes G, Motta Dias da Silva M, Tizianel Aguilar G, et al. Evaluation of behavioral change intentions among primary healthcare physicians with the use of telemedicine. Fam Pract. 2025;42(4):cmaf039. doi: 10.1093/fampra/cmaf039 [ DOI ] [ PubMed ] [ Google Scholar ] 52. Schwamm LH, Chumbler N, Brown E, Fonarow GC, Berube D, Nystrom K. Recommendations for the implementation of telehealth in cardiovascular and stroke care: a policy statement from the American Heart Association. Circulation. 2017. [cited 2025 Oct 24];135(7). Available from: https://www.ahajournals.org/doi/10.1161/CIR.0000000000000475 [ DOI ] [ PubMed ] [ Google Scholar ] 53. Kuan PX, Chan WK, Fern Ying DK, Rahman MAA, Peariasamy KM, Lai NM, et al. Efficacy of telemedicine for the management of cardiovascular disease: a systematic review and meta-analysis. Lancet Digit Health. 2022;4(9):e676–91. doi: 10.1016/S2589-7500(22)00124-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Muehlensiepen F, Knitza J, Marquardt W, Engler J, Hueber A, Welcker M. Acceptance of telerheumatology by rheumatologists and general practitioners in Germany: nationwide cross-sectional survey study. J Med Internet Res. 2021;23(3):e23742. doi: 10.2196/23742 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Brauns HJ, Loos W. Telemedizin in Deutschland: Stand – Hemmnisse – Perspektiven. Bundesgesundheitsbl. 2015;58(10):1068–73. [ DOI ] [ PubMed ] [ Google Scholar ] 56. Gerke S, Stern AD, Minssen T. Germany’s digital health reforms in the COVID-19 era: lessons and opportunities for other countries. NPJ Digit Med. 2020;3:94. doi: 10.1038/s41746-020-0306-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 57. Raes S, Annemans L, Willems R, Trybou J. Physicians’ views on optimal use and payment system for telemedicine: a qualitative study. BMC Health Serv Res. 2023;23(1):292. doi: 10.1186/s12913-023-09314-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Vaid A, Chan L, Chaudhary K, Jaladanki SK, Paranjpe I, Russak A, et al. Predictive approaches for acute dialysis requirement and death in COVID-19. Clin J Am Soc Nephrol. 2021;16(8):1158–68. doi: 10.2215/CJN.17311120 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Wang L, Zhao C, Liu X, Chen X, Li C, Wang T, et al. Non-linear effects of the built environment and social environment on bus use among older adults in China: an application of the XGBoost model. Int J Environ Res Public Health. 2021;18(18):9592. doi: 10.3390/ijerph18189592 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] PLOS Digit Health. doi: 10.1371/journal.pdig.0001359.r001 Decision Letter 0 Onicio Leal-Neto Onicio Leal-Neto Academic Editor Find articles by Onicio Leal-Neto Author information Copyright and License information Roles Onicio Leal-Neto : Academic Editor © 2026 Onicio Leal-NetoOnicio Leal-Neto This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 26 Aug 2025 Response to Reviewers '. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.'. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.-->-->* A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled ' Revised Manuscript with Track Changes '.'.-->-->* An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled ' Manuscript '.'.-->--> -->-->If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.-->--> -->-->We look forward to receiving your revised manuscript.-->--> -->-->Kind regards,-->--> -->-->Onicio Batista Leal-Neto-->-->Academic Editor-->-->PLOS Digital Health-->--> -->-->Onicio Leal-Neto-->-->Academic Editor-->-->PLOS Digital Health-->--> -->-->Leo Anthony Celi-->-->Editor-in-Chief-->-->PLOS Digital Health-->-->orcid.org/0000-0001-6712-6626-->--> -->--> Journal Requirements: -->--> -->-->If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. -->--> -->--> Additional Editor Comments (if provided): -->--> -->--> -->--> -->-->[Note: HTML markup is below. Please do not edit.]-->--> -->--> Reviewers' Comments: -->--> -->-->Reviewer's Responses to Questions Comments to the Author 1. Does this manuscript meet PLOS Digital Health’s publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: No Reviewer #2: Yes Reviewer #3: Yes ********** 2. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: I don't know Reviewer #2: No Reviewer #3: Yes ********** 3. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: No Reviewer #2: No Reviewer #3: No ********** 4. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes ********** Reviewer #1: The authors incorporated statistical and machine learning approaches to investigate the factors associated with the use of telemedicine using the survey data from 112 healthcare professionals. Univariate analysis was first employed to select the factor most strongly associated with the TM use. This is followed by multivariate analyses, including Bayesian Model Averaging, and Shapley value analysis using the XGBoost. While this study raises an interesting question, its utility, and its additional benefit it brings to the clinical community compared to existing studies, remain unclear. In addition, the methodologies in this manuscript were not described in sufficient details, thus making judging the soundness of their methodologies difficult in its current form. Overall, this manuscript requires substantial improvement before it can be considered for publication. Major concerns: 1) Essentially, this study employed standard statistical and machine learning approaches to analyze the survey data collected from a previous study from the same group, and arrived at a very similar conclusion. Although the authors have provided extensive discussions, the additional benefits this study brings compared to existing studies still need to be further clarified and strengthened. For example, what are the strengths of the methodologies in this study compared to those from the previous studies? Or are there any novel insights that this study provides that were not found by the analytical approaches used in previous studies? 2) Overall, the methodology and the analyses in this study need to be described in much more specific details in order to properly assess the validity of this study, as various crucial information seems to be missing, and some of the descriptions are also unclear and confusing. My specific concerns about these descriptions will be detailed in the following comments, but here are some of the most relevant concerns: a) To properly evaluate a machine learning model, the cohort used for reporting the performance of the predictive models should never be used during the variable selection and the model training. While the author used cross-validation in the training process, it is unclear whether the variable selection via univariate analysis was conducted using the entire cohort, or within the training cohort of each fold. If it is the former case, the testing partition would have been inadvertently used during the variable selection stage, which would result in information leakage and thereby inflate the performance metrics. Please clarify the independence of data partition in each stage of the analysis. If there is potential information leakage, the authors could either conduct variable selection only within the training data of each fold, or use an independent evaluation cohort for assessing model performance. b) The 34 items of the questionnaire, which is the central part of the study design, should be provided in the manuscript. In addition to providing the complete set of questionnaire items, how the items related to TM barriers were defined requires further clarification. According to the survey questions in reference 19, the items related to the TM barriers are under the question “What prevents you from using telemedicine? (Multiple selections were possible.)“ Does answering this question directly indicate that the participant did not use telemedicine? If so, the link between TM use and these barrier-related variables should be self-evident. In addition, there are some barrier-related variables that somehow “promote” the use of TM for these participants (according to Table 4). If these barrier-related variables reflect the participating clinicians’ personal use of TM rather than their opinion on the TM use as a public agenda (as indicated by reference 19), this finding will be very counterintuitive and contradicts the authors’ own conclusions in the Discussion section that “TM remains underutilized due to inadequate reimbursement.” Specific comments: 1) In the Methods section, Variable Selection subsection: a) “One-hot encoding was applied to all categorical variables with more than two categories.” Judging from the results presented by the authors, ordinal variables were also treated as categorical variables, which limited the explainability of the results. For example, only the age group between 40 and 49 years was considered relevant to predicting TM use, whereas how younger and older age groups are associated with TM use remained unknown. Using approaches that properly handle ordinal variables would greatly improve the explainability of their findings. b) As mentioned in the Major Concerns, the cohort used for reporting the final model performance (Table 2) should not be used during variable selection and the model training. Please clarify if this is the case in the manuscript. 