Comparison of machine learning methods for prediction of venous thromboembolism among hospitalized adults.

Hospital-acquired venous thromboembolism (HA-VTE) is a leading cause of morbidity and mortality among hospitalized adults. Numerous prognostic models have been developed to identify those patients with elevated risk of HA-VTE. None, however, has met the necessary criteria to guide clinical decision-making. This study outlines a protocol for refining and validating a general-purpose prognostic model for HA-VTE, designed for real-time automation within the electronic health record (EHR) system. A retrospective cohort of 132,561 inpatient encounters (89,586 individual patients) at a large academic medical center will be collected, along with clinical and demographic data available as part of routine care. Data for temporal, geographic, and domain external validation cohorts will also be collected. Logistic regression will be used to predict occurrence of HA-VTE during an inpatient encounter. Variables considered for model inclusion will be based on prior demonstrated association with HA-VTE and their availability in both retrospective EHR data and routine clinical care. Least absolute shrinkage and selection operator (LASSO) with tenfold cross-validation will be used for initial variable selection. Variables selected by the LASSO procedure, along with those deemed necessary by clinicians, will be used in an unpenalized multivariable logistic regression model. Discrimination and calibration will be reported for the derivation and validation cohorts. Discrimination will be measured using Harrell's C statistic. Calibration will be measured using calibration intercept, calibration slope, Brier score, integrated calibration index, and visual examination of non-linear calibration curve. Model reporting will adhere to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis guidelines for clinical prediction models using machine learning methods (TRIPOD + AI). We describe methods for developing, evaluating, and validating a prognostic model for HA-VTE using routinely collected EHR data. By combining best practices in statistical development and validation, knowledge engineering, and clinical domain knowledge, the resulting model should be well suited for real-time clinical implementation. Although this protocol describes our development of a model for HA-VTE, the general approach can be applied to other clinical outcomes.

The increase in hospitalizations of patients with noncommunicable diseases remains a key challenge for healthcare systems. A significant portion of these hospitalizations is potentially preventable with timely outpatient monitoring and effective primary care. Predictive models based on data from health information systems (HIS) and electronic health records (EHRs) make it possible to identify patients at increased risk of hospitalization and improve their management strategies. The article systematizes and summarizes international experience in using hospitalization risk prognostic models, developed on the basis of HIS and EHR data. We made a systematic review of publications presented in Russian and foreign databases (Scopus, PubMed, BMC Health Services Research, BMJ Research, Google Scholar, Elibrary, Oxford Academic, Wiley Online Library) for the period 1993-2023. The analysis included studies that used EHR and HIS data to develop and/or validate prognostic models of hospitalization risk. The information is systematized by following predictor categories: diseases and comorbidities, prescriptions, interaction with healthcare organizations, sociodemographic and laboratory parameters. The final analysis included 14 studies describing 28 models for predicting the hospitalization risk. Most studies used logistic regression. Machine learning methods (gradient boosting, random forest, and Bayesian ensemble models) were used in some studies. The categories with the highest prognostic value were "diseases and comorbidities" (42,3%), "medication prescriptions" (23,6%), and "interaction with health organizations" (19,8%). Including temporal characteristics (frequency and recency of hospitalizations, visits to primary care facilities, and laboratory tests) increased the model accuracy. Predictive models based on EHR and HIS data demonstrate high discriminatory power and enable the assessment of hospitalization risk at the primary care level. Machine learning methods in some studies demonstrated the possibility of a more in-depth analysis of the relationships between predictors and improved prediction accuracy. A promising area of research is the further development, external validation, and adaptation of models using expanded EHR and HIS data sets for its use in outpatient monitoring and preventing noncommunicable diseases.

Palliative care has been demonstrated to have positive effects for patients, families, health care providers, and health systems. Early identification of patients who are likely to benefit from palliative care would increase opportunities to provide these services to those most in need. This study predicted all-cause mortality of patients as a surrogate for patients who could benefit from palliative care. Claims and electronic health record (EHR) data for 59,639 patients from a large integrated health care system were utilized. A deep learning algorithm-a long short-term memory (LSTM) model-was compared with other machine learning models: deep neural networks, random forest, and logistic regression. We conducted prediction analyses using combined claims data and EHR data, only claims data, and only EHR data, respectively. In each case, the data were randomly split into training (80%), validation (10%), and testing (10%) data sets. The models with different hyperparameters were trained using the training data, and the model with the best performance on the validation data was selected as the final model. The testing data were used to provide an unbiased performance evaluation of the final model. In all modeling scenarios, LSTM models outperformed the other 3 models, and using combined claims and EHR data yielded the best performance. LSTM models can effectively predict mortality by using a combination of EHR data and administrative claims data. The model could be used as a promising clinical tool to aid clinicians in early identification of appropriate patients for palliative care consultations.

