In-hospital Mortality Risk Prediction for Myocardial Infarction Patients: A Complete Machine Learning Pipeline for Real-world Clinical Scenarios

This article presents an end-to-end machine learning project for predicting in-hospital mortality risk in myocardial infarction patients for real-world clinical scenarios. It focuses on addressing challenges in clinical data such as class imbalance, high-dimensional sparse features, multicollinearity of physiological indicators, and non-random missing data, compares the application effects of various regularization methods and nonlinear modeling techniques, and demonstrates responsible medical AI practices.

Myocardial infarction (MI) is one of the leading causes of death globally. Early and accurate assessment of patients' in-hospital mortality risk is crucial for treatment plan formulation, medical resource allocation, and prognosis improvement. This project addresses challenges in clinical data science by building a complete machine learning pipeline, demonstrating responsible ML practices under real clinical constraints.

The HOSP_ADMIT dataset is used, covering features such as demographics, medical history, physiological indicators, and electrocardiogram results. Core challenges include: class imbalance (84% of patients survived), high-dimensional sparse categorical features (ECG leads), multicollinearity of physiological indicators, and non-random missing data (MNAR).

A leak-proof preprocessing strategy is adopted. Missing values are handled via ColumnTransformer, and missing value indicators are combined to retain biological signals. All preprocessing steps are stateful designs—fitted on training data and reused—to avoid data leakage and ensure model generalization ability.

Various modeling methods are compared: regularized linear models (L1/L2 to handle EPV constraints and multicollinearity), generalized additive models (GAMs to capture nonlinear physiological risk profiles), and random forests (to capture feature interaction effects), providing a basis for clinical feature selection and risk understanding.

A clinically oriented evaluation framework is used: PR-AUC (sensitive to class imbalance, focusing on positive case identification), Brier score (evaluating probability calibration), and stratified cross-validation (ensuring consistent class proportions), avoiding misleading results from traditional accuracy.

Engineering-wise, modular structure, state persistence (export and reuse of preprocessors and models), and complete documentation are adopted. Insights include: preprocessing needs to encode clinical knowledge, model selection should serve clinical problems, and evaluation metrics should align with business goals.

This project demonstrates a complete ML pipeline for real-world clinical scenarios. While addressing technical challenges, it embodies a responsible medical AI attitude, providing a reference practical example for medical ML applications.