2) In the Methods section, Machine learning analysis: a) On “A balanced 70%/30% train/test split was created using downsampling, ensuring no HCPs appeared in both training and test datasets.”: Please be more specific about how the downsampling was performed. i) By creating a “balanced” training/testing split, did the author intend to convey that the number of HCPs who use TM equals those who do not use TM? ii) Was the downsampling conducted completely at random? In addition, was the random downsampling performed independently for each fold of the cross-validation? Or done only once prior to cross-validation? iii) Was the testing set also downsampled? If so, it may inadvertently introduce unwanted selection bias, and the reported values for positive/negative will no longer be valid since it alters the data distribution iv) “ensuring no HCPs appeared in both training and test datasets” seems to suggest that there are cases where an HCP has filled multiple questionnaires. Please specify if this is true and justify the use of multiple questionnaires per HCP if needed. b) It was suggested in the abstract that nested cross-validation was used. However, how such nested cross-validation was conducted was either not described in sufficient detail or was described inaccurately. A truly nested cross-validation would consist of an outer loop and an inner loop of the cross-validation. Nevertheless, there is no clear description of how such nested cross-validation was conducted. Additionally, regarding “Final tuning hyperparameter values were chosen based on the best out-of-fold AUROC. Model performance was assessed using AUROC and F1 scores.” Please confirm if the data partition used for selecting the hyperparameter is different from the partition used for assessing the final model performance. c) Since there are multiple XGBoost models from the 10-fold cross-validation, please clarify how the Shapley values were combined and presented. 3) In the Results section: a) The purpose of each analysis was not clearly stated in the main text, and the writing should be improved to provide a coherent flow of logic. b) On the subsection title, “Patient characteristics”: the participants in this study are not patients but healthcare professionals instead. c) Methods such as BMA and XGBoost should be cited upon their first occurrence. d) Table 2: The purpose of McNemar's test was not described in the manuscript. e) Some of the findings also seem to be self-contradictory. Most notably, both “TM use relevant in acute events” and “TM use less suitable in acute events” were considered as limiting factors in Table 3. f) Table 4 shows a marked divergence between the findings of the two analytical approaches. Although the authors partially addressed this by focusing solely on the congruent findings across these two approaches, it would be more methodologically sound to conduct dedicated sensitivity analyses to investigate how much the findings vary between different initialization criteria or different folds of the cross-validation. 4) In the Discussion section, Strengths and Limitations subsection: a) “advanced statistical methodologies” is an overstatement, since BMA has been used for decades. b) “The five models with the highest posterior probabilities (CPPs) together accounted for only 24.1% of the total model weight…This reinforces the importance of our BMA approach, which averages across the full set of plausible models, thereby providing effect estimates that are more robust and less sensitive to arbitrary variable inclusion decisions.” A low CPP when using the top-five models does not automatically lead to the conclusion that BMA can provide more robust estimates when using the full set of plausible models. Quantitative evidence needs to be provided. Reviewer #2: Petit and colleagues submitted the manuscript : "Predicting telemedicine use among healthcare professionals in cardiology: a Bayesian model averaging and machine learning analysis of German survey data". Usage of telemedicine in modern medicine is a relevant topic. This manuscript is nicely written in logical order and easy to follow. It analyzes variables that influence implementation of telemedicine in Germany. However, this title doesn't fully match the text of the manuscript. The authors mentioned machine learning (ML) models such as XGBoost, but the results of the ML testing were not fully analyzed nor explained in details in the manuscript. ML model was not meaningfully incorporated in this study, it's not clear why did authors perform this ML analysis. This study could be completely explained by traditional statistical models. Bayes method is one of them, but significance of these findings is not well explained. If the authors want to keep ML analysis in the text, they need to better elaborate the findings and to compare results with at least 1-2 similar studies where ML models were used. It's not easy to recognize contribution of this manuscript. Most of findings are well known. I would suggest the authors to complete ML part with better explanations especially in the discussion section. And comparison of ML findings with other similar studies. Or the alternative is to completely remove the ML study from the text and expand statistical analysis and provide significance of the findings using adequate statistical methods. After correction of the text as suggested they could resubmit the manuscript. Reviewer #3: The paper presents a secondary analysis of survey data from German healthcare professionals to determine the factors that influence and forecast the use of telemedicine in cardiology. The work is technically sound and provides insightful information about the uptake of digital health from a data science standpoint. Below are my suggestions to the authors to improve the paper: 1. To meet PLOS's data sharing guidelines, think about storing anonymised survey data in a public repository. 