ObjectivesTo develop and compare prediction models for 30-day and 1-year mortality in Heart failure with preserved ejection fraction (HFpEF) using EHR data, utilizing both traditional and machine learning (ML) techniques.BackgroundHFpEF represents 1 in 2 heart failure patients. Predictive models in HFpEF, specifically those derived from electronic health record (EHR) data, are less established.MethodsUsing MIMIC-IV EHR data from 2008−2019, patients aged ≥ 18 years admitted with a primary diagnosis of HFpEF were identified using ICD-9 and 10 codes. Demographics, vital signs, prior diagnoses, and lab data were extracted. Data was partitioned into 80% training, 20% test sets. Prediction models from seven model classes (Support Vector Classifier (SVC), Logistic Regression, Lasso Regression, Elastic Net, Random Forest, Histogram-based Gradient Boosting Classifier (HGBC), and eXtreme Gradient Boosting (XGBoost)) were developed using various imputation and oversampling techniques with 5-fold cross-validation. Model performance was compared using several metrics, and individual feature importance assessed using SHapley Additive exPlanations (SHAP) analysis.ResultsAmong 3,235 hospitalizations for HFpEF, 30-day mortality was 6.3%, and 1- year mortality was 29.2%. Logistic regression performed well for 30-day mortality (Area Under the Receiver operating characteristic curve (AUC) 0.83), whereas Random Forest (AUC 0.79) and HGBC (AUC 0.78) for 1-year mortality. Age and NT-proBNP were the strongest predictors in SHAP analyses for both outcomes.ConclusionModels derived from EHR data can predict mortality after HFpEF hospitalization with comparable performance to models derived from registry or trial data, highlighting the potential for clinical implementation.

To develop and compare prediction models for 30-day and 1-year mortality in Heart failure with preserved ejection fraction (HFpEF) using EHR data, utilizing both traditional and machine learning (ML) techniques. HFpEF represents 1 in 2 heart failure patients. Predictive models in HFpEF, specifically those derived from electronic health record (EHR) data, are less established. Using MIMIC-IV EHR data from 2008-2019, patients aged ≥ 18 years admitted with a primary diagnosis of HFpEF were identified using ICD-9 and 10 codes. Demographics, vital signs, prior diagnoses, and lab data were extracted. Data was partitioned into 80% training, 20% test sets. Prediction models from seven model classes (Support Vector Classifier (SVC), Logistic Regression, Lasso Regression, Elastic Net, Random Forest, Histogram-based Gradient Boosting Classifier (HGBC), and eXtreme Gradient Boosting (XGBoost)) were developed using various imputation and oversampling techniques with 5-fold cross-validation. Model performance was compared using several metrics, and individual feature importance assessed using SHapley Additive exPlanations (SHAP) analysis. Among 3,235 hospitalizations for HFpEF, 30-day mortality was 6.3%, and 1- year mortality was 29.2%. Logistic regression performed well for 30-day mortality (Area Under the Receiver operating characteristic curve (AUC) 0.83), whereas Random Forest (AUC 0.79) and HGBC (AUC 0.78) for 1-year mortality. Age and NT-proBNP were the strongest predictors in SHAP analyses for both outcomes. Models derived from EHR data can predict mortality after HFpEF hospitalization with comparable performance to models derived from registry or trial data, highlighting the potential for clinical implementation.

