A machine learning model to simplify recognition of patients with atrial fibrillation based on diagnostic codes in Swedish primary health care - 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 BMC Med Inform Decis Mak . 2026 Apr 17;26:126. doi: 10.1186/s12911-026-03491-4 Search in PMC Search in PubMed View in NLM Catalog Add to search A machine learning model to simplify recognition of patients with atrial fibrillation based on diagnostic codes in Swedish primary health care Anders Norrman Anders Norrman 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 2 Academic Primary Health Care Centre, Region Stockholm, Stockholm, Sweden Find articles by Anders Norrman 1, 2 , Caroline Wachtler Caroline Wachtler 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 2 Academic Primary Health Care Centre, Region Stockholm, Stockholm, Sweden Find articles by Caroline Wachtler 1, 2 , Per Wändell Per Wändell 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 3 Center for Primary Health Care Research, Lund University, Malmö, Sweden Find articles by Per Wändell 1, 3 , Julia Eriksson Julia Eriksson 4 Division of Biostatistics, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden Find articles by Julia Eriksson 4 , Toralph Ruge Toralph Ruge 5 Department of Clinical Sciences Malmö, Lund University, Lund, Sweden 6 Department of Internal Medicine, Skåne University Hospital, Malmö, Sweden Find articles by Toralph Ruge 5, 6 , Boel Brynedal Boel Brynedal 7 Centre for Epidemiology and Community Medicine, Region Stockholm, Stockholm, Sweden 8 Department of Global Public Health, Karolinska Institutet, Stockholm, Sweden Find articles by Boel Brynedal 7, 8 , Jan Hasselström Jan Hasselström 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 2 Academic Primary Health Care Centre, Region Stockholm, Stockholm, Sweden Find articles by Jan Hasselström 1, 2 , Thomas Kahan Thomas Kahan 9 Division of Cardiovascular Medicine, Department of Clinical Sciences, Karolinska Institutet, Danderyd Hospital, Stockholm, Sweden Find articles by Thomas Kahan 9 , Axel C Carlsson Axel C Carlsson 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 2 Academic Primary Health Care Centre, Region Stockholm, Stockholm, Sweden Find articles by Axel C Carlsson 1, 2, ✉ Author information Article notes Copyright and License information 1 Division of Family Medicine and Primary Care, Department of Neurobiology, Care Sciences and Society, Karolinska Institutet, Huddinge, Sweden 2 Academic Primary Health Care Centre, Region Stockholm, Stockholm, Sweden 3 Center for Primary Health Care Research, Lund University, Malmö, Sweden 4 Division of Biostatistics, Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden 5 Department of Clinical Sciences Malmö, Lund University, Lund, Sweden 6 Department of Internal Medicine, Skåne University Hospital, Malmö, Sweden 7 Centre for Epidemiology and Community Medicine, Region Stockholm, Stockholm, Sweden 8 Department of Global Public Health, Karolinska Institutet, Stockholm, Sweden 9 Division of Cardiovascular Medicine, Department of Clinical Sciences, Karolinska Institutet, Danderyd Hospital, Stockholm, Sweden ✉ Corresponding author. Received 2025 Dec 7; Accepted 2026 Apr 10; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13097664 PMID: 41998624 Abstract Background Atrial fibrillation (AF) is a major risk factor for atherothrombotic complications but is often asymptomatic and undiagnosed. This study aimed to develop a machine learning model to distinguish between individuals with low and high risk of AF, using routinely collected diagnostic data from Swedish primary health care. Methods Cases ( n = 42,607, aged ≥ 45 years) with diagnosed new onset AF and controls ( n = 427,169) matched by age and sex. Machine learning models stratified for age (45–69 and ≥ 70 years) and sex were developed using stochastic gradient boosting, based on number of primary health care visits during the year before the index AF diagnosis, age, and ICD-10 codes from electronic medical records 2014–2019. Performance was evaluated by AUC, sensitivity and specificity, and key predictors ranked by normalized relative influence (NRI) and odds ratios for marginal effects. Results The most influential predictors were the number of visits (NRI: 29.9–46.3%) and age (NRI: 6.2–15.9%), followed by risk factors for AF such as heart failure, hypertension, and cardiac arrhythmias. Model AUC ranged from 0.77 to 0.79 across subgroups. Sensitivity was 0.76–0.80, and specificity 0.58–0.66, with higher sensitivity in older groups and higher specificity in younger ones. The models correctly identified 95–98% of individuals without known AF. Conclusions The models show good predictive ability, effectively ruling out low-risk patients while identifying known risk factors. With AUC values comparable to more complex models, our approach using only visit frequency, age, and diagnoses may support initial risk assessment in primary health care for identifying individuals at risk of AF. Supplementary Information The online version contains supplementary material available at 10.1186/s12911-026-03491-4. Keywords: Artificial intelligence, Atrial fibrillation, Gradient boosting, Normalized relative influence, Prediction Introduction Atrial fibrillation (AF) affects an estimated 2–3% of the general population, with prevalence rising steeply with advancing age; however, a considerable fraction of cases remains undetected [ 1 ]. AF is a major cause of atherothrombotic stroke, and early detection is crucial, as adequate anticoagulant therapy substantially reduces the risk of stroke [ 1 ]. In Sweden, the prevalence of detected AF is estimated to be 2.9% among adults above the age of 20 [ 2 ]. However, in the STROKESTOP study, 3.0% of 75- and 76-year-olds in Sweden who participated in AF screening were found to have previously undiagnosed AF, contributing to a total AF prevalence of 12.3% in that screened population [ 3 ]. In the Swedish context, population-based screening with single-lead handheld ECG has been shown to increase the previously known prevalence of AF by approximately 30%, indicating that a substantial proportion of cases remain undetected without active screening [ 4 ]. To detect AF, various strategies have been proposed, including general, targeted, and opportunistic screening, but