Bridging accuracy and explainability in AI-ECG: A multi-stage hyperparameter tuning framework for reinforcement learning-based random forests.

BACKGROUND

Deep learning applied to ECG has demonstrated strong diagnostic performance for acute coronary syndrome (ACS), but widespread clinical adoption remains limited by lack of interpretability and model robustness. Classical machine learning approaches, such as random forests (RF), offer greater explainability but often underperform compared with deep learning models.

OBJECTIVE

To develop and evaluate a structured, multi-stage hyperparameter optimization framework for a reinforcement learning-enhanced RF (RL-RF) model to improve the accuracy, stability, and interpretability of ACS classification from 12‑lead ECGs.

METHODS

We analyzed 10,393 ECGs from 7397 patients (15.6% ACS prevalence), partitioned into training (80%), validation (10%), and test (10%) sets using stratified sampling. A five-stage optimization pipeline was implemented, including (1) initial hyperparameter exploration, (2) reinforcement learning-based optimization using a Light Double Deep Q-Network, (3) multi-local refinement on the full training set, (4) stability evaluation via multi-seed averaging, and (5) final ensemble construction. Model performance was compared against conventional RF and untuned RL-RF baseline using area under ROC curve, average precision (AP), and F1-score. Performance metrics were computed with a threshold that maximized the F1-score for each model.

RESULTS

The tuned RL-RF model achieved superior performance on the independent test set, with an AUROC of 0.86, AP of 0.69, and F1-score of 0.58. The model demonstrated improved specificity and positive predictive value compared with baseline models, indicating enhanced ability to reduce false-positive classifications, with a modest decrease in sensitivity.

CONCLUSIONS

A multi-stage reinforcement learning-guided optimization framework significantly improves the performance and stability of explainable RF models for ACS classification. This approach bridges the gap between accuracy and interpretability, supporting the development of clinically trustworthy AI-ECG systems.