2. The rigor section would be strengthened by a brief explanation of possible overfitting hazards despite nested cross-validation, given the relatively small sample size (n=112) and large number of initial variables (n=53). 3. Although the authors point out that there isn't an independent test set, the significance of the paper would be increased with a clear plan for future validation. 4. While handling missing data as distinct categories makes sense, a sensitivity analysis could bolster robustness assertions even further. 5. Although figures are instructive, interpretability might be improved by including confidence intervals for SHAP value ranks or by including an additional plot that illustrates feature interaction effects. ********** what does this mean? ). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files. Do you want your identity to be public for this peer review? If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public. For information about this choice, including consent withdrawal, please see our Privacy Policy ..--> Reviewer #1: Yes: Shih-Yen LinShih-Yen LinShih-Yen LinShih-Yen Lin Reviewer #2: No Reviewer #3: Yes: Mohammad Ali TakallouMohammad Ali TakallouMohammad Ali TakallouMohammad Ali Takallou ********** Figure resubmission: -->--> -->-->While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/ . PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at [email protected]. Please note that Supporting Information files do not need this step. If there are other versions of figure files still present in your submission file inventory at resubmission, please replace them with the PACE-processed versions.-->--> -->--> Reproducibility: -->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->?> PLOS Digit Health. 2026 Apr 20;5(4):e0001359. doi: 10.1371/journal.pdig.0001359.r002 Author response to Decision Letter 1 Article notes Copyright and License information Collection date 2026 Apr. PMC Copyright notice 30 Oct 2025 Attachment Submitted filename: Reviewers_Comments_241025.docx pdig.0001359.s008.docx (167.8KB, docx) PLOS Digit Health. doi: 10.1371/journal.pdig.0001359.r003 Decision Letter 1 Mathew Kiang Mathew Kiang Section Editor Find articles by Mathew Kiang Author information Copyright and License information Roles Mathew Kiang : Section Editor © 2026 Mathew KiangMathew Kiang This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 10 Mar 2026 Response to Reviewers '. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.'. This file does not need to include responses to any formatting updates and technical items listed in the 'Journal Requirements' section below.-->-->* A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled ' Revised Manuscript with Track Changes '.'.-->-->* An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled ' Manuscript '.'.-->--> -->-->If you would like to make changes to your financial disclosure, competing interests statement, or data availability statement, please make these updates within the submission form at the time of resubmission. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.-->--> -->-->We look forward to receiving your revised manuscript.-->--> -->-->Kind regards,-->--> -->-->Mathew V. Kiang, PhD-->-->Section Editor-->-->PLOS Digital Health-->--> -->-->Mathew Kiang-->-->Section Editor-->-->PLOS Digital Health-->--> -->-->Leo Anthony Celi-->-->Editor-in-Chief-->-->PLOS Digital Health-->-->orcid.org/0000-0001-6712-6626-->--> -->--> -->--> Journal Requirements: -->--> -->-->If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. -->--> -->--> Additional Editor Comments (if provided): -->--> -->--> -->--> -->-->[Note: HTML markup is below. Please do not edit.]-->--> -->--> Reviewers' Comments: -->--> -->-->Reviewer's Responses to Questions Comments to the Author Reviewer #1: (No Response) Reviewer #3: All comments have been addressed ********** publication criteria ? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.? Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe methodologically and ethically rigorous research with conclusions that are appropriately drawn based on the data presented.-->?> Reviewer #1: Yes Reviewer #3: Yes ********** 3. Has the statistical analysis been performed appropriately and rigorously?-->?