the 2024 European Society of Cardiology (ESC) Guidelines for the management of AF recommend selective approaches and do not support universal population-based screening [ 1 ]. The use of clinical data to identify risk for AF improves the detection rates of previously unrecognized AF. Many established risk factors and risk markers are associated with incident AF, including older age, male sex, low socioeconomic status, hypertension, heart failure, coronary heart disease, prior atherothrombotic stroke, diabetes mellitus, elevated B-type natriuretic peptide (BNP), elevated C-reactive protein, low estimated glomerular filtration rate (eGFR), body mass index (BMI), and electrocardiographic (ECG) abnormalities [ 1 ]. The use of machine learning (ML) methods is becoming increasingly common in studies aiming to enhance the classification of patients at high risk of undetected chronic diseases, including AF [ 5 ], potentially offering a complement or an alternative to traditional risk prediction models [ 6 – 8 ]. Previous ML models for AF prediction have been developed using diverse data types, including electronic medical records, biomarkers, imaging findings, and electrophysiological parameters, reflecting the evolving landscape of AF risk stratification [ 9 ]. Many of these models, however, rely on extensive sets of clinical variables, reducing their feasibility for clinical application [ 10 – 12 ]. In primary health care (PHC), efficient use of limited resources and time are important creating a need to limit screening those at the highest risk of having undiagnosed AF -- those who would benefit the most from timely detection and treatment [ 13 ]. There is no population-based AF screening program in Stockholm County; however, patients undergoing evaluation for atherothrombotic stroke, acute coronary syndromes, or other cardiovascular symptoms routinely undergo ECG assessment in secondary care, and pulse palpation during blood pressure measurement is widely recommended. We hypothesized that routinely collected register data, such as number of PHC visits, age, and International Statistical Classification of Diseases and Related Health Problems (ICD-10) codes, might predict individuals who would benefit from further evaluation, such as ECG or pulse palpation. These features are standardized, widely available, and easily extracted, which could make them suitable for scalable applications in real-world PHC settings. Thus, the aim of this study was to develop ML models based solely on routinely collected PHC data to distinguish between individuals at low and high risk of AF, supporting more efficient targeted screening strategies, which may ultimately contribute to prevent future atherothrombotic stroke. Methods Study design This study employed a case–control design using routinely collected electronic medical records from PHC in Region Stockholm, Sweden to examine whether a limited set of standardized clinical variables can distinguish between individuals at low and high risk of AF. Study context Data were collected from patients in the Stockholm Region, which had a population of approximately 2.4 million residents at the time of the study. Data for this study was gathered from the Stockholm Region regional health care data warehouse (known as VAL-databasen), where healthcare data, including diagnoses (according to ICD-10), medical visits, and hospitalization at all levels of care are systematically recorded. A dataset of comprehensive healthcare information from all levels of care within the Stockholm Region health care is available at Karolinska Institutet for research purposes (Multimorbidity Integrated Registry Across Care Levels in Stockholm, MIRACLE-S). The study adheres to the TRIPOD-AI reporting guidelines for transparent reporting of multivariable prediction models developed for individual prognosis or diagnosis [ 14 ]. Participants Data for this study was gathered from cases and controls were identified based on electronic medical records data from the years 2014–2019. All individuals who had received an ICD-10 diagnosis of atrial fibrillation or atrial flutter (I48) recorded in connection with any healthcare visit, in primary or secondary care, between 2010 and 2013 were excluded. Atrial fibrillation or atrial flutter (I48), identified either in primary or secondary care settings, are collectively referred to as AF in this study. In Swedish PHC, atrial fibrillation and atrial flutter are not consistently distinguished in routine documentation; therefore, reliable separation between the two conditions was not feasible in this dataset, and at the stage of risk stratification, such differentiation is not clinically meaningful. Cases were defined as individuals aged ≥ 45 years, registered at one or more PHC centers in Region Stockholm [ 15 ], who had received the ICD-10 code for atrial fibrillation or atrial flutter (I48) recorded in connection with health care visit during 2014–2019. The lower age limit of 45 years was chosen based on Swedish population data demonstrating that the prevalence of AF is very low below 45 years of age and increases markedly with advancing age thereafter [ 16 ]. Between 2014 and 2019, 43,771 individuals were identified with a newly registered diagnosis of AF. For each identified case, up to ten age- and sex-matched controls were selected from the same time period (2014–2019). In some instances, fewer than ten suitable controls were available. If more than ten eligible matched candidates were available, the ten controls were selected randomly. Sampling was performed without replacement, meaning that no individual was included more than once in the control group. After limiting the population to only include individuals ≥ 45 years of age, we ended up with a study population of 42,607 newly diagnosed patients with AF and 427,169 controls. Whether this diagnosis was made for the first time in PHC or had been established previously elsewhere in Sweden or abroad could not be determined. The correctness of the AF diagnosis was not validated. The first time the diagnosis was recorded was defined as the index date. How a documented diagnosis was established and whether the controls underwent any investigations to screen for AF is not evident from the available data. Age and male sex are known risk factors for