P1945Machine learning to identify high-risk clusters of patients undergoing cardiac resynchronization therapy

Background: Cardiac Resynchronization Therapy (CRT) has been shown to improve mortality and cardiac function in patients with chronic systolic heart failure (HF) and prolonged QRS. However, the beneficial response is multifactorial. Consequently, a reliable method is warranted to identify vulnerable patients by risk stratification for long-term mortality. Purpose: We sought to generate an unsupervised machine learning model based on baseline clinical characteristics of pre-procedural assessment to risk-stratify patients undergoing CRT implantation. Methods: After CRT implantation, 958 patients (67±10 years, 244 [25%] females) were followed up for median of 1325 days. Prior to implantation, all patients underwent comprehensive clinical evaluation including assessment of cardiovascular risk factors, laboratory tests, and echocardiographic examination. Lasso penalty regularized Cox Proportional Hazard model was used to identify features with high predictive power to perform hierarchical clustering. Results: During the follow up period 527 (55%) deaths occurred. The selected parameters included gender (male, Hazard Ratio [HR]=1.62, 95% Confidence Interval [CI], 1.27–2.08, p<0.001), age (HR=1.03, 95% CI, 1.02–1.04, p<0.001), valvular heart disease (HR=1.56, 95% CI, 1.19–2.05, p=0.001), diabetes mellitus (HR=1.23, 95% CI, 1.03–1.47, p=0.021), ischemic etiology (HR=1.35, 95% CI, 1.08–1.68, p=0.008), hemoglobin (HR=0.91, 95% CI, 0.87–0.96, p<0.001), blood urea nitrogen (HR=1.03, 95% CI, 1.01–1.05, p<0.001), early diastolic transmitral flow velocity (HR=1.01, 95% CI, 1.01–1.02, p=0.007), late diastolic transmitral flow velocity (HR=0.99, 95% CI, 0.98–0.99, p=0.005), CRT-Defibrillator (HR=0.72, 95% CI, 0.59–0.88, p<0.001), digoxin- (HR=1.44, 95% CI, 1.18–1.76, p<0.001), and statin usage (HR=1.23, 95% CI, 1.01–1.50, p=0.035). The unsupervised clustering algorithm identified two distinct clusters. The survival curves of the clusters (Figure 1) differed significantly (Log-rank test, p<0.001). The risk of mortality more than doubled in patients in the second cluster (HR=2.19, 95% CI, 1.83–2.62, p<0.001). Patients in the first cluster was found to have higher left ventricular ejection fraction (cluster 1 vs cluster 2, 30±7 vs 28±7%, p<0.001) and Tricuspid Annular Plane Systolic Excursion (22±5 vs 18±5mm, p<0.001), but lower N-terminal pro B-type Natriuretic Peptide values (2264±2333 vs 3155±4017 pmol/L, p=0.013).