Explainable machine learning for predicting mechanical ventilation in Alzheimer's patients with pneumonia: a SHAP-guided XGBoost nomogram based on MIMIC-IV.

Explainable machine learning for predicting mechanical ventilation in Alzheimer’s patients with pneumonia: a SHAP-guided XGBoost nomogram based on MIMIC-IV - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide 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 Sci Rep . 2026 Apr 13;16:17250. doi: 10.1038/s41598-026-48214-x Explainable machine learning for predicting mechanical ventilation in Alzheimer’s patients with pneumonia: a SHAP-guided XGBoost nomogram based on MIMIC-IV Xuhui Liu 1, # , Xujie Wang 2, # , Rongfei Xie 2 , Zhaohui Liu 3 , Biyun Wang 4, ✉ , Honglin Zhai 2, ✉ Author information Article notes Copyright and License information 1 Department of Neurology, The Second Hospital of Lanzhou University, 82 Cuiying Men, Chengguan District, Lanzhou, 730030 Gansu China 2 Department of Emergency ICU, The Affiliated Hospital of Qinghai University, 29Tongren Road, Chengxi District, Xining, 810001 Qinghai China 3 Clinical Medical College of Tianjin Medical University, 167 Dagang Xueyuan Road, Binhai New Area, Tianjin, 300270 China 4 Department of Biobank, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, 510120 China ✉ Corresponding author. # Contributed equally. Received 2025 Oct 18; Accepted 2026 Apr 7; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, 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 you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. 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-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13233829  PMID: 41974983 Abstract Alzheimer’s disease (AD) patients are particularly vulnerable to pneumonia and subsequent respiratory failure due to neurodegeneration-induced dysphagia and immunosenescence. Accurate prediction of mechanical ventilation (MV) need in this population remains a critical but unmet clinical challenge. This retrospective cohort study utilized the MIMIC-IV database to identify 793 critically ill AD patients diagnosed with pneumonia. Clinical variables were extracted and filtered using LASSO regression and multivariate logistic regression to identify key predictors. Five machine learning algorithms, including XGBoost, were constructed and compared to evaluate MV prediction performance. Model interpretability was enhanced using SHAP analysis, and a nomogram was developed based on the best-performing model. Among the included patients, 213 (26.9%) required mechanical ventilation during ICU admission. The XGBoost model outperformed all other algorithms, achieving the highest AUC in both the training (0.849) and validation (0.744) sets, as well as superior F1-score (0.507) and balanced accuracy (0.776). SHAP analysis identified COPD, temperature, and gender as the most influential predictors. The final nomogram incorporating seven key variables (Gender, AKI, COPD, OASIS, Temperature, PLT, and antibiotic use) achieved robust discrimination ( validation Set AUC = 0.780), excellent calibration, and consistent clinical benefit in decision curve analysis. The derived NomoScore stratified patients into quartiles with escalating MV risk ( p < 0.001), and retained predictive accuracy with Cohen’s kappa values of 0.40 (training) and 0.363 (validation). The model demonstrated superior net benefit over conventional scoring systems and has the potential to support bedside decision-making in cognitively impaired ICU populations. This study presents a validated, interpretable XGBoost-based model and clinical nomogram for predicting mechanical ventilation risk in AD patients with pneumonia. The framework offers a precision tool to aid ICU triage and guide ventilatory strategies. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-48214-x. Keywords: Alzheimer’s disease, Pneumonia, Mechanical ventilation, Machine learning, XGBoost, Nomogram, SHAP, MIMIC-IV Subject terms: Computational biology and bioinformatics, Diseases, Health care, Medical research, Neurology, Risk factors Introduction Alzheimer’s disease (AD), the most prevalent neurodegenerative disorder worldwide, has seen a rising incidence paralleling global population aging. According to the World Health Organization, approximately 55 million individuals currently live with AD, a figure projected to reach 139 million by 2050 1 . This population faces heightened susceptibility to infectious diseases such as pneumonia due to immunosenescence, dysphagia, and prolonged immobilization, and dysphagia, which specifically drives the aspiration of colonized oropharyngeal material—leading to aspiration pneumonia rather than transient pneumonitis 2 . Studies demonstrate that AD patients exhibit a 2- to threefold increased incidence of pneumonia compared to age-matched non-AD cohorts, with 30% progressing to critical illness requiring mechanical ventilation (MV) due to impaired neuroregulatory mechanisms 3 – 5 . Although MV sustains vital functions, it carries substantial risks of delirium, ventilator-associated pneumonia, and in-hospital mortality rates as high as 45% in AD populations 5 – 7 . These clinical challenges underscore the urgent need for precision prediction of ventilation requirements. Current clinical decision-making for MV in severe pneumonia relies heavily on conventional scoring systems such as the Sequential Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Evaluation II (APACHE-II). However, these tools demonstrate critical limitations in AD populations: First, they fail to incorporate AD-specific neurocognitive deficits, despite evidence that frontal lobe atrophy-induced cough reflex impairment independently exacerbates pneumonia severity 8 . This impairment disrupts the high-speed exhalatory clearing of secretions from airway walls, a vital protective mechanism whose failure significantly accelerates the progression toward mechanical ventilation 9 .Second, baseline physiological parameters (e.g., blood pressure, oxygenation indices) in dementia patients often deviate from normative ranges, introducing systemic bias into traditional scores 10 . Notably, the integration of the Glasgow Coma Scale (GCS) in these systems often confounds chronic cognitive deficits with acute neurological failure, further compromising their diagnostic precision in the AD population. Furthermore, APACHE-II exhibits declining predictive accuracy for long-term outcomes (e.g., 90-day mortality) with advancing age, yielding suboptimal AUC values of 0.62–0.68 in AD cohorts 11 . Consequently, clinicians face therapeutic dilemmas—premature extubation risks respiratory failure, while prolonged ventilation may accelerate cognitive decline 12 , 13 . Advances in machine learning (ML) offer transformative solutions to these challenges. Compared to conventional logistic regression, ensemble algorithms (e.g., XGBoost, LightGBM) excel at modeling high-dimensional nonlinear interactions, as evidenced by their superior performance in sepsis prediction 14 . For AD patients, MV risk likely arises from multimodal factor interplay: β-amyloid deposition may amplify pulmonary inflammation via microglial activation 15 , 16 , while cholinesterase inhibitors modulate autonomic tone 15 . ML models can leverage feature engineering to decode such biomarker-clinical variable synergies, enabling robust prediction frameworks. Notably, interpretability algorithms like SHAP (Shapley Additive Explanations) illuminate decision pathways of "black-box" models, identifying critical predictors and threshold effects 17 . This breakthrough bridges the gap between algorithmic complexity and clinical interpretability. For translational implementation, nomograms emerge as ideal tools for bedside risk stratification. These visual scoring systems successfully translate complex models into rapid decision aids in fields like breast oncology 18 . Yet no validated nomogram exists for AD-associated pneumonia. Current research predominantly focuses on unidimensional biomarkers (e.g., serum IL-6, cerebrospinal tau protein) 19 – 21 , neglecting integrative analysis of multidimensional data streams. Leveraging critical care databases (e.g., MIMIC-IV) to develop ML-driven nomograms could address this gap. Methods Study population This retrospective cohort study was conducted using data from the MIMIC-IV (version 3.1) database, a large, publicly accessible critical care dataset maintained by the Massachusetts Institute of Technology. The database includes de-identified clinical records from ICU admissions at Beth Israel Deaconess Medical Center between 2008 and 2019. One author (Xujie Wang) obtained certified access (Certification ID: 13,931,651) and performed all data extraction. Patients with a diagnosis of Alzheimer’s disease (AD) were identified using ICD-9 and ICD-10 codes (331.0, G30.0, G30.1, G30.8, G30.9). From an initial cohort of 875 AD patients, we excluded those younger than 60 years at ICU admission (n = 15), those with multiple ICU stays (n = 58), and those with ICU lengths of stay less than 6 h (n = 9). After exclusions, 793 patients were included for analysis as shown in (Fig. 1 ).This final cohort was randomly divided into training (70%) and validation (30%) subsets. Subsequent steps—including variable selection, predictive modeling using multiple machine learning algorithms, and nomogram construction—were conducted as outlined in the study workflow (Fig. 1 ), and detailed in the Statistical Analysis section. Fig. 1. Open in a new tab Study Flowchart. Data collection All clinical data were extracted from the MIMIC-IV (version 3.1) database using PostgreSQL (version 13.7.2) and Navicat Premium (version 16), with structured query language (SQL) employed to access and filter patient-level variables. Data extraction was performed by an author (Xujie Wang), who had certified access to the database. Variables were categorized into four domains for analysis: (1) Demographics, including age, sex, and weight; (2) Comorbidities, such as acute kidney injury (AKI), chronic obstructive pulmonary disease (COPD), and others based on ICD-coded diagnoses; (3) Clinical and laboratory parameters, including Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), and Mean Blood Pressure (MBP);body temperature, platelet count (PLT), ABG-derived oxygenation indices, and antibiotic metrics—specifically Abx (a binary indicator of any administration) and Abxtotalval (cumulative dose in mg)—captured within the initial 24 h of ICU admission to serve as high-fidelity proxies for baseline clinical suspicion ; white blood cell count (WBC), hemoglobin (HB), glucose, and serum electrolytes; and (4) Severity scores, including the Sequential Organ Failure Assessment (SOFA) and Oxford Acute Severity of Illness Score (OASIS), also computed within the first 24 h. Categorical variables were binary-encoded for model input; specifically, gender was encoded as 1 for male and 0 for female. Variables with more than 20% missingness were excluded to minimize imputation bias. For variables with ≤ 20% missingness, multiple imputation was conducted using chained equations with a random forest algorithm, implemented via the “mice” package in R (Supplementary Table S1 ) 27 , 28 . Outcome definition The primary outcome of this study was the occurrence of invasive mechanical ventilation ( IMV ) during ICU stay in patients with Alzheimer’s disease complicated by pneumonia. This endpoint was selected because MV serves as a decisive clinical juncture that significantly influences the long-term cognitive trajectory and survival of AD patients. Pneumonia diagnoses were identified using a comprehensive list of ICD-9/10 codes validated in previous research, including: 041.2, 486, J18.9, 507.0, J69.0, B96.1, 482.9, V03.82, J15.9, Z87.01, 482.1, 482.42, 482.41, V06.6, 487.0, 482.83, 516.8, J18.8, J15.211, 481, J15.1, J15.6, 482.0, J15.212, 484.6, J13, J18.1, and V12.61. Mechanical ventilation status was determined using procedural and event records from both the chartevents and procedureevents tables in MIMIC-IV, supplemented by ICD-9/10 procedural codes including 96.70, 96.71, and 96.72. These criteria ensured high specificity in identifying ventilated patients and allowed accurate modeling of risk prediction. To maintain temporal precedence, variables for early-ventilated cases were restricted strictly to pre-IMV records. Statistical analysis All statistical analyses were conducted using R software (version 4.2.2) and Python (version 3.10.9). Descriptive statistics were used to summarize baseline characteristics. Continuous variables were presented as medians with interquartile ranges (IQRs) and compared using the Wilcoxon rank-sum test. Categorical variables were expressed as frequencies and percentages and compared using the chi-squared or Fisher’s exact test as appropriate. To ensure balanced model training and generalizability, the cohort was randomly split into a training set (70%) and a Validation Set (30%). To address high-dimensionality and multicollinearity among candidate variables, we first applied least absolute shrinkage and selection operator (LASSO) regression using tenfold cross-validation on the training set. The optimal penalization parameter (λ) was selected based on the minimum binomial deviance (λ < sub > min < /sub >). Variables with non-zero coefficients at λ < sub > min < /sub > were retained for further modeling. Subsequently, backward stepwise multivariate logistic regression was performed based on the Akaike Information Criterion (AIC) to further refine the model. Odds ratios (ORs) and 95% confidence intervals (CIs) were reported for each retained predictor. Multicollinearity was assessed using variance inflation factor (VIF), with a VIF > 5 considered indicative of collinearity. We then constructed five supervised machine learning models using the selected predictors: logistic regression (LR), support vector machine (SVM), decision tree (DT), extreme gradient boosting (XGBoost), and light gradient boosting machine (LGBM). Model performance was evaluated in both the training and Validation Sets using multiple metrics, including area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, recall, F1-score, and balanced accuracy. Calibration curves were plotted to assess agreement between predicted and observed outcomes, and decision curve analysis (DCA) was performed to evaluate net clinical benefit across threshold probabilities. Brier scores were calculated to evaluate model calibration and probability estimation accuracy. To enhance model transparency and clinical interpretability, SHAP (Shapley Additive Explanations) was used to quantify the contribution of each feature to the XGBoost model’s predictions. Global interpretability was visualized through gain-based feature importance plots, SHAP summary plots, and SHAP dependence plots. Individualized risk explanations were generated via SHAP waterfall and force plots for representative high- and low-risk cases. Based on the final XGBoost model developed on the training set, a nomogram was constructed using the training cohort (n = 556) to provide an intuitive graphical tool for individualized risk prediction. Model-derived point totals were converted into estimated probabilities of mechanical ventilation. The nomogram’s performance was assessed using AUC, accuracy, sensitivity, specificity, and DCA in both the training and Validation Sets. Furthermore, a composite Nomogram Score (NomoScore) was calculated and stratified into quartiles for risk classification. Associations between NomoScore quartiles and ventilation risk were visualized using forest plots and boxplots. Diagnostic performance of the NomoScore system was further evaluated using AUC and Cohen’s kappa statistics. All statistical tests were two-sided, and P -values < 0.05 were considered statistically significant. Results A total of 793 patients diagnosed with Alzheimer’s disease (AD) and concomitant pneumonia were included in the present study. Based on the occurrence of mechanical ventilation