Machine learning based risk stratification of patients undergoing ventricular tachycardia ablation

Abstract Background Monomorphic ventricular tachycardia (VT) is a potentially life-threatening condition. Although radiofrequency catheter ablation represents an effective treatment method for many of these patients, significant variability is observed in postprocedural mortality, which is attributable to multiple factors, including the high burden of comorbidities. Therefore, there is a great demand for an accurate risk stratification system. Purpose We sought to implement a machine learning pipeline to predict 1-year all-cause mortality in patients undergoing VT ablation. Methods For 265 consecutive patients who underwent VT ablation at our center, we retrospectively collected demographics, medical history, cardiovascular risk factors, laboratory results, echocardiographic measurements, and VT ablation-related parameters. To predict 1-year all-cause mortality based on these features, several supervised machine learning models were trained and evaluated using 5-fold cross-validation. We applied a recursive elimination technique to identify the optimal subset of input features. The area under the receiver operating characteristic curve (AUC) with a 95% confidence interval (CI) was calculated to quantify the models' performance. We also identified the most important predictors of mortality using Shapley values. As the final step, we used topological data analysis to discern and visualize patient subgroups with different mortality risk. Results 57 (22%) patients died during the 1-year follow-up period. In predicting all-cause mortality, the best performance was achieved by a random forest model utilizing 18 input features [AUC: 0.73 (95% CI: 0.68–0.78)]. This model significantly outperformed other previously published risk scores such as the I-VT [AUC: 0.63 (95% CI: 0.55–0.70), p<0.001 vs. random forest] or the PAINESD [AUC: 0.63 (95% CI: 0.55–0.71), p=0.009 vs. random forest]. The most important predictors of mortality were mitral E-wave deceleration time, cardiac resynchronization therapy, age, electrical storm, and hemoglobin concentration. In the topological network created based on the 18 input features of the best-performing random forest model, we could identify five patient subsets with different clinical characteristics and 1-year mortality rates (Figure 1). Conclusions Our machine learning model could efficiently predict 1-year all-cause mortality in patients undergoing VT ablation. Thus, it could facilitate the prompt identification of high-risk patients and the personalization of treatment and follow-up strategies, ultimately leading to improved outcomes. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Heart Program, as part of the National Research, Development and Innovation Fund of HungaryThematic Excellence Programme of the Ministry for Innovation and Technology in Hungary