A structured machine learning approach using XGBoost achieved AUROC of 0.845 for predicting 30-day mortality in CAD patients, with key predictive features including the Charlson Comorbidity Index, hemoglobin level, emergency room admission status, age, and creatinine level.
Key Findings
Results
XGBoost achieved the best predictive performance for the overall CAD cohort and acute CAD cohort with AUROCs of 0.845 and 0.820, respectively.
Multiple machine learning algorithms were compared using stratified fivefold cross-validation
XGBoost outperformed other models for both the overall cohort (AUROC 0.845) and the acute CAD subgroup (AUROC 0.820)
Logistic regression performed best for the chronic CAD subgroup (AUROC 0.766)
Performance metrics evaluated included AUC, accuracy, sensitivity, specificity, NPV, PPV, and F1 score
Results
The top predictive features for 30-day mortality in CAD patients were the Charlson Comorbidity Index, hemoglobin level, emergency room admission status, age, and creatinine level.
These features were identified across models as the key drivers of mortality prediction
Charlson Comorbidity Index ranked as the leading predictive feature, reflecting overall comorbidity burden
Both laboratory values (hemoglobin, creatinine) and clinical presentation factors (ER admission status) were among the top predictors
Age was also identified as a significant independent predictor
Methods
The study included 23,267 CAD patients with an overall 30-day mortality rate of 5.2%, with notably higher mortality in the external validation cohort.
Total cohort: 23,267 patients with mean age 64.9 years
Overall deaths: 1,215 (5.2%)
Internal cohort (n=15,510): 570 deaths (3.7%)
External validation cohort from Shuang Ho Hospital (n=7,757): 645 deaths (8.3%), more than double the mortality rate of the internal cohort
Methods
The study employed a retrospective multicenter design using data from the Taipei Medical University Clinical Research Database (TMUCRD) with external validation at Shuang Ho Hospital.
Retrospective cohort study design was used
Data sourced from the TMUCRD, a structured electronic health records database
External validation was performed using an independent cohort from Shuang Ho Hospital (n=7,757)
Both acute and chronic CAD patients were analyzed as separate subgroups in addition to the overall cohort
Model performance was evaluated using stratified fivefold cross-validation
Results
Machine learning models demonstrated promising discriminative performance for predicting 30-day mortality across both acute and chronic CAD patient populations.
Acute CAD cohort best model (XGBoost): AUROC 0.820
Chronic CAD cohort best model (logistic regression): AUROC 0.766
The authors noted this performance provides 'valuable insights that could enhance personalized care and inform clinical decisions'
Different algorithms performed best for acute versus chronic CAD, suggesting disease subtype may influence optimal modeling approach
What This Means
This research suggests that machine learning models can reliably predict which patients hospitalized with coronary artery disease (CAD) are at risk of dying within 30 days. The study analyzed electronic health records from over 23,000 patients across multiple hospitals in Taiwan. Among several artificial intelligence approaches tested, a method called XGBoost performed best overall, correctly distinguishing high-risk from low-risk patients about 84.5% of the time—a level of accuracy that could be clinically meaningful for guiding care decisions.
The study found that five factors were most important for predicting early death: a score measuring how many serious health conditions a patient has (Charlson Comorbidity Index), blood hemoglobin levels, whether the patient was admitted through the emergency room, age, and kidney function as measured by creatinine levels. Notably, patients admitted through the emergency room had a much higher death rate (8.3%) compared to those in the general internal cohort (3.7%), highlighting the severity of acute presentations. Interestingly, different algorithms worked best for different types of CAD—XGBoost for acute cases and traditional logistic regression for chronic cases—suggesting that the nature of the disease matters when choosing a predictive tool.
This research matters because early identification of CAD patients at highest risk of dying could allow doctors to prioritize intensive monitoring and treatment resources for those who need them most. By using routinely collected electronic health record data, this type of model could potentially be integrated into hospital systems without requiring additional testing. However, the authors note this was a retrospective study at specific hospitals in Taiwan, and further validation in broader populations would be needed before such models could be widely adopted in clinical practice.
Melisa S, Phuc P, Chien S, Chiang K, Hsu M, Nguyen P, et al.. (2026). Machine learning prediction of 30-day mortality in coronary artery disease: a retrospective multicenter study using electronic health records.. International journal of cardiology. https://doi.org/10.1016/j.ijcard.2026.134745