Elderly heart attack patients face complex risk profiles that traditional scoring systems struggle to assess accurately. Researchers at Tehran Heart Center have developed machine learning models that significantly outperform conventional risk stratification tools, potentially enabling more personalized care for this vulnerable population undergoing percutaneous coronary intervention (PCI).
Key Points
- The study evaluated eight machine learning algorithms to predict one-year major adverse cardiovascular events (MACE) in 1,358 elderly patients (aged 65 years and older) who underwent PCI for ST-elevation myocardial infarction between 2015 and 2021.
- Random Forest and XGradient Boosting models achieved exceptional discrimination with area under the curve (AUC) values of 95% and 94% respectively, addressing documented limitations of traditional risk scores like TIMI and GRACE, which prior research has shown achieve AUCs of 85% and 78% in general AMI populations and demonstrate suboptimal performance in geriatric patients due to their heterogeneous risk profiles.
- Key predictive features identified through SHAP analysis included pre-PCI ejection fraction, age, creatinine levels, fasting blood sugar, body mass index, and LDL/HDL cholesterol ratio, with Random Forest demonstrating 79.3% sensitivity and 96.7% specificity.
- This study is limited by its retrospective, single-center design without external validation, survival bias potentially excluding the frailest patients, and lack of data on medication adherence and frailty indices that could influence outcomes.
Machine learning models demonstrate strong capability for personalized risk prediction in elderly post-PCI patients, though external validation and prospective studies are needed before clinical implementation can support individualized treatment strategies.
The Data
- The retrospective cohort study used Synthetic Minority Oversampling Technique (SMOTE) to address the class imbalance problem, with only 152 patients (11.2%) experiencing MACE during follow-up.
- Random Forest achieved the highest overall performance with an AUC of 0.953 (95% CI: 0.943–0.964), accuracy of 90.9%, precision of 92.3%, and specificity of 96.7%, while maintaining reasonable sensitivity at 79.3%.
- XGradient Boosting showed robust performance with an AUC of 0.941 (95% CI: 0.930–0.950), demonstrating a balanced trade-off between precision (84.9%) and sensitivity (77.5%) with 93.5% specificity.
- Models were validated using 5-fold cross-validation with feature selection performed within each fold to prevent information leakage, and all performance metrics represent averaged results across validation folds.
- Traditional models performed substantially worse, with Logistic Regression, Neural Networks, and Support Vector Machines achieving AUCs around 0.81, with Logistic Regression and SVM showing sensitivity below 53% and Neural Networks at 59.2%, highlighting limitations of linear approaches for this complex prediction task.
- Study limitations included the class imbalance with only 11.2% MACE incidence, potential survival bias excluding early deaths, lack of medication adherence data, absence of inflammatory markers, single-center design limiting generalizability, and no external validation to confirm model performance across different healthcare settings.
Industry Context
By combining high-performance models with SHAP-based explanations, this approach supports transparent, clinically actionable risk stratification and personalized care.
Amir Ghaffari Jolfayi and colleagues, Tehran Heart Center
This research addresses a critical gap for elderly patients, who constitute over 60% of acute myocardial infarction cases yet remain underrepresented in clinical trials. Traditional risk scores like TIMI and GRACE were developed primarily on younger populations and demonstrate suboptimal performance in elderly patients. Kwon and colleagues (2019) showed deep learning models achieved AUCs of 0.905 versus 0.851 for GRACE and 0.781 for TIMI in general populations, but prior studies lacked elderly-specific focus or interpretability analysis.
The research received ethical approval (IR.TUMS.THC.REC.1399.045) with no funding sources or conflicts declared. Using SHAP analysis to make models interpretable, the study revealed pre-PCI ejection fraction, creatinine, and fasting blood sugar as the strongest predictors. Next steps include multicenter validation, prospective trials, and integration into clinical decision support systems.
The study, “Applied machine learning to predict 1-year major adverse cardiovascular events in elderly patients after percutaneous coronary intervention,” was published in BMC Medical Informatics and Decision Making, October 2025 (DOI: 10.1186/s12911-025-03238-7).