AF, and the number of consultations has proven to be a strong predictor in similar ML-models for other chronic diseases [ 17 ]. Including all diagnoses above a minimum frequency has been shown to improve model performance in similar models for the detection of hypertension and diabetes, demonstrating high ability to discriminate between cases and non-cases [ 17 , 18 ]. Data collected include age and sex, the number of PHC visits for 12 months preceding the index date, and all registered primary and secondary diagnoses, categorized by three-digit ICD-10 codes, recorded during all visits to a PHC centre between 2014 and 2019. First, a three-year retrospective collection of diagnoses was performed for all individuals, starting from the date of the new AF diagnosis or the corresponding index date for controls, meaning that the index date for controls could occur on a date that did not coincide with a healthcare visit generating a diagnosis. To minimize this potential bias, we included a three-year retrospective collection of diagnoses. A 3-year window was chosen as a 1-year window has been shown to catch 56% of patients with diabetes and 59% of patients with hypertension compared to a 5-year window and a 5-year window would limit the time to find AF events [ 19 ]. Diagnoses were chosen from the 2,000 most frequently registered ICD-10 codes in PHC. Of these, only diagnoses with at least 50 occurrences in the studied population were included, resulting in a total of 1,309 diagnoses [ 20 ]. This was done to exclude rare diagnostic codes. These were identified as predictors alongside number of PHC visits 12 months prior AF diagnosis or index date for controls and age. Variables To further strengthen the likelihood that the cases studied represent incident AF, in addition to excluding patients with AF diagnosis between 2010 and 2013, we also excluded diagnoses indicating a potential history of post-stroke AF as predictors. Specifically, we excluded the diagnoses a personal history of medical treatment ( Z92), including anticoagulant use (Z92.1), as well as those with a personal history of certain other conditions (ICD-10 Z86), such as previous TIA (Z86.6) or stroke (Z86.7c). This exclusion was based on preliminary analyses in which these variables showed disproportionately high NRI values, suggesting that they primarily captured prior disease history rather than incident risk, and were therefore removed to improve transparency and interpretability of the final model. Several diagnoses were merged and treated as single diagnostic predictor categories in the model. A complete list of these merged categories, including diabetes (E10–E13), hypertension (I10–I15), and coronary heart disease (I20, I21, I23, I24, I25) is provided in Supplementary Table 1 . Type 1 and type 2 diabetes were analyzed together, as many individuals had multiple ICD-10 codes for diabetes, making separation based solely on diagnostic coding unreliable [ 19 ]. Statistical analyses Data were analyzed using stochastic gradient boosting (SGB), a ML method that has previously been shown to be effective in predictive modelling of other chronic conditions within PHC settings [ 21 , 22 ]. We developed four separate models to account for the substantial differences in search patterns, diagnostic characteristics, and incidence rates of AF observed across age and sex [ 20 ]. To account for differences in AF prevalence and the distribution of risk factors, the study population was stratified by sex and age group: women and men aged 45–69 years and ≥ 70 years. Machine learning models The four SGB models used a Bernoulli loss function fitted to 20 000 trees, each having a maximum depth of 5 interactions, with a shrinkage (learning rate) of 0.001, a minimum of 10 observations in the terminal nodes of the trees, and a subsampling rate (bag fraction) of 0.5. The optimal number of trees to use for prediction was estimated using 10-fold cross validation. The optimal number of trees was used on the test data to predict AF. Model validation Within each SGB model, the data was further divided into a 50% training set and a 50% test set, to ensure robust model development and evaluation. The performances of the four models were assessed using ROC curves, sensitivity, and specificity. The SGB models were then applied to each test data set to obtain patient specific probabilities of being newly diagnosed with AF. The probabilities that maximized the sum of sensitivity and specificity were used as a cut-off value, such that patients with a probability higher than this cut-off were classified as being newly diagnosed with AF. Using the SGB models, we identified and ranked the predictors most strongly associated with newly diagnosed AF, represented by their normalized relative influence (NRI) scores and corresponding odds ratios of marginal effects (OR ME ). For each predictor, the odds ratio was calculated using the probabilities of being newly diagnosed with AF obtained by integrating out all other variables in the model using the weighted tree traversal method. All four models were assessments of their (ROC) curve performance and evaluated the NRI of predictors, including registered ICD-codes in PHC, age, and the number of PHC visits 12 months preceding the index date. For each of the four training datasets we included age, number of PHC visits 12 months prior and diagnoses with at least 10 occurrences. This resulted in 316 diagnoses for women aged 45–69, 449 for women aged ≥ 70, 427 for men aged 45–69, and 468 for men aged ≥ 70. The analyses were performed using R version 4.2.1 [ 23 ]. Results General A total of 42,607 cases, 15,298 individuals aged 45–69 years and 27,309 aged 70 years and older, were diagnosed with AF. A summary of included cases and controls according to sex and age group is presented in Table 1 . Table 1. Participant characteristics for the cases and controls included in the overall study population, and in the training and test datasets Variable Total Training dataset Test dataset Cases Controls Cases Controls Cases Controls All participants 42,607 427,169 21,528 213,362 21,079 213,807 Age at index date, mean (SD) a 73.0 (9.3) 69.9 (9.4) Women 45–69 years, n (%) 5,189 76,216 2,623 38,080 2,566 38,136 Women ≥ 70 years, n (%) 13,986 118,425 7,050 59,156 6,936 59,269 Men 45–69 