during hospitalization, patients were divided into two groups: 580 (73.1%) did not undergo mechanical ventilation, while 213 (26.9%) required ventilatory support. As shown in Supplementary Table 1 , the median age of the cohort was 84.0 years [IQR: 79.0–89.0], with no significant age difference between groups ( p = 0.879). Similarly, gender distribution, body weight, blood pressure, and common laboratory indices, including white blood cell count and hemoglobin levels, did not differ significantly between the two groups (all P > 0.05; see Table 1 for details). However, several clinical variables exhibited statistically significant differences between the two groups. Patients in the ventilated group had higher rates of acute kidney injury (AKI) ( p < 0.001), cerebral bleeding (CB) ( p = 0.004), heart failure (HF) ( p = 0.018), and chronic obstructive pulmonary disease (COPD) ( p < 0.001). Among the 213 patients who required ventilatory support, the median duration of invasive mechanical ventilation was 31.3 h (IQR: 17.0–65.0), reflecting the high clinical severity and the substantial respiratory burden in this cognitively impaired population. These differences highlight the potential value of these variables in predicting the need for mechanical ventilation in this vulnerable population. For model development and validation, the dataset was randomly split into a training cohort (n = 556, 70%) and a Validation cohort (n = 237, 30%). As summarized in Table 1 , the age and gender distributions were balanced between the training and Validation Sets ( p = 0.163 and p = 1.000, respectively). The incidence of mechanical ventilation was also comparable across both subsets (26.4% vs. 26.6%, p = 0.771), supporting the appropriateness of the data partition for subsequent model training and performance evaluation. Table 1. Baseline Characteristics of the Training and Validation Sets. [ALL] N = 793 Validation Validation N = 237 Train N = 556 p.overall Clinical Grouping 0.27 (0.44) 0.26 (0.44) 0.27 (0.45) 0.771 Age 84.00 [79.00;89.00] 85.00 [80.00;90.00] 84.00 [79.00;89.00] 0.163 Gender 1.000 Female 447 (56.37%) 134 (56.54%) 313 (56.29%) Male 346 (43.63%) 103 (43.46%) 243 (43.71%) Weight 67.70 [58.20;79.90] 67.70 [57.00;80.00] 67.70 [58.68;79.75] 0.851 HTN 0.604 No 401 (50.57%) 116 (48.95%) 285 (51.26%) Yes 392 (49.43%) 121 (51.05%) 271 (48.74%) AKI 0.561 No 489 (61.66%) 142 (59.92%) 347 (62.41%) Yes 304 (38.34%) 95 (40.08%) 209 (37.59%) CVA 0.674 No 674 (84.99%) 199 (83.97%) 475 (85.43%) Yes 119 (15.01%) 38 (16.03%) 81 (14.57%) CKD 0.093 No 633 (79.82%) 180 (75.95%) 453 (81.47%) Yes 160 (20.18%) 57 (24.05%) 103 (18.53%) T2DM 0.657 No 579 (73.01%) 170 (71.73%) 409 (73.56%) Yes 214 (26.99%) 67 (28.27%) 147 (26.44%) CB: 1.000 No 745 (93.95%) 223 (94.09%) 522 (93.88%) Yes 48 (6.05%) 14 (5.91%) 34 (6.12%) HF: 0.111 No 600 (75.66%) 170 (71.73%) 430 (77.34%) Yes 193 (24.34%) 67 (28.27%) 126 (22.66%) MI: 0.911 No 739 (93.19%) 220 (92.83%) 519 (93.35%) Yes 54 (6.81%) 17 (7.17%) 37 (6.65%) IHD 0.075 No 549 (69.23%) 153 (64.56%) 396 (71.22%) Yes 244 (30.77%) 84 (35.44%) 160 (28.78%) COPD 0.564 No 699 (88.15%) 206 (86.92%) 493 (88.67%) Yes 94 (11.85%) 31 (13.08%) 63 (11.33%) Ventilation_Hour 12.00 [0.00;31.35] 12.68 [0.00;33.02] 11.10 [0.00;30.98] 0.529 SOFA 4.00 [3.00;6.00] 4.00 [3.00;6.00] 4.00 [3.00;6.00] 0.942 APSIII 46.00 [36.00;61.00] 46.00 [36.00;62.00] 46.00 [36.00;61.00] 0.772 SIRS 2.00 [2.00;3.00] 2.00 [2.00;3.00] 2.00 [2.00;3.00] 0.951 SAPSII 40.00 [34.00;49.00] 40.00 [34.00;50.00] 41.00 [35.00;49.00] 0.806 OASIS 35.00 [30.00;40.00] 34.00 [30.00;39.00] 35.00 [30.00;40.00] 0.333 GCS 13.00 [10.00;14.00] 14.00 [11.00;14.00] 13.00 [10.00;14.00] 0.187 Charlson 6.00 [5.00;8.00] 7.00 [6.00;8.00] 6.00 [5.00;8.00] 0.002 HR(bpm) 86.00 [71.00;99.00] 87.00 [72.00;99.00] 85.00 [71.00;99.25] 0.456 SBP (mmHg) 126.00 [109.00;145.00] 125.00 [109.00;142.00] 127.00 [110.00;145.25] 0.779 DBP (mmHg) 67.00 [56.00;82.00] 69.00 [57.00;85.00] 67.00 [55.00;81.00] 0.099 MBP (mmHg) 83.00 [70.00;97.00] 83.00 [71.00;98.00] 83.00 [69.75;96.00] 0.269 RR (breaths/min) 19.00 [16.00;23.00] 19.00 [16.00;24.00] 19.00 [16.00;23.00] 0.377 Spo2(%) 97.00 [95.00;100.00] 97.00 [95.00;100.00] 97.00 [95.00;100.00] 0.686 Temperature(°C) 98.10 [97.50;98.70] 98.10 [97.50;98.70] 98.10 [97.60;98.70] 0.939 Hematocrit(%) 33.70 [29.70;37.20] 33.70 [29.20;37.30] 33.70 [29.80;37.00] 0.812 HB(g/dL) 10.90 [9.60;12.20] 11.00 [9.60;12.10] 10.90 [9.70;12.22] 0.653 PLT(× 10⁹/L) 192.00 [150.00;246.00] 192.00 [150.00;251.00] 192.00 [149.75;243.00] 0.688 RDW(%) 14.40 [13.50;15.50] 14.40 [13.60;15.70] 14.40 [13.50;15.30] 0.413 RBC(× 10 12 /L) 3.65 [3.25;4.09] 3.66 [3.32;4.08] 3.65 [3.23;4.09] 0.478 WBC(× 10⁹/L) 10.30 [8.00;13.80] 10.30 [8.30;14.10] 10.30 [7.90;13.50] 0.262 Anion_gap (mmol/L) 14.00 [12.00;16.00] 14.00 [12.00;16.00] 14.00 [12.00;16.00] 0.257 Calcium_Total (mmol/L) 8.40 [8.00;8.90] 8.50 [8.00;9.00] 8.40 [8.00;8.90] 0.130 Chloride(mmol/L) 106.00 [103.00;110.00] 106.00 [103.00;110.00] 106.00 [103.00;110.00] 0.875 Glucose(mg/dL) 127.00 [105.00;159.00] 127.00 [108.00;158.00] 127.00 [105.00;160.00] 0.712 Potassium (mmol/L) 4.00 [3.70;4.50] 4.00 [3.70;4.50] 4.00 [3.70;4.50] 0.359 Sodium (mmol/L) 140.00 [138.00;143.00] 140.00 [138.00;143.00] 140.00 [138.00;143.00] 0.809 Creatinine (mg/dL) 1.00 [0.80;1.40] 1.00 [0.80;1.50] 1.00 [0.80;1.40] 0.547 Urea_Nitrogen (mg/dL) 22.00 [16.00;35.00] 22.00 [16.00;35.00] 22.00 [15.75;36.00] 0.479 Abx: 0.510 No 198 (24.97%) 55 (23.21%) 143 (25.72%) Yes 595 (75.03%) 182 (76.79%) 413 (74.28%) Abxtotalval (mg) 1004.50 [1.00;2771.00] 1002.25 [1.00;3000.00] 1005.25 [0.00;2756.69] 0.777 Open in a new tab SOFA, Sequential Organ Failure Assessment; APSIII, Acute Physiology Score III; SIRS, Systemic Inflammatory Response Syndrome; SAPSII, Simplified Acute Physiology Score II; OASIS, Oxford Acute Severity of Illness Score; GCS, Glasgow Coma Scale; HR, heart rate;SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure; RR, respiratory rate (breaths/min); SpO₂, oxygen saturation; Temperature, body temperature; Hematocrit, red cell volume percentage; HB, hemoglobin; PLT, platelet count; RDW, red cell distribution width; RBC, red blood cell count; WBC, white blood cell count; Anion_gap, serum anion gap; Calcium_Total, total serum calcium; Chloride, serum chloride; Glucose, serum glucose; Potassium, serum potassium; Sodium, serum sodium; Creatinine, serum creatinine; Urea_Nitrogen, blood urea nitrogen; Abxtotalval, total antibiotic exposure. Variable selection using LASSO regression To identify the most predictive variables for mechanical ventilation in patients with Alzheimer’s disease (AD) and pneumonia, we applied least absolute shrinkage and selection operator (LASSO) regression with tenfold cross-validation. The optimal penalty parameter was determined at the value of λ that minimized binomial deviance (λ min = 0.0215). As shown in Fig. 2 A, the trajectories of standardized coefficients revealed that most variables were shrunk toward zero with increasing penalization. Figure 2 B illustrates the cross-validation curve, where λ min was selected based on the minimum mean deviance. At this threshold, a total of 11 variables with non-zero coefficients were retained: Gender, AKI, HF, COPD, SOFA, OASIS, HR, Temperature, PLT, and Abx, in addition to the model intercept. These variables were considered most predictive and were subsequently used in model construction. For transparency, the complete variable list and their corresponding coefficients at λ < sub > min < /sub > are detailed in Supplementary Table 2 . Additional visualizations of coefficient magnitude and variable selection are presented in Supplementary Fig. 2 A–B, which further confirm the relative contributions of each feature and the robustness of the λ min -based selection process. Fig. 2. Open in a new tab LASSO-Based Variable Selection ( A ) Coefficient profiles of all candidate variables across log(λ). ( B ) Cross-validation curve showing binomial deviance; the dashed line marks λ min used for