Sex-Specific Patterns of Mortality Predictors Among Patients Undergoing Cardiac Resynchronization Therapy: A Machine Learning Approach

Abstract

Background: The relative importance of variables explaining sex-related differences in outcomes is scarcely explored in patients undergoing cardiac resynchronization therapy (CRT). We sought to implement and evaluate machine learning (ML) algorithms for the prediction of 1- and 3-year all-cause mortality in CRT patients. We also aimed to assess the sex-specific differences in predictors of mortality utilizing ML. Methods: Using a retrospective registry of 2,191 CRT patients, ML models were implemented in 6 partially overlapping patient subsets (all patients, females, or males with 1- or 3-year follow-up). Each cohort was randomly split into training (80%) and test sets (20%). After hyperparameter tuning in the training sets, the best performing algorithm was evaluated in the test sets. Model discrimination was quantified using the area under the receiver-operating characteristic curves (AUC). The most important predictors were identified using the permutation feature importances method. Results: Conditional inference random forest exhibited the best performance with AUCs of 0.728 (0.645-0.802) and 0.732 (0.681-0.784) for the prediction of 1- and 3-year mortality, respectively. Etiology of heart failure, NYHA class, left ventricular ejection fraction, and QRS morphology had higher predictive power, whereas hemoglobin was less important in females compared to males. The importance of atrial fibrillation and age increased, while the importance of serum creatinine decreased from 1- to 3-year follow-up in both sexes. Conclusions: Using ML techniques in combination with easily obtainable clinical features, our models effectively predicted 1- and 3-year all-cause mortality in CRT patients. Sex-specific patterns of predictors were identified, showing a dynamic variation over time.

Keywords: cardiac resynchronization therapy; heart failure; machine learning; mortality prediction; sex differences.

Figure 1

The schematic outline of the data analysis pipeline. The data analysis pipeline included three major steps: (1) data pre-processing, (2) machine learning model development and evaluation, and (3) the calculation of feature importances. During data pre-processing, feature selection was performed, patients with a high proportion of missing data were excluded, missing values were imputed using MICE, and z-transformation was performed. Then, machine learning models were implemented in the 6 partially overlapping subsets of patients (in all patients, females, or males of the 1- and 3-year cohorts). Before model training, each patient subset was split into training and test cohorts (80:20 ratio). Hyperparameter tuning was performed with 10-fold CV in each training cohort. Models' discriminatory power was estimated using the area under the receiver-operating characteristic curves. Each of the 6 models was retrained in the given training cohort, and its performance was evaluated in the corresponding test cohort. Finally, to identify the most important predictors of mortality in each subset, permutation feature importances were computed from each of the 6 final models. See text for further details. AUC, area under the receiver operating characteristic curve; CRT, cardiac resynchronization therapy; CV, cross-validation; MICE, Multiple Imputation by Chained Equations.

Figure 2

Flowchart illustrating the steps of patient selection. For each patient who underwent successful CRT implantation at our center, pre-implantation clinical characteristics and procedural parameters were collected retrospectively from paper-based or electronic medical records and entered to our structured database. After excluding patients with ≥30% missing values, machine learning models were implemented to predict 1- and 3-year all-cause mortality in the entire cohort, in males and females separately (altogether 6 separate binary classification tasks). CRT, cardiac resynchronization therapy.

Figure 3

Kaplan-Meier curves for males and females in the 1- (A) and 3-year (B) cohorts. Kaplan-Meier curve analysis illustrates the difference in the survival of male and female CRT patients during 1- and 3-year follow-up. Cox proportional hazards models were used to compute hazard ratios with 95% confidence intervals. Hazard ratios were adjusted for age (at implantation), QRS morphology, etiology of heart failure, the type of the implanted device, and the type of atrial fibrillation. CI, confidence interval; CRT, cardiac resynchronization therapy; HR, hazard ratio.

Figure 4

The most important predictors of 1- and 3-year all-cause mortality in patients undergoing CRT implantation. The importance of each feature was quantified with the permutation feature importances method, which measures the importance of a feature by calculating the mean decrease in the model's performance (area under the receiver-operating characteristic curve) after permuting its values 10 times (see text for further details). To keep the data comparable between the different models, we identified the top 5 predictors in each model and took the union of these features; then, we plotted the results on radar charts. AF, atrial fibrillation; LBBB, left bundle branch block; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association.

Figure 5

Effect of the most important features on the predicted probability of 1-year all-cause mortality in the training cohorts. The probability of death was calculated for each patient in the training cohort with 10-fold cross-validation. The predicted probability is plotted for each patient, and second-order polynomial trendlines are fitted to their values. *p < 0.05 vs. non-ischemic/non-LBBB morphology/NYHA class II/no AF, unpaired Student's t-test or Mann-Whitney U test. Abbreviations as in Figure 4.

Figure 6

Effect of the most important features on the predicted probability of 3-year all-cause mortality in the training cohorts. The probability of death was calculated for each patient in the training cohort with 10-fold cross-validation. The predicted probabilities are plotted for each patient, and second-order polynomial trendlines are fitted to their values. *p < 0.05 vs. non-ischemic/non-LBBB morphology/ NYHA class II/no AF, unpaired Student's t-test or Mann-Whitney U test. Abbreviations as in Figure 4.