years, n (%) 10,109 146,292 5,104 73,097 5,005 73,195 Men ≥ 70 years, n (%) 13,323 86,236 6,751 43,029 6,572 43,207 Open in a new tab Model performance The sensitivities, specificites, and AUC for predicting new AF in the test set by sex and age group are shown in Table 2 . The sensitivity ranged between 0.76 and 0.80 (highest in the ≥ 70 year age groups for both women and men), and specificity ranged between 0.58 and 0.66 (highest in the 45–69 years age groups). Overall, the four models demonstrated similar AUCs across age and sex groups, ranging from 0.77 to 0.79. Table 2. Statistical performance of the models in the test data sets Group Sensitivity Specificity AUC Women 45–69 years 0.78 0.64 0.785 (0.776–0.793) Women ≥ 70 years 0.80 0.58 0.765 (0.760–0.771) Men 45–69 years 0.77 0.66 0.788 (0.783–0.794) Men ≥ 70 years 0.83 0.71 0.778 (0.772–0.783) Open in a new tab Estimates for sensitivity and specificity for the age- and sex‐stratified machine learning models, with area under curve (AUC) and 95% confidence intervals (CI) The confusion matrices for women and men by age group, together with the corresponding positive and negative predictive values, are presented in Table 3 . The models demonstrated 95–98% ability to correctly classify patients in all groups without a recorded diagnosis of AF. Positive predictive values ranged from 14% to 37% in the models, with the highest fraction among men ≥ 70 years. All four SGB models reached the optimal number of trees, ranging between 1900 and 3764 (see Supplementary Table 2 ). Table 3. Confusion matrix for predicting presence of incident atrial fibrillation using the optimal stochastic gradient in the test data set Predicted Observed No atrial fibrillation Atrial fibrillation Total Women 45–69 years No atrial fibrillation n (%) 24,205 (97.8) 553 (2.2) 24,758 Atrial fibrillation n (%) 13,931 (87.4) 2,013 (12.6) 15,944 Total n (%) 38,136 (93.7) 2,566 (6.3) 40,702 Positive predictive value 0.13 Negative predictive value 0.98 Women ≥ 70 years No atrial fibrillation n (%) 34,374 (96.2) 1,363 (3.8) 35,737 Atrial fibrillation n (%) 24,895 (81.7) 5,573 (18.3) 30,468 Total n (%) 59,269 (89.5) 6,936 (10.5) 66,205 Positive predictive value 0.18 Negative predictive value 0.96 Men 45–69 years No atrial fibrillation n (%) 48,315 (97.6) 1,176 (2.4) 49,491 Atrial fibrillation n (%) 24,880 (86.7) 3,829 (13.3) 28,709 Total n (%) 73,195 (93.6) 5,005 (6.4) 78,200 Positive predictive value 0.13 Negative predictive value 0.98 Men ≥ 70 years n (%) No atrial fibrillation n (%) 26,846 (95.1) 1,398 (4.9) 28,244 Atrial fibrillation n (%) 16,361 (76.0) 5,174 (14.0) 21,535 Total n (%) 43,207 (86.8) 6,572 (13.2) 49,779 Positive predictive value 0.14 Negative predictive value 0.95 Open in a new tab Variable importance The number of PHC visits in the 12 months preceding an AF diagnosis emerged as the single strongest predictor of AF, with high effects among older individuals (e.g., OR 8.0 for men ≥ 70 years). Age at diagnosis, together with recorded diagnoses of cardiovascular conditions such as cardiac failure (I50), hypertension (I10–I15), abnormalities of heart beat (R00), and other arrhythmias (I49), showed consistently high NRI in the models, and were all strongly associated with OR ME values above 1. Overall, age contributed an NRI between 5% and 16% across the four models. Tables 4 and 5 present the ten predictors with the highest NRI, along with their corresponding OR ME values, by sex and stratified by age group. Encountering health services (Z76) was among the 10 most influential predictors in all four models and had an OR ME slightly below 1 in all models. Table 4. The top 10 predictors, including diagnoses defined by ICD-10 codes recorded in the electronic medical record, with highest normalized relative influence for predicting incident atrial fibrillation among women, along with the odds ratios for the marginal effects Predictors Relative influence (%) OR ME Women 45–69 years Number of visits 12 months before diagnosis 29.9 3.4 Age at diagnosis 15.9 0.4 Abnormalities of heart beat (R00) 5.9 3.0 Cardiac failure (I50) 4.6 3.5 Other cardiac arrhythmias (I49) 3.0 2.8 Presence of cardiac and vascular implants and grafts (Z95) 2.0 2.5 Hypertension (I10-I15) 1.48 1.2 Overweight and obesity (E66) 1.14 1.5 Transplanted organ and tissue status (Z94) 1.1 6.2 Encountering health services (Z76) 1.1 0.9 Women ≥ 70 years Number of visits 12 months before diagnosis 37.3 3.5 Age at diagnosis 11.9 1.6 Cardiac failure (I50) 5.1 1.7 Hypertension (I10-I15) 2.6 1.3 Abnormalities of heart beat (R00) 2.5 2.2 Other cardiac arrhythmias (I49) 2.3 3.0 Convalecence (Z54) 1.5 1.6 Encountering health services (Z76) 1.0 0.9 Observation (Z03) 0.99 1.0 Paroxysmal tachycardia (I47) 0.94 3.6 Open in a new tab The optimal stochastic gradient boosting (SGB) model with 1,900 optimal trees for women 45–69 years and 3,148 trees for women ≥ 70. Z76 denotes: Persons encountering health services in other circumstances, Z03 Encounter for medical observation for suspected diseases and conditions ruled out, NRI normalized relative influence, OR ME marginal effects Table 5. The top 10 predictors, including diagnoses defined by ICD-10 codes recorded in the electronic medical record, with highest normalized relative influence for predicting incident atrial fibrillation among men, along with the odds ratios for the marginal effects Predictors Relative influence (%) OR ME Men 45–69 years Number of visits 12 months before diagnosis 33.7 5.5 Age at diagnosis 12.4 0.97 Cardiac failure (I50) 5.2 2.3 Abnormalities of heart beat (R00) 3.1 2.6 Other cardiac arrhythmias (I49) 2.7 4.0 Presence of cardiac and vascular implants and grafts (Z95) 1.97 1.6 Encountering health services (Z76) 1.27 0.9 Hypertension (I10-I15) 1.19 1.0 Paroxysmal tachycardia (I47) 0.94 4.6 Observation (Z03) 0.85 0.9 Men ≥ 70 years Number of visits 12 months before diagnosis 46.3 8.0 Age at diagnosis 6.2 1.34 Cardiac failure (I50) 3.9 1.7 Presence of cardiac and vascular implants and grafts (Z95) 1.4 1.6 Hypertension (I10-I15) 1.3 1.2 Other cardiac arrhythmias (I49) 1.18 2.0 Convalecence (Z54) 0.87 1.4 Abnormalities of breathing (R06) 0.83 1.3 Encountering health services (Z76) 0.82 0.9 Observation (Z03) 0.76 1.0 Open in a new tab The optimal stochastic gradient boosting (SGB) model with 3,764 