Toward Reliable ECG Arrhythmia Classification: A Quantile-Based Selective Prediction Approach

Abstract

CG arrhythmia classification is one task that has shown to be well suited for deep learning techniques, but the predictions are commonly deterministic and this is not desirable when it comes to safety-critical clinical applications. This limitation is overcome by designing a complementary, uncertainty-aware framework for selective prediction for the ECG arrhythmia classification task to boost the robustness of the classification without re-training the underlying model. Variance estimates derived from Monte Carlo dropout are used to quantify the predictive uncertainty, and an adaptive thresholding strategy based on quantiles is used to identify the unreliable predictions. The proposed framework is assessed using a deep learning arrhythmia classifiers prediction level output on the basis of Risk–Coverage analysis, comparisons of accepted versus rejected predictions, simulation based evaluations of class-wise coverage for confident vs non-confident predictions, and for simulations aimed at deployment. Empirical findings show that the classification risk is consistently reduced with decreasing coverage with selective prediction and accuracy and F1-score are significantly higher for the accepted predictions. In calibration analysis, there is also a good match between the predicted level of confidence and the actual accuracy, and deployment simulations reveal the dramatic impact of small declines in automation rates on error reduction. The results showed that uncertainty-aware selective prediction is an effective and model-agnostic solution for enhancing the reliability and clinical trustworthiness of arrhythmia classification systems, while allowing the safe use of AI in clinical practice.