selection. Multivariate logistic regression analysis To further refine the predictive model, we applied backward stepwise logistic regression based on the Akaike Information Criterion (AIC), using the variables retained from the LASSO selection. The final model included seven predictors: Gender, AKI, COPD, OASIS, Temperature, PLT, and Abx, as shown in Table 2 . Among them, Gender (OR = 1.654, 95% CI: 1.082–2.538, p = 0.021), OASIS (OR = 1.070, 95% CI: 1.041–1.101, p < 0.001), Temperature (OR = 1.118, 95% CI: 1.005–1.245, p = 0.041), PLT (OR = 1.002, 95% CI: 1.000–1.004, p = 0.042), and Abx (OR = 36.749, 95% CI: 8.027–168.280, p < 0.001) were independently associated with the risk of mechanical ventilation. Notably, Temperature ( p = 0.178) and PLT ( p = 0.108) were retained in the final model because they were identified as critical contributors during the AIC-based backward stepwise selection process, ensuring an optimal balance between model complexity and predictive fit . The stepwise process and goodness-of-fit are visualized in Supplementary Fig. 3 , while multicollinearity was evaluated using variance inflation factor (VIF). As presented in Supplementary Table 3 , all VIF values were below 1.1, indicating negligible multicollinearity among included predictors. The final model achieved an AIC of 533.6 and BIC of 568.2, supporting optimal balance between model complexity and fit. Table 2. Multivariate Logistic Regression Results After Backward Stepwise Selection. characteristics B SE OR CI Z P (Intercept) − 18.462 8.20322 0 9.596 (6.849–0.001) − 2.251 < 0.05 Gender 0.503 0.21726 1.654 1.654 (1.082–2.538) 2.316 < 0.05 AKI 0.408 0.21905 1.504 1.503 (0.978–2.312) 1.863 0.062 COPD 0.504 0.3129 1.655 1.654 (0.889–3.050) 1.61 0.107 OASIS 0.068 0.0143 1.07 1.070 (1.041–1.101) 4.742 < 0.001 Temperature 0.112 0.08294 1.118 1.118 (0.994–1.320) 1.347 0.178 PLT 0.002 0.00112 1.002 1.001 (0.999–1.004) 1.607 0.108 Abx 3.603 0.72924 36.701 36.70 (11.18–226.8) 4.94 < 0.001 Open in a new tab Machine learning model construction and evaluation To optimize predictive performance for mechanical ventilation among AD patients with pneumonia, we developed and compared five supervised machine learning algorithms: logistic regression (LR), decision tree (DT), support vector machine (SVM), extreme gradient boosting (XGBoost), and light gradient boosting machine (LGBM). These models were constructed using variables identified through LASSO regression and backward stepwise logistic regression. As shown in Fig. 3 A–B, receiver operating characteristic (ROC) curves demonstrated that XGBoost consistently achieved the highest discriminatory ability in both the training set (AUC = 0.849) and the Validation Set (AUC = 0.744), followed by LR and SVM. Calibration curves (Fig. 3 C–D) showed excellent concordance between predicted and observed probabilities, particularly for the XGBoost and LR models. Decision curve analysis (DCA, Fig. 3 E–F) further confirmed the superior net clinical benefit of XGBoost and LR across a wide range of threshold probabilities, highlighting their practical value in individualized decision-making. Comprehensive model performance metrics—including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, F1-score, and balanced accuracy—are presented in Table 3 . Among all algorithms, XGBoost yielded the highest F1-score (0.507) and balanced accuracy (0.776) in the validation cohort, indicating superior classification performance across multiple evaluation domains. To further assess the calibration and reliability of predicted probabilities, we calculated Brier scores for each model. As summarized in Supplementary Table 4 , XGBoost and DT demonstrated the lowest Brier scores in both the training and Validation Sets, reflecting the most accurate probability estimates with minimal prediction error. Visual comparisons of performance metrics across models are presented in Supplementary Fig. 5 A–D. Panels A and B display heatmaps of seven key evaluation metrics—sensitivity, specificity, PPV, NPV, precision, recall, and F1-score—for the training and Validation Sets, respectively. These visualizations clearly show that XGBoost and LR outperformed other models across most metrics. Panels C and D further depict bar plots comparing AUC values across all models, reaffirming the consistently superior discriminatory capacity of XGBoost. Collectively, these visualizations provide an intuitive, multidimensional confirmation of the statistical findings. Fig. 3. Open in a new tab Discrimination, Calibration, and Clinical Utility of Machine Learning Models ( A – B ) ROC curves for the training and Validation Sets across all models. ( C – D ) Calibration curves showing agreement between predicted and observed risk. ( E – F ) Decision curve analysis (DCA) for net clinical benefit assessment. Table 3. Comparative Performance Metrics of Machine Learning Models in the Training and validation Cohorts. model Dataset (Development/Validation) Sensitivity Specificity Pos Pred Value Neg Pred Value Precision Recall F1 LR Training Set 0.748 0.716 0.496 0.884 0.496 0.748 0.596 DT Training Set 0.753 0.815 0.424 0.948 0.424 0.753 0.542 XGB Training Set 0.762 0.79 0.575 0.899 0.575 0.762 0.655 SVM Training Set 0.278 0.269 0.124 0.5 0.124 0.278 0.172 LGBM Training Set 0.609 0.768 0.495 0.841 0.495 0.609 0.546 LR Validation Set 0.71 0.714 0.468 0.874 0.468 0.71 0.564 DT Validation Set 0.448 0.764 0.21 0.909 0.21 0.448 0.286 XGB Validation Set 0.581 0.749 0.45 0.834 0.45 0.581 0.507 SVM Validation Set 0.355 0.257 0.145 0.529 0.145 0.355 0.206 LGBM Validation Set 0.452 0.771 0.412 0.799 0.412 0.452 0.431 Open in a new tab Model interpretation and individualized explanation To enhance interpretability of the XGBoost model, we conducted a comprehensive SHAP-based (Shapley Additive Explanations) analysis. As shown in Fig. 4 A, feature importance was quantified using the “gain” metric, revealing that COPD, Temperature, and Gender were the top three contributors to model performance. To visualize the distribution and directionality of feature effects, a SHAP beeswarm plot (Fig. 4 B) was generated. This plot highlights both the magnitude and direction of SHAP values across the entire cohort. For example, high values of Temperature and PLT tended to decrease predicted risk, whereas the presence of COPD, AKI, and Antibiotic use (Abx) were strongly associated with increased risk of mechanical ventilation. Model classification performance was further illustrated using confusion matrices. Figure 4 C presents the confusion matrix for the training set, indicating good separation between outcome categories with minimal misclassification. The corresponding matrix in the Validation Set (Fig. 4 D) also confirmed strong generalizability, with balanced predictive accuracy across both classes. To facilitate clinical application and individualized decision support, SHAP-based explanations were constructed for selected patients, as shown in Fig. 5 . In Fig. 5 A–B, A representative individual is visualized using a SHAP waterfall plot and force plot, respectively, demonstrating how cumulative feature contributions lead from the expected model output (base value) to the final prediction. For this individual, factors such as Abx, COPD, and elevated OASIS scores had the strongest positive SHAP contributions, pushing the risk higher. In addition, Fig. 5 C–D depicts explanation plots for a contrasting patient with lower predicted risk. Here, features such as low Temperature and female sex pulled the risk downward, while comorbidities such as COPD and AKI partially offset this protection. These visualizations illustrate how the same model accommodates individualized feature contributions and clinical heterogeneity. Finally, global SHAP dependence plots for each variable are shown in Supplementary Fig. 6 . These plots provide insight into the nonlinear relationships between feature values and predicted risk. For instance, PLT exhibited a threshold effect with lower counts substantially increasing SHAP