optimal trees for men 45–69 years and 3,306 trees for men ≥ 70. Z76 denotes Persons encountering health services in other circumstances, Z03 Encounter for medical observation for suspected diseases and conditions ruled out, NRI normalized relative influence, OR ME marginal effects Discussion Main findings This study demonstrates that newly diagnosed AF can be predicted using SGB models incorporating the number of PHC visits, age, and diagnostic codes from PHC during the three years preceding diagnosis. The models showed overall good discrimination (AUC 0.77–0.79), high negative predictive values (0.95–0.98), and a clinically valuable positive predictive values (0.13–0.18), with performance comparable to previously published machine learning models [ 16 – 18 ]. The high negative predictive values and the lower positive predictive values indicate that the model is suited as a risk strafification tool, enabling more efficient use of PHC resources while still identifying patients with AF. Our study confirms previous evidence that AF can be predicted for risk stratification from electronic medical records, and extends these findings by demonstrating feasibility using Swedish PHC data, although with somewhat lower AUC values (0.77–0.79) than those reported in previous studies. A study evaluating different predictive models for AF prediction using harmonized electronic health record data, demonstrated that a single-layer neural network with random oversampling performed best, achieving an AUC of 0.80, solely based on diagnoses with a known association to incident AF, along with age and sex [ 12 ]. Previous studies have explored different ML models to identify undiagnosed AF using routinely collected primary care data, reporting AUC of 0.83 in a UK study [ 10 ], and AUC of 0.83 in a study based on clinical practice data in a German study [ 11 ]. Among the evaluated models, the model for men aged ≥ 70 years demonstrated the highest sensitivity (0.83) and specificity (0.71). A possible explanation is the higher prevalence of cardiovascular disease and the lower burden of competing non-cardiovascular diagnoses in this group compared with younger men and women in both age groups [ 20 ], which may contribute to improved diagnostic discrimination. Our study suggests that risk stratification may be achievable using the models. In the current study, as shown previously [ 10 – 12 ], findings suggest that age and sex, together with a limited set of key diagnoses, particularly hypertension, heart failure, and cardiac arrhythmias, are sufficient to identify individuals with risk of AF as well as individuals at low risk of AF. The variable with the highest NRI was Number of visits 12 months before diagnosis and this is likely explained by an increase in the likelihood of detection of AF at every visit. The diagnoses contributing most to the prediction models were conditions with well-established association with incident AF, such as cardiac arrhythmias [ 11 ], hypertension [ 6 , 7 ], and heart failure [ 6 , 7 ]. Abnormalities in heartbeat (R00) was a predictor of AF and based on our understanding of Swedish PHC and the present findings, patients with R00 should be screened for AF. R00 is sometimes coded during a remote consultation and should be followed up with an in-person visit to rule out AF and other underlying conditions. We observed that advancing age was associated with an increased risk of AF in both women and men aged ≥ 70 years, consistent with previous studies [ 6 – 8 , 24 ]. However, in contrast to this, the models for women and men aged 45–69 years demonstrated an inverse association between increasing age and AF risk. We believe that this is an effect of the matching procedure, where individuals were matched for age. Men showed a higher incidence of AF compared to women [ 25 ]. We were not able to account for ethnicity [ 26 ] and socioeconomic status [ 27 ], both recognized as important determinants of AF risk, due to a lack of available data. This study was unable to duplicate previous strong associations between AF and several chronic diseases, but this could be due to the nature of the data. Chronic kidney disease, hyperlipidemia, diabetes mellitus, peripheral artery disease, and coronary artery disease, also contributed to the models but did not rank among the top ten strongest predictors of AF risk. This likely reflects the underuse or underreporting of several chronic diagnoses in PHC and suggest that a five-year time window is sometimes needed to capture a larger proportion of chronic conditions [ 28 ]. Obesity [ 6 , 7 ] was identified as a contributing factor but not a diagnosis of alcohol dependence; however, other lifestyle and behavioral variables including the level of alcohol consumption, smoking, physical inactivity, and family history of AF were absent from our dataset. There are many potential explanations for why the number of PHC visits was the most contributing factor to prediction in all of the models. This aligns with previous findings in the prediction of type 2 diabetes [ 17 ], likely reflecting a higher burden of chronic comorbidities associated with risk of incident AF. Frequent healthcare use is associated with multiple chronic conditions, medically unexplained symptoms, and mental health issues such as anxiety and depression [ 29 – 31 ]. These patients also have more frequent opportunities to have AF recognized through clinical assessments. Persons encountering health services in other circumstances (Z76), which often represents non–face-to-face contacts such as prescription renewals, showed an OR ME below 1 across all models. Although this indicates that the model may be incomplete as a pure disease-risk predictor, the finding is of interest, as it highlights that patients with frequent non–face-to-face contacts but few in-person visits may represent a group where additional opportunistic screening could be warranted. The high influence of the number of PHC visits likely reflects its status as a continuous variable, which provides more information than dichotomous diagnoses and allows the SGB models to identify both linear and nonlinear patterns. Strengths and limitations The findings of this study should be interpreted in the context of its quality as an AI-based prediction model, taking into account five