Venous thromboembolism (VTE) risk scores assist providers in determining the relative benefit of prophylaxis for individual patients. While automated risk calculation using simpler electronic health record (EHR) data is feasible, it lacks clinical nuance and may be less predictive. Automated calculation of the Padua Prediction Score (PPS), requiring more complex input such as recent medical events and clinical status, may save providers time and increase risk score use. We developed the Automated Padua Prediction Score (APPS) to auto-calculate a VTE risk score using EHR data drawn from prior encounters and the first 4 hours of admission. We compared APPS to standard practice of clinicians manually calculating the PPS to assess VTE risk. Cohort study of 30,726 hospitalized patients. APPS was compared to manual calculation of PPS by chart review from 300 randomly selected patients. Prediction of hospital-acquired VTE not present on admission. Compared to manual PPS calculation, no significant difference in average score was found (5.5 vs. 5.1, P = 0.073), and area under curve (AUC) was similar (0.79 vs. 0.76). Hospital- acquired VTE occurred in 260 (0.8%) of 30,726 patients. Those without VTE averaged APPS of 4.9 (standard deviation [SD], 2.6) and those with VTE averaged 7.7 (SD, 2.6). APPS had AUC = 0.81 (confidence interval [CI], 0.79-0.83) in patients receiving no pharmacologic prophylaxis and AUC = 0.78 (CI, 0.76- 0.82) in patients receiving pharmacologic prophylaxis. Automated calculation of VTE risk had similar ability to predict hospital-acquired VTE as manual calculation despite differences in how often specific scoring criteria were considered present by the 2 methods. Journal of Hospital Medicine 2017;12: 231- 237.

Heart failure is a critical side effect of many cancer treatments. Identifying cancer patients at a high risk of cardiotoxicity before cancer treatment is a critical step towards early detection and possible prevention. This study seeks to examine how genetic data can be used with Electronic Health Record (EHR) data to identify cancer patients at risk of treatment-related heart failure. We explored four machine learning models, including Logistic Regression (LR), Support Vector Machines (SVMs), Random Forest (RF), and Gradient Boost (GB) for heart failure prediction using EHR data linked with genetic data from the UK Biobank. We first identified the best machine learning model using only EHR data and then further added genetic data to the model using three different strategies. The experimental results show that the GB model combining EHR data and genetic data achieved the best area under the curve (AUC) score of 0.7781, outperforming the machine learning models using only EHR data. Among the three strategies of including genetic data, the genome-wide association study (GWAS) method achieved the best performance. Our study shows that genomic data can be used to improve the performance of heart failure prediction among cancer patients.

INTRODUCTION: Heart failure with preserved ejection fraction (HFpEF) represents 50% of heart failure patients. Predictive models in HFpEF, specifically those derived from electronic health record (EHR) data, are less established. HYPOTHESIS: EHR models can predict 30-day and 1-year mortality in HFpEF patients. Machine learning (ML) models compared with traditional may show superior performance. AIM: Develop and compare prediction models for 30-day and 1-year mortality in HFpEF using EHR data, using traditional and ML techniques. Methods: Using MIMIC-IV data (EHR data from the Beth Israel Deaconess Medical Center in Boston) from 2008-2019, we identified patients aged ≥ 18 years admitted with a primary diagnosis of HFpEF using ICD-9 and 10 codes. Demographics, vital signs, prior diagnoses, and lab data were extracted. After preprocessing, data was partitioned into 70% training and 30% test. Prediction models from seven model classes were developed using a variety of imputation and oversampling techniques using 5-fold cross validation - Support Vector Classifier (SVC), Logistic Regression, Lasso Regression, Elastic Net, Random Forest, HistGradient Boosting Classifier, and XGBoost. Model performance was compared using several metrics. SHapley Additive exPlanations (SHAP) analysis was used to assess individual feature influence on model output. Results: Among 3910 individual hospitalizations for HFpEF the 30-day mortality was 6.3% and 1-year mortality was 29.2%. Performance metrics of the best model in each class is shown in Table 1. By AUCs, Lasso and Elastic Net Regression models provided superior prediction for 30-day mortality (AUCs of 0.79, each), whereas Random Forest and HistGradient Boosting Classifier for 1-year mortality (AUCs of 0.78 each). Overall, considering all metrics, Elastic Net and Lasso regression models gave a good combination of sensitivity, specificity and AUC enhancing their clinical applicability. SHAP analyses showed age, NT-proBNP and sodium levels, WBC, and platelet counts as key drivers of these predictions. Conclusion: Models derived from EHR data, which is important for implementability, predict mortality after HFpEF hospitalization with good performance comparable to models derived from registry or trial data.

