Interpretable machine learning models for arrhythmia risk assessment in patients with cardiac-related mutations

Introduction

Rare cardiac mutations are the leading cause of sudden cardiac death (SCD) in children and young adults. Due to its high risk and limited risk stratification, prophylactic implantable cardioverter-defibrillators (ICDs) are often used but carry surgical and device-related risks. These challenges highlight the need for methods to identify genuinely at-risk patients. Electrocardiography (ECG), particularly heart rate variability (HRV), offers a non-invasive tool with proven predictive value for arrhythmic events.

Methods

We analyzed 118 ECG recordings from patients with high-risk cardiac mutations. Ten-minute segments were extracted at consistent times to account for circadian variations. HRV features were computed using a phased approach—first focusing on the most prevalent mutation, then grouping by phenotype, and finally combining all mutations. Emphasizing interpretability, we tested multiple AI models and combined them in several ensemble frameworks, introducing clinical metrics to protect high-risk patients while limiting unnecessary ICDs.

Results and Discussion

In four scenarios—single-mutation (KCNH2), pathophysiological (potassium-channel), phenotypic (Long QT Syndrome), and the entire dataset—models varied in identifying at-risk patients. Single-mutation analysis showed robust AUCs (>0.85) for Multi-Layer Perceptron model and a voting ensemble, achieving high recall while identifying half of the non-risk group and capturing all high-risk cases. These models remained strong (AUC >0.80) for potassium-channel mutations, where logistic regression also excelled. In Long QT Syndrome, performance dipped, likely due to phenotypic heterogeneity. Across all mutations, the models reached AUCs of ~0.85–0.90. Setting a 20% cutoff, Logistic Regression emerged as the optimal model, achieving an effect index of 0.38 while maintaining high recall, thereby effectively balancing patient safety with minimizing unnecessary ICD implantations.

Conclusions

Scenario-specific training is crucial for SCD risk prediction. While maximizing recall protects at-risk patients, adjusting cutoffs can spare low-risk individuals from unnecessary ICDs. Larger, more detailed datasets could further refine model performance, enabling more precise, individualized care for rare cardiac mutations.