values, while OASIS showed a near-linear association with rising risk. Categorical features such as Gender, COPD, and AKI demonstrated consistent directionality, further reinforcing their interpretive clarity and clinical plausibility. Fig. 4. Open in a new tab Global Interpretability of the XGBoost Model ( A ) Variable importance ranked by gain. ( B ) SHAP beeswarm plot displaying the distribution and direction of feature effects. ( C ) Confusion matrix for the training set. ( D ) Confusion matrix for the Validation Set. Fig. 5. Open in a new tab SHAP-Based Individual Risk Explanations ( A ) Waterfall plot for a representative high-risk patient. ( B ) Force plot for the same patient. ( C ) Waterfall plot for a low-risk patient. ( D ) Force plot for the low-risk patient. Nomogram construction and model evaluation To facilitate clinical deployment of the XGBoost model, we developed an integrated nomogram based on the seven key predictors: Gender, AKI, COPD, OASIS, Temperature, PLT, and Antibiotic use (Abx). The resulting nomogram is presented in Fig. 6 , offering an intuitive graphical interface to estimate the probability of mechanical ventilation based on individual patient characteristics. The performance of the nomogram was evaluated using both discrimination and clinical utility metrics. As shown in Supplementary Fig. 7 A–B, ROC curves for the training and Validation Sets demonstrated good discrimination, with AUCs of 0.796 and 0.780, respectively. Decision curve analysis (DCA, Supplementary Fig. 7 C–D) indicated a higher net clinical benefit for the nomogram compared to "treat-all" or "treat-none" strategies across a range of threshold probabilities. Evaluation metrics including AUC, accuracy (ACC), sensitivity, specificity, and predictive values are summarized in Supplementary Table 5 , confirming consistent performance across datasets. To assess the predictive value of the Nomogram Score (NomoScore) derived from the point-based nomogram system, we further stratified patients into four risk quartiles. As shown in Fig. 7 , higher quartiles of NomoScore were strongly associated with increased odds of mechanical ventilation, with a dose–response pattern of risk escalation across quartiles. Box plots comparing NomoScores between outcome groups are provided in Supplementary Fig. 8 A–B, illustrating significant score separation in both training and validation cohorts. Predictive performance metrics of the NomoScore-based classification system are listed in Supplementary Table 6 , showing high calibration and diagnostic reliability with AUCs of 0.796 (training) and 0.712 (validation), and kappa coefficients of 0.40 and 0.363, respectively. Fig. 6. Open in a new tab Clinical Nomogram for Predicting the Risk of Mechanical Ventilation. Fig. 7. Open in a new tab Forest Plot of Mechanical Ventilation Risk According to NomoScore Quartiles. Discussion In this study, we developed and internally validated a robust and interpretable machine learning model to predict the need for mechanical ventilation among critically ill patients with Alzheimer’s disease (AD) complicated by pneumonia. Among the five supervised learning algorithms evaluated, the XGBoost model demonstrated significantly superior and most consistent predictive performance, with an AUC of 0.849 in the training set and 0.744 in the Validation Set, alongside superior F1-score, calibration, and net clinical benefit across a range of thresholds as assessed by decision curve analysis (DCA). SHAP (Shapley Additive Explanations) analysis further revealed that clinical variables such as COPD, OASIS score, body temperature, and platelet count were the most influential contributors to the model’s output. Importantly, we translated the final model into a user-friendly nomogram based on the training cohort data, enabling individualized risk estimation at the bedside. XGBoost’s superiority over traditional scores underscores its capacity to integrate non-linear AD-specific stressors that generic, linear metrics frequently overlook. The model’s performance remained robust in both internal validation and subgroup analysis, supporting its potential clinical applicability in ICU triage and early intervention planning for vulnerable older adults. Although not directly measured, the identified predictors likely exhibit AD-specific pathophysiological significance through potential multifaceted mechanisms, serving as macro-level surrogates for the following biological pathways . Specifically, the high predictive weight of respiratory rate (RR) reflects the unsustainable increase in the work of breathing (WoB) necessitated by reduced lung compliance, signaling an impending mechanical failure of the ventilatory pump in this elderly cohort. Although Diabetes Mellitus is known to increase infection susceptibility 22 , its predictive weight for IMV in our oldest-old cohort ( p = 0.657) was overshadowed by acute physiological insults such as AKI and COPD. Chronic obstructive pulmonary disease (COPD) comorbidity amplifies mechanical ventilation (MV) risk via hypoxia-inducible factor 1α (HIF-1α)-mediated β-secretase upregulation, which concurrently accelerates amyloid-β deposition in pulmonary macrophages and hippocampal neurons—a dual pathology confirmed by bronchoalveolar lavage studies showing 4.7-fold higher Aβ42 levels in AD-COPD patients compared to AD-only controls 23 – 25 . This lung-brain axis dysfunction is further exacerbated by matrix metalloproteinase-9 (MMP-9) activation during COPD exacerbations, which proteolytically cleaves both lung elastin and neuronal tau proteins, creating a self-perpetuating cycle of shunt-induced respiratory failure and cognitive deterioration 26 – 28 , exacerbated by age-related declines in respiratory mechanics and gas-exchange surface area, which culminates in systemic tissue dysoxia and multi-organ strain. Acute kidney injury (AKI) demonstrates disproportionate impact in AD populations through neuro-renal crosstalk mechanisms. Uremic toxins like indoxyl sulfate penetrate the compromised blood–brain barrier, activating microglial NLRP3 inflammasomes that potentiate neuroinflammation and ventilator dyssynchrony 29 . Multimodal imaging studies reveal synchronized amyloid-β deposition in renal tubules and default mode network regions, suggesting shared clearance dysfunction along the neuro-renal axis 30 . Clinically, AKI-stage III prolongs MV duration in AD patients through uremia-induced cholinergic dysfunction, a phenomenon less pronounced in non-AD cohorts 31 , 32 .The U-shaped relationship between Oxygenation, Age, and Survival in Severe Pneumonia (OASIS) scores and MV risk reflects AD’s unique neurohomeostatic disruption. Over-resuscitation (OASIS > 40) induces cerebral hyperperfusion, exacerbating blood–brain barrier leakage as quantified by cerebrospinal fluid/serum albumin ratios exceeding 15 × 10 3 32 . Conversely, under-resuscitation (OASIS < 20) elevates neurotoxic kynurenine metabolites 4.1-fold, directly impairing brainstem respiratory drive—a metabolic derangement detectable through transcranial Doppler monitoring of middle cerebral artery flow velocities 33 – 35 .Platelet dynamics emerge as critical predictors through neurovascular coupling mechanisms. Thrombocytopenia (< 100 × 10 3 /μL) triggers AD-specific vascular crises via platelet-derived serotonin deficiency, which impairs blood–brain barrier repair through 5-hydroxytryptamine receptor 2B (5-HT2B) downregulation 36 . Proteomic analyses identify δ-secretase activation in AD platelets, simultaneously cleaving amyloid precursor protein (APP) to generate Aβ42 and promoting fibrinogen-driven microthrombosis—a dual pathology ameliorated in murine models through platelet transfusion protocols that restore hippocampal oxygenation 37 , 38 .Thermoregulatory dysfunction manifests as temperature variability (ΔT > 1.5 °C/24 h), reflecting AD-related hypothalamic neurodegeneration. Consequently, our model leverages the absence or attenuation of a febrile response—rather than fever elevation—as