In patients with diabetes, current models for predicting the risk of readmission within 30 days of hospital discharge vary in performance. We previously published the Diabetes Early Readmission Risk Indicator (DERRI TM), a logistic regression (LR) model based on 10 simple features with modest predictive performance (C-statistic 0.69). The current study aims to develop a more accurate model using deep learning on electronic health record (EHR) data. We electronically abstracted data from 36,563 patients with diabetes and at least 1 hospitalization at an urban, academic medical center between 7/1/2010 and 12/31/2020. One hospitalization per patient (index hospitalization) was randomly selected for analysis. A deep learning long short-term memory Recurrent Neural Network (RNN) was developed and compared to traditional linear and non-linear models: LR, AdaBoost, and Random Forest (RF). Models to predict unplanned, all-cause readmission were developed using demographics, vital signs, diagnostic and procedure codes, medications, laboratory tests, and administrative data as defined by the National Patient-Centered Clinical Research Network (PCORnet) Common Data Model. Unplanned readmissions were identified according to the Centers for Medicare and Medicaid (CMS) definition. A look-back time of 1 year before the index hospitalization and up to 60 previous ambulatory and hospital visits were used for learning and inference. Data dimensionality was reduced to 3,000 features by Singular Value Decomposition. The RNN model C-statistic is significantly greater than those of the traditional models (RNN 0.78, AdaBoost 0.71, RF 0.73, and LR 0.71, p<0. 0001). Likewise, the F1-score is numerically greater for the RNN model (RNN 0.76, AdaBoost 0.75, RF 0.75, LR 0.72). Direct comparison to the DERRI TM is limited by lack of EHR data on two of the component variables (employment status and zip code). The deep learning RNN model outperforms the DERRI TM and is based on more generalizable EHR data. This new model may provide the basis for a more useful readmission risk prediction tool for patients with diabetes. Deep learning models may outperform traditional models at predicting readmission risk in this population.Presentation: No date and time listed

Adverse drug events (ADEs) pose a significant challenge in current clinical practice. Machine learning (ML) has been increasingly used to predict specific ADEs using electronic health record (EHR) data. This systematic review provides a comprehensive overview of the application of ML in predicting specific ADEs based on EHR data. A systematic search of PubMed, Web of Science, Embase, and IEEE Xplore was conducted to identify relevant articles published from the inception to 20 May 2024. Studies that developed ML models for predicting specific ADEs or ADEs associated with particular drugs were included using EHR data. A total of 59 studies met the inclusion criteria, covering 15 drugs and 15 ADEs. In total, 38 machine learning algorithms were reported, with random forest (RF) being the most frequently used, followed by support vector machine (SVM), eXtreme gradient boosting (XGBoost), decision tree (DT), and light gradient boosting machine (LightGBM). The performance of the ML models was generally strong, with an average area under the curve (AUC) of 76.68% ± 10.73, accuracy of 76.00% ± 11.26, precision of 60.13% ± 24.81, sensitivity of 62.35% ± 20.19, specificity of 75.13% ± 16.60, and an F1 score of 52.60% ± 21.10. The combined sensitivity, specificity, diagnostic odds ratio (DOR), and AUC from the summary receiver operating characteristic (SROC) curve using a random effects model were 0.65 (95% CI: 0.65-0.66), 0.89 (95% CI: 0.89-0.90), 12.11 (95% CI: 8.17-17.95), and 0.8069, respectively. The risk factors associated with different drugs and ADEs varied. Future research should focus on improving standardization, conducting multicenter studies that incorporate diverse data types, and evaluating the impact of artificial intelligence predictive models in real-world clinical settings. https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42024565842, identifier CRD42024565842.