key criteria: transparent reporting and reproducibility, a clearly defined intended use, rigorous internal and external validation, an adequate sample size, and openness of data and software [ 32 ]. This study is reported according to the TRIPOD-AI guidelines [ 14 ], which provide updated standards for transparent reporting of prediction models developed with regression or ML methods. The reproducibility of our results is supported by our findings confirming results from similar ML models [ 10 – 12 ]. The study has several strengths. First, the development and validation of the model on a large, representative dataset. Secondly, a strength of the study is that it provides a well-described and detailed account of the model, including how predictors were selected, which ML algorithms were used, and the assumptions under which they were applied. However, the study was based solely on data from Stockholm County, and external validation in independent populations and health care systems are required. Third, an adequate sample size of this study was ensured by including all patients in Stockholm County aged 45 years and older with a new diagnosis of AF. This represents nearly one quarter of the Swedish population and resulted in more than 40,000 cases. Forth, the model was specifically designed to predict incident rather than post-stroke AF, achieved by excluding patients with recorded diagnoses of prior TIA/stroke (Z86) or anticoagulant treatment (Z92). However, patients prescribed oral anticoagulants were not excluded, even though oral anticoagulant use might better capture prior TIA/stroke, which is not always coded. Similarly, patients with sequelae of cerebrovascular disease (I69) were not excluded; however, this variable showed low influence in the model. Fifth, apart from regression methods, where colinearity of similar variables may be limiting, in tree-based ML models collinearity is not an issue. The study has several limitations. First, the matched case–control design may lead to overestimation of the true prevalence and positive predictive value, as a substantial proportion of the underlying population was excluded and the analyses were conducted in a selected population with a higher prevalence than that of the general population within the studied age groups. Second, a major limitation concerns how previously diagnosed AF was defined. We excluded all patients with a recorded AF diagnosis between 2010 and 2013. However, there is a risk that some individuals may have received an AF diagnosis in another healthcare system or before 2010, which represents a potential source of misclassification. Third, case identification relied on diagnostic codes without information about how AF was detected or whether diagnoses were confirmed by ECG. Controls were not systematically screened, and it is likely that undiagnosed AF was present in this group. This could lead to misclassification and reduced model performance. Fourth, no systematic screening was performed. Consequently, we cannot determine whether the model predicts the future development of AF itself or merely the likelihood of receiving a clinical diagnosis of AF. Fifth, the study was based solely on data from Stockholm County, and the lack of external validation represents another major limitation; therefore, validation in independent populations and healthcare systems is essential before the model can be considered for clinical implementation. Sixth, relates to the retrospective feature window starting from the date of the new AF diagnosis (and the corresponding index date for controls). As diagnoses recorded during the same encounter as the AF diagnosis were included, some predictors may reflect the diagnostic workup leading to AF detection rather than predictive features. Consequently, part of the model performance may be driven by signals related to the diagnostic encounter itself, potentially leading to an overestimation of predictive performance. Seventh, among women and men, aged 45–69 years, the observed odds ratio for age in both groups were below 1.0 suggested an inverse association with increasing age. This counterintuitive finding is unlikely to represent a true protective effect of age but rather reflects the matching design and resulting within-stratum comparisons. Consequently, the estimated effect of age within this subgroup should be interpreted with caution. Eight, the dominance of visit frequency may influence the interpretability of the models, as it can overshadow other predictors and reflect healthcare utilization patterns rather than disease-specific risk. Ninth. the model does not incorporate well-established clinical risk markers for AF, which may constrain predictive accuracy. Several established factors associated with incident AF, including blood pressure values, BMI, ethnicity, tobacco use, and socioeconomic status were not available. Furthermore, diagnostic codes in our study were assigned by physicians at their discretion and are therefore subject to an inherent risk of incomplete or inaccurate reporting; a limitation that has been shown to result in underreporting of conditions such as obesity, hyperlipidemia, chronic kidney disease, and peripheral artery disease in Swedish PHC [ 17 ]. Tenth, the sample sizes of the age- and sex-stratified test datasets differed, which may have resulted in minor variation in the precision of performance estimates between strata; however, all subgroups were large and demonstrated narrow confidence intervals for AUC, indicating stable model performance. Implications and further directions Findings from this study indicate that a limited set of routinely available variables in electronic medical records, age, a few key diagnoses, and the number of PHC visits, allow reliable identification of individuals unlikely to have AF. Such easily obtainable factors may thus help identify individuals who would benefit from targeted AF screening. Furthermore, such information could be integrated into automated decision-support tools within electronic medical record systems. The ML models could serve as a decision-support tool to identify patients who should undergo pulse palpation or ECG. This ML models could be accessible to all PHC staff, for example during routine visits such as suture removal, increasing task shifting