Predição de infarto agudo do miocárdio e insuficiência cardíaca com aprendizado de máquina - uma análise comparativa no MIMIC-IV

Abstract

Cardiovascular diseases are among the leading causes of morbidity and mortality, particularly acute myo cardial infarction and heart failure, whose clinical severity calls for early identification of at-risk patients. This study developed and compared four supervised classifiers (Logistic Regression, Support Vector Machine, Decision Tree, and K-NearestNeighbors)topredictCVDusingMIMIC-IVdata. Atotalof546,028hospitaladmissionswereanalyzed,witha binary outcome defined by ICD-9/ICD-10 codes; 16.3% were positive cases. Features comprised demographic variables, vital signs, and laboratory tests; preprocessing included removal of physiologically implausible values, median imputa tion, standardization, and a stratified split into training (70%) and testing (30%). Models were tuned via grid search with stratified cross-validation and assessed using accuracy, precision, recall, F1-score, and AUC-ROC, complemented by con fusion matrices and interpretability analyses. The decision tree achieved the best overall balance (AUC-ROC 0.868; recall 0.788; lowest false negatives, 5,664), while the SVM was competitive at higher computational cost. Logistic regression provided stable performance with coefficient-based interpretability, and KNN, despite the highest accuracy, showed low recall, underscoring the limits of accuracy under class imbalance. Interpretability results highlighted age, urea, troponin T, and NT-proBNP as the most influential variables, supporting the clinical plausibility of the learned patterns.