a key indicator of physiological exhaustion and impending respiratory failure. This makes temperature a critical proxy for the patient’s neurological resilience and the integrity of the lung-brain feedback loop during acute infection. Postmortem analyses link MV-associated febrile episodes to accelerated tau propagation in thermoregulatory nuclei, while targeted temperature management trials demonstrate hypothermia-mediated stabilization of brainstem serotonergic neurons, reducing MV duration by 41% 39 , 40 . Furthermore, the blunted febrile response in older AD patients can lead to significant diagnostic delays, masking the severity of infection until physiological exhaustion necessitates ventilatory support 41 . Notably, the high odds ratio for Abx reflects its role as a clinical surrogate for baseline severity rather than a causative driver. In AD patients with blunted symptoms, early antimicrobial therapy serves as a high-fidelity proxy for "biological gravity," integrating expert clinical suspicion of impending respiratory failure. In comparison to widely used severity scoring systems such as the Sequential Organ Failure Assessment (SOFA) and the Acute Physiology and Chronic Health Evaluation (APACHE II/III), our XGBoost-based model exhibited superior predictive performance and interpretability for forecasting mechanical ventilation in AD patients with pneumonia. Traditional scoring systems were not specifically designed for neurodegenerative populations and often fail to account for disease-specific physiological responses or geriatric syndromes. Moreover, while SOFA and APACHE rely on fixed, linear weightings of variables, our machine learning approach accommodates complex, non-linear interactions and variable importance tailored to this high-risk subgroup. Prior studies applying SOFA or APACHE to AD or older ICU populations have reported modest predictive capabilities (AUCs typically between 0.60–0.75) when assessing outcomes such as mortality or ventilation need 42 , 43 . In contrast, our XGBoost model achieved an AUC of 0.849 in the training set and 0.744 in the Validation Set, with strong calibration and net benefit, as demonstrated by Brier score and decision curve analysis. Head-to-head benchmarking further revealed XGBoost’s superior discriminative power over standalone clinical scores, including SOFA, OASIS, and SAPS II (all P < 0.05; Supplementary Table 7 ).Importantly, our model integrates model-agnostic interpretability through SHAP values, which is absent from conventional risk scoring tools. This study offers several methodological and translational strengths. First, it leverages a large, high-resolution critical care dataset (MIMIC-IV), enhancing the statistical power and generalizability within a real-world ICU setting. Second, we employed a rigorous analytic pipeline that incorporated LASSO regression for feature selection, multivariate logistic regression for statistical inference, and multiple machine learning algorithms for comparative benchmarking, culminating in the selection of the best-performing model. Third, comprehensive evaluation across multiple metrics—discrimination (AUC), calibration (Brier score), clinical utility (DCA), and explainability (SHAP)—provides a robust assessment of model validity. Fourth, the inclusion of SHAP analysis enables both global model transparency and individualized prediction decomposition, addressing the "black-box" criticism often directed at machine learning applications in healthcare 44 , 45 . Finally, we translated our model into a point-based nomogram, offering clinicians an intuitive, bedside-compatible risk estimation tool without the need for complex computational infrastructure. Collectively, these features position our model as a practical, explainable, and clinically relevant tool for the quantification of disease severity and to guide mechanical ventilation decisions in cognitively vulnerable ICU populations. This study has several limitations that should be acknowledged. First, the dataset was derived from a single academic center (MIMIC-IV), which may limit the generalizability of our findings to broader ICU populations. Although the database is extensive and well-curated, validation in diverse healthcare settings and patient cohorts—particularly non-U.S. populations—is essential to confirm external applicability. Second, although we implemented multiple imputation using chained equations (MICE) with a random forest algorithm to address missing data, imputation cannot fully account for potential biases introduced by non-random missingness or unmeasured confounders 46 . Third, while our model incorporated a comprehensive set of structured clinical variables, it did not include potentially important features such as advanced diagnostic techniques for precise shunt quantification or imaging data (e.g., chest radiography), laboratory biomarkers (e.g., procalcitonin, troponin), functional status, or caregiver dependence—all of which may influence the clinical course of AD patients and their risk of requiring mechanical ventilation. Finally, as with all retrospective observational studies, residual confounding cannot be entirely excluded, and causal inferences should be interpreted with caution.Future research should focus on several directions. External validation using independent datasets, such as the eICU Collaborative Research Database or multi-national ICU cohorts, is necessary to assess the robustness and transportability of our model. Second, the model could be prospectively integrated into electronic health record (EHR) systems to enable real-time risk prediction and clinical decision support. Third, the inclusion of multimodal data—including radiologic imaging, genomic information, and natural language processing of clinical notes—may further enhance the model’s predictive accuracy and personalization capacity 47 , 48 . Lastly, prospective interventional studies are needed to determine whether implementing such a model in clinical workflows improves patient-centered outcomes, optimizes ventilation strategies, or enhances ICU resource allocation in cognitively impaired populations. Additionally, social and home-care factors were not captured, despite their known impact on geriatric clinical trajectories. Conclusion This study presents a validated and interpretable machine learning model for predicting mechanical ventilation (MV) in Alzheimer’s disease (AD) patients with pneumonia, based on real-world ICU data from the MIMIC-IV database. Among five algorithms tested, the XGBoost model showed the best predictive performance and calibration, outperforming traditional scores like SOFA and APACHE-II. Key predictors—including COPD, OASIS score, platelet count, and temperature—were identified through SHAP analysis, offering mechanistic insights into AD-specific risk factors for MV. We further translated the model into a practical nomogram, enabling bedside risk estimation and individualized decision-making. This tool could reduce unnecessary ventilation and support timely intervention in high-risk patients. By bridging algorithmic precision with clinical applicability, this model offers a novel decision-support strategy for managing respiratory deterioration in elderly AD patients. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (32.1KB, docx) Supplementary Material 2 (5.6MB, docx) Acknowledgements Not applicable. The authors have no additional acknowledgements to declare. Abbreviations AD Alzheimer’s disease AKI Acute kidney injury AIC Akaike Information Criterion AUC Area under the receiver operating characteristic curve CB Cerebral bleeding CI Confidence interval COPD Chronic obstructive pulmonary disease DCA Decision curve analysis DT Decision tree EHR Electronic health record GCS Glasgow Coma Scale HF Heart failure ICU Intensive care unit IHD Ischemic heart disease IQR Interquartile range LASSO Least absolute shrinkage and selection operator LGBM Light gradient boosting machine LR Logistic regression MI Myocardial infarction ML Machine learning MIMIC-IV Medical Information Mart for Intensive Care IV MV Mechanical ventilation NomoScore Nomogram-derived risk score OASIS Oxford Acute Severity of Illness Score OR Odds ratio PLT Platelet count PPV Positive predictive value NPV Negative predictive value RDW Red cell distribution width RBC Red blood cell count SHAP Shapley Additive Explanations SOFA Sequential Organ Failure Assessment SVM Support vector machine WBC White blood cell count XGBoost Extreme gradient boosting Author contributions Xuhui Liu and Xujie Wang contributed equally to this work. Xuhui Liu designed the study, performed data analysis, interpreted the results, and drafted the manuscript. Xujie Wang obtained certified access to the MIMIC-IV database, conducted data extraction and preprocessing, and participated in statistical modeling. Rongfei Xie contributed to data cleaning, visualization, and model validation. Zhaohui Liu provided methodological supervision, assisted with result interpretation, and critically reviewed the manuscript. Biyun Wang supervised the overall study, guided the research framework, and revised the manuscript for important intellectual content. Honglin Zhai, as the corresponding author, coordinated the research team, ensured data integrity and analysis accuracy, and approved the final version for submission. All authors read and approved the final manuscript. Funding This research was supported by the 2025 Graduate Research and Practice Innovation Project of Qinghai University (Project No. 2025-GMKY-14). The funding body had no role in the design of the study, collection, analysis, interpretation of data, or writing of the manuscript. Data availability The datasets generated and/or analyzed during the current study are available in the MIMIC-IV (Medical Information Mart for Intensive Care IV) database at ( https://physionet.org/content/mimiciv/ ). Declarations Competing interests The authors declare that they have no competing interests or financial relationships that could be perceived as potential conflicts of interest regarding this research. Ethics approval and consent to participate This study utilized data from the publicly available MIMIC-IV database, which contains de-identified health information. As such, the use of this dataset is exempt from institutional review board (IRB) approval and individual informed consent in accordance with applicable guidelines and regulations. Access to the MIMIC-IV database was granted to one author (Xujie Wang) after completion of the Collaborative Institutional Training Initiative (CITI) program (Certification ID: 13931651). Consent for publication Not applicable. This study used de-identified data from the MIMIC-IV database, and no individual patient data or images are reported. Clinical trial number Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xuhui Liu and Xujie Wang contributed equally to this work. Contributor Information Biyun Wang, Email: wangby35@mail.sysu.edu.cn. Honglin Zhai, Email: 13919686237@163.com. References 1. Meyers, E. A., Sexton, C., Snyder, H. M. & Carrillo, M. C. Impact of Alzheimer’s association support and engagement in the AD/ADRD research community through the COVID-19 pandemic and beyond. Alzheimer’s Dementia 19 (7), 3222–3225 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Simpson, A. J. et al. BTS clinical statement on aspiration pneumonia. Thorax 78 (Suppl 1), s3–s21 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 3. Kalia, M. Dysphagia and aspiration pneumonia in patients with Alzheimer’s disease. Metabolism 52 (10 Suppl 2), 36–38 (2003). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Girard, T. D. et al. Long-Term Cognitive Impairment after Hospitalization for Community-Acquired Pneumonia: A Prospective Cohort Study. J. Gen. Intern. Med. 33 (6), 929–935 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Liu, H. et al. Impact of mechanical ventilation on clinical outcomes in ICU-admitted Alzheimer’s disease patients: A retrospective cohort study. Front. Public Health 12 , 1368508 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Lagu, T. et al. Dementia and Outcomes of Mechanical Ventilation. J. Am. Geriatr. Soc. 64 (10), e63–e66 (2016). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Scott L, Winzey KD, Moreira D, Bresee C, Vit JP, Tourtellotte WG, et al., "Microglia ameliorate delirium-like phenotypes in a murine model of acute ventilator-induced lung injury," Journal of neuroinflammation, V. 21, No. 1, Oct 21. 2024, pp. 270. [ DOI ] [ PMC free article ] [ PubMed ] 8. Yamanda S, Ebihara S, Ebihara T, Yamasaki M, Asamura T, Asada M, et al., "Impaired urge-to-cough in elderly patients with aspiration pneumonia," Cough (London, England), V. 4, Nov 19. 2008, pp. 11. [ DOI ] [ PMC free article ] [ PubMed ] 9. Andrani F, Aiello M, Bertorelli G, Crisafulli E, Chetta A, "Cough, a vital reflex. mechanisms, determinants and measurements," Acta bio-medica : Atenei Parmensis, V. 89, No. 4, Jan 15. 2019, pp. 477–80. [ DOI ] [ PMC free article ] [ PubMed ] 10. David MCB, Kolanko M, Del Giovane M, Lai H, True J, Beal E, et al., "Remote Monitoring of Physiology in People Living With Dementia: An Observational Cohort Study," JMIR aging, V. 6, Mar 9. 2023, pp. e43777. [ DOI ] [ PMC free article ] [ PubMed ] 11. Baldwin, M. R. Measuring and predicting long-term outcomes in older survivors of critical illness. Minerva Anestesiol. 81 (6), 650–661 (2015). [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Yang TM, Fang YH, Chang CH, Lin SJ, Liu CC, Tsai DMT, et al., "Prolonged T-Piece Spontaneous Breathing Trial and Extubation Outcomes in Patients Received Prolonged Mechanical Ventilation," Medicina (Kaunas, Lithuania), V. 61, No. 3, Feb 26. 2025. [ DOI ] [ PMC free article ] [ PubMed ] 13. Sasannejad C, Ely EW, Lahiri S, "Long-term cognitive impairment after acute respiratory distress syndrome: A review of clinical impact and pathophysiological mechanisms," Critical care (London, England), V. 23, No. 1, Nov 12. 2019, pp. 352. [ DOI ] [ PMC free article ] [ PubMed ] 14. Zhao, X., Shen, W. & Wang, G. Early Prediction of Sepsis Based on Machine Learning Algorithm. Comput. Intell. Neurosci. 2021 , 6522633 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Combs CK, Karlo JC, Kao SC, Landreth GE, "beta-Amyloid stimulation of microglia and monocytes results in TNFalpha-dependent expression of inducible nitric oxide synthase and neuronal apoptosis," The Journal of neuroscience : The official journal of the Society for Neuroscience, V. 21, No. 4, Feb 15. 2001, pp. 1179–88. [ DOI ] [ PMC free article ] [ PubMed ] 16. da Costa Dias, F. L., Ferreira Lisboa da Silva, R. M., de Moraes, E. N. & Caramelli, P. Cholinesterase inhibitors modulate autonomic function in patients with Alzheimer´s disease and mixed dementia. Curr. Alzheimer Res. 10 (5), 476–481 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Hernández-Monsalves AH, Letelier P, Morales C, Rojas E, Saez MA, Coña N, et al., "A Machine Learning Model for Predicting Intensive Care Unit Admission in Inpatients with COVID-19 Using Clinical Data and Laboratory Biomarkers," Biomedicines, V. 13, No. 5, Apr 24. 2025. [ DOI ] [ PMC free article ] [ PubMed ] 18. Gong, Y. et al. Development and Validation of Nomograms for Predicting Overall and Breast Cancer-Specific Survival in Young Women with Breast Cancer: A Population-Based Study. Translational oncology 11 (6), 1334–1342 (2018). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Wu, Y. Y. et al. Alterations of the neuroinflammatory markers IL-6 and TRAIL in Alzheimer’s disease. Dement. Geriatr. Cogn. Disord. Extra. 5 (3), 424–434 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Griseta, C. et al. Serum levels of IL-6 are associated with cognitive impairment in the salus in apulia population-based study. Heliyon 9 (3), e13972 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Chen Z, Balachandran YL, Chong WP, Chan KWY, "Roles of Cytokines in Alzheimer’s Disease," International journal of molecular sciences, V. 25, No. 11, May 26. 