and optimize work flows in PHC. Comparisons with established AF prediction scores, such as C2HEST, would be of considerable interest. However, further studies are needed to determine whether screening strategies based on such tools can improve clinical outcomes, such as reducing the risk of stroke, as suggested, but not confirmed, by recent trials [ 33 ]. Conclusions The study demonstrates that ML models to predict incident AF in a PHC setting based on the number of visits during the 12 months preceding diagnosis, age, and diagnostic codes from PHC provide a high negative predictive value, indicating potential utility for risk stratification to guide targeted AF screening. However, their clinical applicability remains uncertain, as cost-effectiveness, integration into clinical workflows, and their impact on preventing atherothrombotic stroke have not been evaluated. In addition, the validity of the SGB prediction models is limited by uncertainty about whether the diagnostic data truly reflect underlying AF status and therefore warrants further study. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (154.5KB, docx) Supplementary Material 2 (27.1KB, docx) Acknowledgements None. Abbreviations AF Atrial fibrillation AI Artificial intelligence AUC Area under the receiver operating characteristic curve BMI Body mass index BNP B–type natriuretic peptide CRP C–reactive protein ECG Electrocardiogram ESC European Society of Cardiology eGFR Estimated glomerular filtration rate ICD 10–International Statistical Classification of Diseases and Related Health Problems, 10th Revision ML Machine learning NRI Normalized relative influence ORME Odds ratio of marginal effects PHC Primary health care ROC Receiver operating characteristic SGB Stochastic gradient boosting SD Standard deviation TIA Transient ischemic attack TRIPOD AI–Transparent Reporting of a multivariable prediction model for Individual Author contributions AN researched, drafted the manuscript and revised the manuscript. CW data aquisation, revised the manuscript. PW researched and revised the manuscript. JE data management, formal analysis and revised the manuscript. TR study design and revised the manuscript. BB supervised, researched and revised the manuscript. JH supervised, researched and revised the manuscript. TK supervised, funded, researched and revised the manuscript. ACC supervised, study design, researched, funded, researched and revised the manuscript. Funding Open access funding provided by Karolinska Institute. Grants from the Swedish state under the agreement between the Swedish government and the county council in Stockholm, SLL-NSV project (FoUI-990472 and 988787), funding provided by Region Stockholm to Axel C Carlsson (FoUI-973001), and by Karolinska Institutet Research Foundations (Stockholm, Sweden; FS-2024:0017). Data availability The data from this study can be accessed for research by qualified researchers who have been trained in confidentiality protocols for human subjects, following ethical approval from Region Stockholm at [email protected]. Analytical code and programs can be obtained from (mailto: [email protected]). Declarations Ethical approval The study adheres to the Helsinki declaration. All data was pseudonymized. The use of the data was approved by the Swedish Ethical Review Authority without patient consent (reference number 2021 − 01016 with amendments 2021-05735-02, 2022–06729, 2023-07166-0202 and 2024-05462-02). The need for consent to participate was waived by Institutional Review Board at Karolinska Institutet. Consent for publication There was no patient or public involvement in the design, conduct, reporting, interpretation, or dissemination of this study. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Van Gelder IC, Rienstra M, Bunting KV, Casado-Arroyo R, Caso V, Crijns H, et al. 2024 ESC Guidelines for the management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS). Eur Heart J. 2024;45(36):3314–414. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Friberg L, Bergfeldt L. Atrial fibrillation prevalence revisited. J Intern Med. 2013;274(5):461–8. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Svennberg E, Engdahl J, Al-Khalili F, Friberg L, Frykman V, Rosenqvist M. Mass Screening for Untreated Atrial Fibrillation: The STROKESTOP Study. Circulation. 2015;131(25):2176–84. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Engdahl J, Andersson L, Mirskaya M, Rosenqvist M. Stepwise screening of atrial fibrillation in a 75-year-old population: implications for stroke prevention. Circulation. 2013;127(8):930–7. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Petzl AM, Jabbour G, Cadrin-Tourigny J, Pürerfellner H, Macle L, Khairy P, et al. Innovative approaches to atrial fibrillation prediction: Should polygenic scores and machine learning be implemented in clinical practice? Europace. 2024;26(8). [ DOI ] [ PMC free article ] [ PubMed ] 6. Schnabel RB, Sullivan LM, Levy D, Pencina MJ, Massaro JM, D’Agostino RB, Sr, et al. Development of a risk score for atrial fibrillation (Framingham Heart Study): a community-based cohort study. Lancet. 2009;373(9665):739–45. [ DOI ] [ PMC free article ] [ PubMed ] 7. Chamberlain AM, Agarwal SK, Folsom AR, Soliman EZ, Chambless LE, Crow R, et al. A clinical risk score for atrial fibrillation in a biracial prospective cohort (from the Atherosclerosis Risk in Communities [ARIC] study). Am J Cardiol. 2011;107(1):85–91. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Alonso A, Krijthe BP, Aspelund T, Stepas KA, Pencina MJ, Moser CB, et al. Simple risk model predicts incidence of atrial fibrillation in a racially and geographically diverse population: the CHARGE-AF consortium. J Am Heart Assoc. 2013;2(2):e000102. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Tseng AS, Noseworthy PA. Prediction of Atrial Fibrillation Using Machine Learning: A Review. Front Physiol. 2021;12:752317. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Hill NR, Ayoubkhani D, McEwan P, Sugrue DM, Farooqui U, Lister S, et al. Predicting atrial fibrillation in primary care using machine learning. PLoS ONE. 2019;14(11):e0224582. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Schnabel RB, Witt H, Walker J, Ludwig M, Geelhoed B, Kossack N, et al. Machine learning-based identification of risk-factor signatures for undiagnosed atrial fibrillation in primary prevention and post-stroke in clinical practice. Eur Heart J Qual Care Clin Outcomes. 