BackgroundDelirium frequently occurs among patients admitted to the intensive care unit (ICU). There is limited evidence to support interventions to treat or resolve delirium in patients who have already developed delirium. Therefore, the early recognition and prevention of delirium are important in the management of critically ill patients.ObjectiveThis study aims to develop and validate a delirium prediction model within 24 hours of admission to the ICU using electronic health record data. The algorithm was named the Prediction of ICU Delirium (PRIDE).MethodsThis is a retrospective cohort study performed at a tertiary referral hospital with 120 ICU beds. We only included patients who were 18 years or older at the time of admission and who stayed in the medical or surgical ICU. Patients were excluded if they lacked a Confusion Assessment Method for the ICU record from the day of ICU admission or if they had a positive Confusion Assessment Method for the ICU record at the time of ICU admission. The algorithm to predict delirium was developed using patient data from the first 2 years of the study period and validated using patient data from the last 6 months. Random forest (RF), Extreme Gradient Boosting (XGBoost), deep neural network (DNN), and logistic regression (LR) were used. The algorithms were externally validated using MIMIC-III data, and the algorithm with the largest area under the receiver operating characteristics (AUROC) curve in the external data set was named the PRIDE algorithm.ResultsA total of 37,543 cases were collected. After patient exclusion, 12,409 remained as our study population, of which 3816 (30.8%) patients experienced delirium incidents during the study period. Based on the exclusion criteria, out of the 96,016 ICU admission cases in the MIMIC-III data set, 2061 cases were included, and 272 (13.2%) delirium incidents occurred. The average AUROCs and 95% CIs for internal validation were 0.916 (95% CI 0.916-0.916) for RF, 0.919 (95% CI 0.919-0.919) for XGBoost, 0.881 (95% CI 0.878-0.884) for DNN, and 0.875 (95% CI 0.875-0.875) for LR. Regarding the external validation, the best AUROC were 0.721 (95% CI 0.72-0.721) for RF, 0.697 (95% CI 0.695-0.699) for XGBoost, 0.655 (95% CI 0.654-0.657) for DNN, and 0.631 (95% CI 0.631-0.631) for LR. The Brier score of the RF model is 0.168, indicating that it is well-calibrated.ConclusionsA machine learning approach based on electronic health record data can be used to predict delirium within 24 hours of ICU admission. RF, XGBoost, DNN, and LR models were used, and they effectively predicted delirium. However, with the potential to advise ICU physicians and prevent ICU delirium, prospective studies are required to verify the algorithm’s performance.

Big data analytics research using heterogeneous electronic health record (EHR) data requires accurate identification of disease phenotype cases and controls. Overreliance on ground truth determination based on administrative data can lead to biased and inaccurate findings. Hospital-acquired venous thromboembolism (HA-VTE) is challenging to identify due to its temporal evolution and variable EHR documentation. To establish ground truth for machine learning modeling, we compared accuracy of HA-VTE diagnoses made by administrative coding to manual review of gold standard diagnostic test results. We performed retrospective analysis of EHR data on 3680 adult stepdown unit patients identifying HA-VTE. International Classification of Diseases, Ninth Revision (ICD-9-CM) codes for VTE were identified. 4544 radiology reports associated with VTE diagnostic tests were screened using terminology extraction and then manually reviewed by a clinical expert to confirm diagnosis. Of 415 cases with ICD-9-CM codes for VTE, 219 were identified with acute onset type codes. Test report review identified 158 new-onset HA-VTE cases. Only 40% of ICD-9-CM coded cases (n = 87) were confirmed by a positive diagnostic test report, leaving the majority of administratively coded cases unsubstantiated by confirmatory diagnostic test. Additionally, 45% of diagnostic test confirmed HA-VTE cases lacked corresponding ICD codes. ICD-9-CM coding missed diagnostic test-confirmed HA-VTE cases and inaccurately assigned cases without confirmed VTE, suggesting dependence on administrative coding leads to inaccurate HA-VTE phenotyping. Alternative methods to develop more sensitive and specific VTE phenotype solutions portable across EHR vendor data are needed to support case-finding in big-data analytics.

BackgroundThe main goal of this study is to explore the use of features representing patient-level electronic health record (EHR) data, generated by the unsupervised deep learning algorithm autoencoder, in predictive modeling. Since autoencoder features are unsupervised, this paper focuses on their general lower-dimensional representation of EHR information in a wide variety of predictive tasks.MethodsWe compare the model with autoencoder features to traditional models: logistic model with least absolute shrinkage and selection operator (LASSO) and Random Forest algorithm. In addition, we include a predictive model using a small subset of response-specific variables (Simple Reg) and a model combining these variables with features from autoencoder (Enhanced Reg). We performed the study first on simulated data that mimics real world EHR data and then on actual EHR data from eight Advocate hospitals.ResultsOn simulated data with incorrect categories and missing data, the precision for autoencoder is 24.16% when fixing recall at 0.7, which is higher than Random Forest (23.61%) and lower than LASSO (25.32%). The precision is 20.92% in Simple Reg and improves to 24.89% in Enhanced Reg. When using real EHR data to predict the 30-day readmission rate, the precision of autoencoder is 19.04%, which again is higher than Random Forest (18.48%) and lower than LASSO (19.70%). The precisions for Simple Reg and Enhanced Reg are 18.70 and 19.69% respectively. That is, Enhanced Reg can have competitive prediction performance compared to LASSO. In addition, results show that Enhanced Reg usually relies on fewer features under the setting of simulations of this paper.ConclusionsWe conclude that autoencoder can create useful features representing the entire space of EHR data and which are applicable to a wide array of predictive tasks. Together with important response-specific predictors, we can derive efficient and robust predictive models with less labor in data extraction and model training.