2024. [ DOI ] [ PMC free article ] [ PubMed ] 22. Holt, R. I. G., Cockram, C. S., Ma, R. C. W. & Luk, A. O. Y. Diabetes and infection: Review of the epidemiology, mechanisms and principles of treatment. Diabetologia 67 (7), 1168–1180 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Tao, B. et al. The relationship between hypoxia and Alzheimer’s disease: An updated review. Front. Aging Neurosci. 16 , 1402774 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Lin TK, Huang CR, Lin KJ, Hsieh YH, Chen SD, Lin YC, et al., "Potential Roles of Hypoxia-Inducible Factor-1 in Alzheimer’s Disease: Beneficial or Detrimental?," Antioxidants (Basel, Switzerland), V. 13, No. 11, Nov 11. 2024. [ DOI ] [ PMC free article ] [ PubMed ] 25. Greve HJ, Dunbar AL, Lombo CG, Ahmed C, Thang M, Messenger EJ, et al., "The bidirectional lung brain-axis of amyloid-β pathology: ozone dysregulates the peri-plaque microenvironment," Brain : A journal of neurology, V. 146, No. 3, Mar 1. 2023, pp. 991–1005. [ DOI ] [ PMC free article ] [ PubMed ] 26. Christopoulou ME, Papakonstantinou E, Stolz D, "Matrix Metalloproteinases in Chronic Obstructive Pulmonary Disease," International journal of molecular sciences, V. 24, No. 4, Feb 14. 2023. [ DOI ] [ PMC free article ] [ PubMed ] 27. Paunikar, S. & Chakole, V. Pulmonary Shunt in Critical Care: A Comprehensive Review of Pathophysiology, Diagnosis, and Management Strategies. Cureus 16 (9), e68505 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Petersson, J. & Glenny, R. W. Gas exchange and ventilation-perfusion relationships in the lung. Eur. Respir. J. 44 (4), 1023–1041 (2014). [ DOI ] [ PubMed ] [ Google Scholar ] 29. Li LC, Chen WY, Chen JB, Lee WC, Chang CC, Tzeng HT, et al., "The AST-120 Recovers Uremic Toxin-Induced Cognitive Deficit via NLRP3 Inflammasome Pathway in Astrocytes and Microglia," Biomedicines, V. 9, No. 9, Sep 17. 2021. [ DOI ] [ PMC free article ] [ PubMed ] 30. Sedaghat, S. et al. The association of kidney function with plasma Amyloid-β levels and brain amyloid deposition. J. Alzheimers Dis. 92 (1), 229–239 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Pang H, Kumar S, Ely EW, Gezalian MM, Lahiri S, "Acute kidney injury-associated delirium: A review of clinical and pathophysiological mechanisms," Critical care (London, England), V. 26, No. 1, Aug 27. 2022, pp. 258. [ DOI ] [ PMC free article ] [ PubMed ] 32. Li Q, Li G, Li D, Chen Y, Zhou F, "Acute kidney injury in elderly patients receiving invasive mechanical ventilation: Early versus late onset," European journal of medical research, V. 29, No. 1, Dec 18. 2024, pp. 590. [ DOI ] [ PMC free article ] [ PubMed ] 33. Verma K, Amitabh, Chandra M, Prasad DN, Debnath C, Mohanty H, et al., "Alteration in cerebral blood flow, kynurenines with respect to mood profile in freshly recruited armed forces personnel," Journal of psychiatric research, V. 149, May. 2022, pp. 155–61. [ DOI ] [ PubMed ] 34. Pamart G, Gosset P, Le Rouzic O, Pichavant M, Poulain-Godefroy O, "Kynurenine Pathway in Respiratory Diseases," International journal of tryptophan research : IJTR, V. 17. 2024, pp. 11786469241232871. [ DOI ] [ PMC free article ] [ PubMed ] 35. Leung, J., Behpour, A., Sokol, N., Mohanta, A. & Kassner, A. Assessment of intracranial blood flow velocities using a computer controlled vasoactive stimulus: A comparison between phase contrast magnetic resonance angiography and transcranial Doppler ultrasonography. J. Magn. Reson. Imaging 38 (3), 733–738 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 36. Lesurtel M, Graf R, Aleil B, Walther DJ, Tian Y, Jochum W, et al., "Platelet-derived serotonin mediates liver regeneration," Science (New York, NY), V. 312, No. 5770, Apr 7. 2006, pp. 104–7. [ DOI ] [ PubMed ] 37. Wang J, Hu HJ, Liu ZK, Liu JJ, Wang SS, Cheng Q, et al., "Pharmacological inhibition of asparaginyl endopeptidase by δ-secretase inhibitor 11 mitigates Alzheimer’s disease-related pathologies in a senescence-accelerated mouse model," Translational neurodegeneration, V. 10, No. 1, Mar 31. 2021, pp. 12. [ DOI ] [ PMC free article ] [ PubMed ] 38. Cortes-Canteli M, Paul J, Norris EH, Bronstein R, Ahn HJ, Zamolodchikov D, et al., "Fibrinogen and beta-amyloid association alters thrombosis and fibrinolysis: A possible contributing factor to Alzheimer’s disease," Neuron, V. 66, No. 5, Jun 10. 2010, pp. 695–709. [ DOI ] [ PMC free article ] [ PubMed ] 39. Tournissac, M. et al. Metabolic determinants of Alzheimer’s disease: A focus on thermoregulation. Ageing Res. Rev. 72 , 101462 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 40. Imataka G, Fujita Y, Kikuchi J, Wake K, Ono K, Yoshihara S, "Brain Hypothermia Therapy and Targeted Temperature Management for Acute Encephalopathy in Children: Status and Prospects," Journal of clinical medicine, V. 12, No. 6, Mar 7. 2023. [ DOI ] [ PMC free article ] [ PubMed ] 41. Norman, D. C. Fever in the elderly. Clin. Infect. Dis. 31 (1), 148–151 (2000). [ DOI ] [ PubMed ] [ Google Scholar ] 42. Qiao, Q., Lu, G., Li, M., Shen, Y. & Xu, D. Prediction of outcome in critically ill elderly patients using APACHE II and SOFA scores. J. Int. Med. Res. 40 (3), 1114–1121 (2012). [ DOI ] [ PubMed ] [ Google Scholar ] 43. Romo Gonzales JE, Silva Obregón J, Martin Dal Gesso C, Gallardo Culebradas P, Saboya Sanchez S, Torralba M, "Evaluation of APACHE II, SAPS II and sofa as predictors of mortality in patients over 80 years admitted to ICU." Intensive Care Med Exp. 2015 Oct 1;3(Suppl 1):A343. 10.1186/2197-425X-3-S1-A343. eCollection 2015 Dec. 44. Luo H, Xiang C, Zeng L, Li S, Mei X, Xiong L, et al., "SHAP based predictive modeling for 1 year all-cause readmission risk in elderly heart failure patients: Feature selection and model interpretation," Scientific reports, V. 14, No. 1, Jul 31. 2024, pp. 17728. [ DOI ] [ PMC free article ] [ PubMed ] 45. Ponce-Bobadilla, A. V., Schmitt, V., Maier, C. S., Mensing, S. & Stodtmann, S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin. Transl. Sci. 17 (11), e70056 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 46. Lin, S., Yang, M., Liu, C., Wang, Z. & Long, X. A pretrain-finetune approach for improving model generalizability in outcome prediction of acute respiratory distress syndrome patients. Int. J. Med. Inform. 186 , 105397 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 47. Liu M, Fan Z, Gao Y, Mubonanyikuzo V, Wu R, Li W, et al., "A two-tier feature selection method for predicting mortality risk in ICU patients with acute kidney injury," Scientific reports, V. 14, No. 1, Jul 22. 2024, pp. 16794. [ DOI ] [ PMC free article ] [ PubMed ] 48. Lin, J. et al. Development and validation of multimodal models to predict the 30-day mortality of ICU patients based on clinical parameters and chest X-Rays. J. Imaging Inform. Med. 37 (4), 1312–1322 (2024). [ DOI ] [ PMC free article ] [ 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 (32.1KB, docx) Supplementary Material 2 (5.6MB, docx) Data Availability Statement The datasets generated and/or analyzed during the current study are available in the MIMIC-IV (Medical Information Mart for Intensive Care IV) database at ( https://physionet.org/content/mimiciv/ ). Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.4 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases 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

• Epanechnikov Nonparametric Kernel Density Estimation Based Feature-Learning in Respiratory Disease Chest X-Ray Images#358887
• Cross-Silo Federated Learning for Predicting Successful Mechanical Ventilation Weaning Across 5 Intensive Care Unit Databases: Retrospective Database Analysis.#615791
• A General Multimodal Probing Framework for Data Fusion and Prediction with Frozen Large Language Models: Evidence from Multimodal EHR Data#1036731
• Development of a Multimodal Deep Learning Model for Pediatric Patients#895616
• Multimodal machine learning predicts type 2 respiratory failure in COPD exacerbations: a multicenter XGBoost model with clinical nomogram.#415448