Risk prediction of inappropriate implantable cardioverter-defibrillator therapy using machine learning

Abstract

Abstract Background Inappropriate shock from implantable cardioverter-defibrillators (ICDs) undermines quality of life and increases all-cause mortality risk. Various predictors of inappropriate ICD shock have also been indicated in previous studies. Nevertheless, no prediction model by machine learning has been reported. Purpose We developed ML models for evaluating the risk of inappropriate ICD therapy, and proposed easily calculable risk scoring. Methods In total, 182 consecutive cases implanted with ICD were enrolled. The predictive models were based on 16 statistically significant factors out of 189 items including electrocardiogram parameters and clinical features after ICD implantation. Statistical analysis and 14 ensemble learning methods were performed. Shapley additive explanation (SHAP) method was performed to extract important features in the best ML model, and easily calculable score was established from the features. Results Twenty-five patients (13.7%) underwent inappropriate therapy. The extra-trees classifier demonstrated the highest area under the receiver operating characteristic curve (AUROC; 0.839). The top six features were history of atrial arrhythmia, ischemic cardiomyopathy (ICM), no diabetes mellitus (DM), cardiac resynchronization therapy (CRT), and V4- and V5 R-wave amplitudes. From the six features, the Cardi45 score was developed (+6.5 points for history of atrial arrhythmia, -2.0 points for ICM, -1.0 points for DM, -0.5 points for CRT, and +0.5 points for V4 T-wave amplitude ≥1200 μV and V5 R-wave amplitude ≥1400 μV). The AUROC for the Cardi45 score was 0.886, which was higher than that of extra-trees classifier. Conclusion Our study demonstrates that the ML approach can provide a good prediction model and Cardi45 score was powerful predictors for inappropriate ICD therapy. Cardi45 score, derived from this model, may be used to judge whether an aggressive rhythm control strategy should be implemented or ICD implantation should be avoided in some patients.Picture 1Picture 2