Objective To assess the performance of machine learning (ML) models in predicting gestational diabetes mellitus (GDM) using electronic health record (EHR) data from the first antenatal visit, and determine whether incorporating previous pregnancies data improves performance. Methods and analysis In this retrospective cohort study, ML models were developed to predict GDM using EHR data (n=27 561, GDM 11.6%). Four ML algorithms: Logistic Regression (LR), Random Forest (RF), XGBoost (XGB) and Explainable Boosting Machine (EBM) were trained with seven to nine top clinical predictors from the EHRs. Models were trained and evaluated in separate first-trimester (n=27 561), nulliparous (n=11 623) and multiparous and past-pregnancy (n=4005) cohorts. Discrimination was measured by the area under the receiver-operating characteristic curve (AUC, 95 % CI), and calibration was assessed by slope and intercept. Decision-curve analysis was performed for models. Results First-trimester models achieved AUC 0.819 (95% CI 0.811-0.827, slope=1.010, intercept=0.013) with LR, similar to more complex models such as XGB (AUC 0.818, slope=1.004, intercept=0.007), EBM (AUC 0.817, slope=0.988, intercept=-0.017) and RF (AUC 0.817, slope=1.062, intercept=0.103). Among nulliparous women, there was little difference between LR (AUC 0.813) and XGB, EBM or RF (0.805–0.814). When past pregnancy features were added to first-trimester data in multiparous women, discrimination improved: EBM AUC 0.885 (95% CI 0.867 to 0.900; slope 0.994), RF 0.878, LR 0.874 and XGB 0.876. Models using data from past pregnancies only achieved good discrimination (AUC 0.860, 95% CI 0.839 to 0.879; slope=1.028). Conclusion A small panel of clinically selected variables provides robust early-pregnancy GDM prediction (AUC ∼0.81) and even stronger performance (AUC ∼0.86–0.89) when past pregnancy information is incorporated in multiparous women. Past pregnancy data alone gives useful preconception risk estimates. These findings highlight the promise of early GDM risk identification in both nulliparous and multiparous populations; however, additional research, including external validation and clinical trials, is needed to determine the models’ practical utility and effect on maternal and neonatal outcomes.

Over 6 million patients in the United States are affected by Alzheimer's Disease and Related Dementias (ADRD). Early detection of ADRD can significantly improve patient outcomes through timely treatment. To develop and validate machine learning (ML) models for early ADRD diagnosis and prediction using de-identified EHR data from the University of Missouri (MU) Healthcare. Retrospective case-control study. The study used de-identified EHR data provided by the MU NextGen Biomedical Informatics, modeled with the PCORnet Common Data Model (CDM). An initial cohort of 380,269 patients aged 40 or older with at least two healthcare encounters was narrowed to a final dataset of 4,012 ADRD cases and 119,723 controls. Six ML classifier models: Gradient-Boosted Trees (GBT), Light Gradient-Boosting Machine (LightGBM), Random Forest (RF), eXtreme Gradient-Boosting (XGBoost), Logistic Regression (LR), and Adaptive Boosting (AdaBoost) were evaluated using Area Under the Receiver Operating Characteristic Curve (AUC-ROC), accuracy, sensitivity, specificity, and F1 score. SHAP (SHapley Additive exPlanations) analysis was applied to interpret predictions. The GBT model achieved the best AUC-ROC scores of 0.809-0.833 across 1- to 5-year prediction windows. SHAP analysis identified depressive disorder, age groups 80-90 yrs and 70-80 yrs, heart disease, anxiety, and the novel risk factors of sleep apnea, and headache. This study underscores the potential of ML models for leveraging EHR data to enable early ADRD prediction, supporting timely interventions, and improving patient outcomes. By identifying both established and novel risk factors, these findings offer new opportunities for personalized screening and management strategies, advancing both clinical and informatics science.