> Reviewer #1: Yes Reviewer #3: Yes ********** 4. Have the authors made all data underlying the findings in their manuscript fully available (please refer to the Data Availability Statement at the start of the manuscript PDF file)??> The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception. The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: No Reviewer #3: No ********** 5. Is the manuscript presented in an intelligible fashion and written in standard English??> Reviewer #1: Yes Reviewer #3: Yes ********** Reviewer #1: In this revision, the authors have addressed most of the concerns raised in the previous round, and the quality of the manuscript has greatly improved. However, several concerns remain: 1. For some reason, the Supplementary Materials mentioned in the response and the revision were not found in the compiled pdf proof (which could either be due to issues of submission system or mistakes during the submission process). This include all supplementary figures and tables, in addition to the full list of questionnaire items. Additionally, the supplementary item corresponding the questionnaire should also be explicitly cited in the main text. 2. I appreciate the authors for clarifying the reasons behind the findings related to the relevance and suitability of TM for acute events. However, I believe this point could benefit from further clarification in the text. The authors suggest that the apparent contradiction is attributable to clinicians perceiving specific subtypes of acute situations as appropriate for TM (e.g., diagnostic support or initial triage) while regarding other acute scenarios as less suitable (e.g., those requiring direct physical intervention). It would be helpful to know whether this is a hypothesis proposed by the authors or actual observations from the experimental results, as I could not find mention of these nuances in the specific acute scenarios (and the questionnaire was not included in this revision). Were these specific acute scenarios explicitly distinguished in the questionnaire? Or is further investigation needed to fully understand the perceived suitability for TM across these scenarios? In either case, I would recommend addressing this distinction clearly in the manuscript. Since the greatest value of this work lies in its actionable insights, clarifying this context will prevent readers from being confused by the same observation. (I also noticed that the explanation provided in the response has not yet been fully incorporated into the manuscript text). 3. The survey data is currently available as a supplemental file in this submission, but the authors stated otherwise in the Data availability Statement and the response to another reviewer. Please kindly confirm whether sharing of survey data is indeed intended. 4. Typos: Introduction, “….where multiple interdependent factors are at play (17) or ML, we applied XGBoost…” Reviewer #3: (No Response) ********** what does this mean? ). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files. Do you want your identity to be public for this peer review? If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public.If you choose “no”, your identity will remain anonymous but your review may still be made public. For information about this choice, including consent withdrawal, please see our Privacy Policy ..--> Reviewer #1: Yes: Shih-Yen LinShih-Yen LinShih-Yen LinShih-Yen Lin Reviewer #3: Yes: Mohammad Ali TakallouMohammad Ali TakallouMohammad Ali TakallouMohammad Ali Takallou ********** Figure resubmission: -->--> -->-->-->While revising your submission, we strongly recommend that you use PLOS’s NAAS tool ( https://ngplosjournals.pagemajik.ai/artanalysis ) to test your figure files. NAAS can convert your figure files to the TIFF file type and meet basic requirements (such as print size, resolution), or provide you with a report on issues that do not meet our requirements and that NAAS cannot fix.-->--> Reproducibility: -->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->--> -->-->To enhance the reproducibility of your results, we recommend that authors of applicable studies deposit laboratory protocols in protocols.io, where a protocol can be assigned its own identifier (DOI) such that it can be cited independently in the future. Additionally, PLOS ONE offers an option to publish peer-reviewed clinical study protocols. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols-->?> PLOS Digit Health. 