Interpretable hybrid-ensemble acute myocardial infarction diagnosis with clinical transparency

Abstract

Background

Single-algorithm machine learning models for acute myocardial infarction (AMI) often struggle with heterogeneous clinical data and lack transparency, limiting their clinical adoption.

Methods

We propose a novel interpretable hybrid-ensemble AMI diagnosis method with clinical transparency. Our method performs nested iteration and global hyperparameter optimization via data reuse, which can improve the generalizability of single-algorithms to a certain extent. By balance criterion between ensemble robustness and performance optimization, some improved pre-ensemble algorithms and optimal ensemble strategy are determined to coordinate multiple algorithms in constructing our hybrid-ensemble model. A game-theoretic marginal contribution analysis (GMCA) framework was established to quantify the feature contribution of prediction results in global and local levels，which can improve the interpretability and clinical transparency of our hybrid-ensemble model. Based on the framework, data-driven dimensionality reduction of clinical feature is performed, while ensuring clinical effectiveness through expert guided validation. Additionally, a series of extended experiments were conducted to further validate the adaptability and generalizability of our proposed method.

Results

Our hybrid-ensemble model achieved 0.944 accuracy and 0.94 F1-score in binary classification, outperforming single-algorithms. The GMCA framework identified six features, achieving dual optimization of data-driven selection and domain knowledge consistency. Extended experiments confirm the model's robust accuracy and generalizability across diverse configurations.

Conclusions

Our study presents a replicable method for developing transparent and clinically actionable AI systems in emergency cardiology. By harmonizing algorithmic diversity with interpretability, our hybrid-ensemble method not only enhances diagnostic accuracy but also fosters clinician trust.