A novel machine learning based algorithm to predict 1-year mortality after ventricular tachycardia ablation

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): -Project no. NVKP_16-1–2016-0017 (’National Heart Program’) has been implemented with the support provided by the Ministry of Innovation and Technology of Hungary from the National Research, Development and Innovation Fund of Hungary, financed under the NVKP_16 funding scheme. -Project no. MEC_R-141548 has been implemented with the support provided by the Ministry of Innovation and Technology of Hungary from the National Research, Development and Innovation Fund, financed under the MEC_R_21 funding scheme. Background Catheter ablation constitutes an established therapeutic option in patients with monomorphic ventricular tachycardia (VT). VT patients have high comorbidity burden and high mortality, however there is big heterogenity among individual patients. Therefore, effective postprocedural risk assessment systems, prediciting the mortality of patients undergoing VT ablation are highly needed. Purpose We aimed to develop a risk stratification algorithm predicting the 1-year all-cause mortality of patients undergoing VT ablation, and to identify the most important input factors of the model. Methods Between 2005 and 2020, 272 consecutive patients underwent VT ablation due to sustained monomorphic VT at our institution. We processed their procedural, demographic and medical history data, in addition to their laboratory and echocardiographic findings. For the training of different supervised learning models, we used 63 pre-procedural and procedural variables. We performed 5-fold cross validation and calculated the area under the receiver operating characteristic (ROC) curve (AUC), to assess the performance of the models. Finally, with calculating Shapley values we determined the most important factors of the mortality prediction for the best performing model. Results After a follow up of one year, total all-cause mortality was 22% (59). In predicting 1-year mortality, the best performance was shown by the random forest model [AUC: 0,73 (0,68-0,78)] among the machine learning models we had trained. This model significantly outperformed the traditional score systems like I-VT [AUC: 0,63 (0,55-0,70) vs. 0,73 (0,68-0,78), p<0,001] and PAINESD [AUC: 0,63 (0,55-0,71) vs. 0,73 (0,68-0,78), p=0,009] in our dataset. The predictive factors with the biggest effect on mortality were mitral E wave deceleration time, presence of cardiac resynchronization therapy, age, electric storm and hemoglobin concentration. Conclusion We were able to establish a supervised machine learning based system which managed to predict the 1-year mortality of VT ablation patients with high accuracy and turned out to be superior when compared to the use of existing risk scores. This enables us to identify the patients in need of a more thorough follow-up, which could reduce their mortality.