2022;9(1):16–23. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Tiwari P, Colborn KL, Smith DE, Xing F, Ghosh D, Rosenberg MA. Assessment of a Machine Learning Model Applied to Harmonized Electronic Health Record Data for the Prediction of Incident Atrial Fibrillation. JAMA Netw Open. 2020;3(1):e1919396. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Martin SA, Johansson M, Heath I, Lehman R, Korownyk C. Sacrificing patient care for prevention: distortion of the role of general practice. BMJ. 2025;388:e080811. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Wändell P, Norrman A, Eriksson J, Ivarsson C, Gudjonsdottir H, Hagströmer M, et al. Differences between private and public primary health care centers and differences between men and women in antihypertensive care and cardiovascular prevention in all patients with hypertension treated in primary care in Stockholm County, Sweden. BMC Prim Care. 2025;26(1):20. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Andersson P, Löndahl M, Abdon NJ, Terent A. The prevalence of atrial fibrillation in a geographically well-defined population in northern Sweden: implications for anticoagulation prophylaxis. J Intern Med. 2012;272(2):170–6. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Wändell P, Carlsson AC, Wierzbicka M, Sigurdsson K, Ärnlöv J, Eriksson J, et al. A machine learning tool for identifying patients with newly diagnosed diabetes in primary care. Prim Care Diabetes. 2024;18(5):501–5. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Norrman A, Hasselström J, Ljunggren G, Wachtler C, Eriksson J, Kahan T, et al. Predicting new cases of hypertension in Swedish primary care with a machine learning tool. Prev Med Rep. 2024;44:102806. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Carlsson AC, Wändell P, Ösby U, Zarrinkoub R, Wettermark B, Ljunggren G. High prevalence of diagnosis of diabetes, depression, anxiety, hypertension, asthma and COPD in the total population of Stockholm, Sweden - a challenge for public health. BMC Public Health. 2013;13:670. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Wändell P, Ljunggren G, Carlsson AC. The most common diagnoses in primary care, and changes over time, in the total population of Stockholm, Sweden. BMC Prim Care. 2025;26(1):235. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Nemlander E, Ewing M, Abedi E, Hasselström J, Sjövall A, Carlsson AC, et al. A machine learning tool for identifying non-metastatic colorectal cancer in primary care. Eur J Cancer. 2023;182:100–6. [ DOI ] [ PubMed ] [ Google Scholar ] 22. Wallensten J, Wachtler C, Bogdanovic N, Olofsson A, Kivipelto M, Jönsson L, et al. Machine learning to detect Alzheimer’s disease with data on drugs and diagnoses. J Prev Alzheimers Dis. 2025;12(5):100115. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Team RC. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. 2016. http://www.R-project.org/ . 24. Lip GYH, Skjøth F, Nielsen PB, Larsen TB. Evaluation of the C(2)HEST Risk Score as a Possible Opportunistic Screening Tool for Incident Atrial Fibrillation in a Healthy Population (From a Nationwide Danish Cohort Study). Am J Cardiol. 2020;125(1):48–54. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Forslund T, Wettermark B, Wändell P, von Euler M, Hasselström J, Hjemdahl P. Risk scoring and thromboprophylactic treatment of patients with atrial fibrillation with and without access to primary healthcare data: experience from the Stockholm health care system. Int J Cardiol. 2013;170(2):208–14. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Ugowe FE, Jackson LR 2nd, Thomas KL. Racial and ethnic differences in the prevalence, management, and outcomes in patients with atrial fibrillation: A systematic review. Heart Rhythm. 2018;15(9):1337–45. [ DOI ] [ PubMed ] [ Google Scholar ] 27. Chung SC, Sofat R, Acosta-Mena D, Taylor JA, Lambiase PD, Casas JP, et al. Atrial fibrillation epidemiology, disparity and healthcare contacts: a population-wide study of 5.6 million individuals. Lancet Reg Health Eur. 2021;7:100157. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Zarrinkoub R, Wettermark B, Wändell P, Mejhert M, Szulkin R, Ljunggren G, et al. The epidemiology of heart failure, based on data for 2.1 million inhabitants in Sweden. Eur J Heart Fail. 2013;15(9):995–1002. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Shamsutdinova D, Das-Munshi J, Ashworth M, Roberts A, Stahl D. Predicting type 2 diabetes prevalence for people with severe mental illness in a multi-ethnic East London population. Int J Med Inf. 2023;172:105019. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Wändell P, Ljunggren G, Wahlström L, Carlsson AC. Diabetes and psychiatric illness in the total population of Stockholm. J Psychosom Res. 2014;77(3):169–73. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Forslund T, Carlsson AC, Ljunggren G, Ärnlöv J, Wachtler C. Patterns of multimorbidity and pharmacotherapy: a total population cross-sectional study. Fam Pract. 2021;38(2):132–40. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. van Royen FS, Asselbergs FW, Alfonso F, Vardas P, van Smeden M. Five critical quality criteria for artificial intelligence-based prediction models. Eur Heart J. 2023;44(46):4831–4. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Kemp Gudmundsdottir K, Svennberg E, Friberg L, Hygrell T, Frykman V, Al-Khalili F, et al. Randomized Invitation to Systematic NT-proBNP and ECG Screening in 75-Year-Olds to Detect Atrial Fibrillation: STROKESTOP II. Circulation. 2024;150(23):1837–46. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (154.5KB, docx) Supplementary Material 2 (27.1KB, docx) Data Availability Statement The data from this study can be accessed for research by qualified researchers who have been trained in confidentiality protocols for human subjects, following ethical approval from Region Stockholm at [email protected]. Analytical code and programs can be obtained from (mailto: [email protected]). Articles from BMC Medical Informatics and Decision Making are provided here courtesy of BMC ACTIONS View on publisher site PDF (973.5 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