2026 Apr 20;5(4):e0001359. doi: 10.1371/journal.pdig.0001359.r004 Author response to Decision Letter 2 Article notes Copyright and License information Collection date 2026 Apr. PMC Copyright notice 20 Mar 2026 Attachment Submitted filename: Reviewers_Comments_241025.pdf pdig.0001359.s009.pdf (263.8KB, pdf) PLOS Digit Health. doi: 10.1371/journal.pdig.0001359.r005 Decision Letter 2 Mathew Kiang Mathew Kiang Section Editor Find articles by Mathew Kiang Author information Copyright and License information Roles Mathew Kiang : Section Editor © 2026 Mathew KiangMathew Kiang This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited., which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice 29 Mar 2026 Telemedicine adoption in cardiology: determinants and predictors identified using Bayesian model averaging and machine learning PDIG-D-25-00449R2 Dear Dr. Muehlensiepen, We are pleased to inform you that your manuscript 'Telemedicine adoption in cardiology: determinants and predictors identified using Bayesian model averaging and machine learning' has been provisionally accepted for publication in PLOS Digital Health. Before your manuscript can be formally accepted you will need to complete some formatting changes, which you will receive in a follow-up email from a member of our team. Please note that your manuscript will not be scheduled for publication until you have made the required changes, so a swift response is appreciated. IMPORTANT: The editorial review process is now complete. PLOS will only permit corrections to spelling, formatting or significant scientific errors from this point onwards. Requests for major changes, or any which affect the scientific understanding of your work, will cause delays to the publication date of your manuscript. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they'll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact [email protected]. Thank you again for supporting Open Access publishing; we are looking forward to publishing your work in PLOS Digital Health. Best regards, Mathew V. Kiang, PhD Section Editor PLOS Digital Health *********************************************************** Additional Editor Comments (if provided): Reviewer Comments (if any, and for reference): Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials S1 Table. Full list of questionnaire items. Complete list of all 34 items included in the pre-validated web-based survey assessing knowledge, acceptance, and utilization of telemedicine among healthcare professionals in cardiology care. (DOCX) pdig.0001359.s001.docx (23.1KB, docx) S2 Table. BMA results from sensitivity analysis. Determinants of telemedicine use identified through Bayesian model averaging after exclusion of participants with missing values, including posterior inclusion probabilities and coefficient estimates. (DOCX) pdig.0001359.s002.docx (16.2KB, docx) S3 Table. XGBoost predictor importance from sensitivity analysis. Top predictors of telemedicine use identified by the XGBoost model in the sensitivity analysis, including mean absolute SHAP values and directionality. (DOCX) pdig.0001359.s003.docx (16.8KB, docx) S1 Fig. BMA radar plot from sensitivity analysis. Profile comparison of telemedicine users versus non-users based on determinants identified through Bayesian model averaging after exclusion of participants with missing values. (TIF) pdig.0001359.s004.tif (110.8KB, tif) S2 Fig. SHAP feature importance plot from sensitivity analysis. Feature importance according to SHAP values for the prediction of telemedicine use in the sensitivity analysis excluding participants with missing values. (TIF) pdig.0001359.s005.tif (53.8KB, tif) S1 Data. Anonymized Survey Data Set. De-identified individual-level response data from the web-based survey among healthcare professionals in cardiology care.(CSV) pdig.0001359.s006.csv (9.7KB, csv) Attachment Submitted filename: Reviewers_Comments_241025.docx pdig.0001359.s008.docx (167.8KB, docx) Attachment Submitted filename: Reviewers_Comments_241025.pdf pdig.0001359.s009.pdf (263.8KB, pdf) Data Availability Statement The anonymized survey dataset underlying the findings of this study, including the participant data and data dictionary, is provided in the Supporting information files. The codes used in the analysis of this study is publicly available at: https://doi.org/10.5281/zenodo.17047856 . Articles from PLOS Digital Health are provided here courtesy of PLOS ACTIONS View on publisher site PDF (911.9 KB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Reduced susceptibility to triclosan and proposed epidemiological cut-off values for disinfectants and heavy metals in Escherichia coli isolated from Thai tilapia farms: Implications for One Health antimicrobial resistance surveillance.#1039988
• Edible coatings and films in meat preservation: Materials, functional mechanisms, and technological challenges.#1039989
• Palm Vein Biometrics A Rapid Patient Identification Tool for Emergency and Trauma Care.#1039990
• Directions for Establishing a Korean National One Health Joint Plan of Action: Focusing on Zoonotic Disease Preparedness and Response.#1039991
• Intensive arable farming calls for resilient wild bee restoration amid severe abundance fluctuations.#1039992