Purpose: Accordingly, our aim was to design and validate a machine learning based risk stratification system to predict 2-year and 5-year mortality from pre-implant parameters of patients undergoing CRT implantation.
Methods: We trained two models separately to predict 2-year (model 1) and 5-year mortality (model 2). As training cohort of model 1 we used 2098 patients (67±10 years, 1574 [75%] males, 1026 [49%] CRT-P, 1072 [51%] CRT-D) undergoing CRT implantation. Out of this population 1650 patients (66±10 years, 1258 [76%] males, 899 [54%] CRT-P, 751 [46%] CRT-D) also had 5-year follow-up data available and they served as the training cohort for model 2. At the time of implantation, demographics, cardiovascular risk factors, medication and laboratory test results were assessed. Forty-seven pre-implant parameters were used to train the models. Our models were designed in a way to tolerate missing values. Among non-linear classifiers, random forest (number of trees: 200) demonstrated the best performance. We validated our models, along with the Seattle Heart Failure Model (SHFM), VALID-CRT risk score and EAARN score on an independent cohort of 136 patients (66±10 years, 110 [81%] males, 114 [84%] CRT-P, 22 [16%] CRT-D).
Results: There were 458 (22%) deaths in the 2-year, 879 (53%) deaths in the 5-year training cohort. In the validation cohort, there were 30 (22%) deaths at 2 years and 58 (43%) deaths at 5 years after CRT implantation. For the prediction of 2-year mortality, the Area Under the Receiver-Operating Characteristic Curve (AUC) for model 1 was 0.77 (95% Confidence Interval [CI]: 0.67-0.87; p=0.002), for SHFM was 0.54 (95% CI: 0.39-0.69; p=0.006), for EAARN was 0.57 (95% CI: 0.46-0.68, p=0.002), and for VALID-CRT was 0.62 (95% CI: 0.52-0.71; p=0.002). To predict 5-year mortality the AUC for model 2 was 0.85 (95% CI: 0.78-0.91; p=0.001), for SHFM was 0.62 (95% CI: 0.51-0.74; p=0.003), for EAARN was 0.61 (95% CI: 0.51-0.70, p=0.002), for VALID-CRT was 0.65 (95% CI: 0.56-0.74; p=0.002). The AUCs of the machine learning based models were significantly higher than the AUCs of the pre-existing scores (DeLong test, all p<0.05).
Conclusion: Our results indicate that machine learning algorithms can outperform the already existing linear model based scores. By capturing the non-linear association of predictors, the utilization of these state-of-the-art approaches may facilitate optimal candidate selection and prognostication of patients undergoing CRT implantation.