Introduction: Machine learning methods such as cluster analysis can identify commonality in patterns of short-term response measures after cardiac resynchronization therapy (CRT) to predict classes of patients with distinct long-term prognoses.Hypothesis: Distinct response clusters identified within 6 months of CRT implantation would provide independent prognostic value relative to known pre-CRT patient characteristics.Methods: Patients with heart failure (HF) undergoing CRT had assessments of left ventricular end-systolic volume fractional change (LVESV-FC), peak VO2, and B-type natriuretic peptide (BNP) based on cardiac magnetic resonance (CMR), echocardiography, exercise testing, and blood tests before and 6 months after CRT. Statistical methods included multivariate multiple linear regression, cluster analysis based on a mixture model, survival analysis, and receiver operating characteristic (ROC) analysis.Results: During a median of 5.0 years of follow-up after CRT, the cohort of 146 patients (age 66.0 ± 11.3 years, 34.9% female) had a death rate of 28.1%. A significant correlation was observed for BNP response and LVESV-FC (r=0.42, p>0.01), but not for the other response comparisons. Three clusters of patients (1: n=27; 2: n=82; 3: n=37) were identified. Kaplan-Meier analysis (Figure) demonstrated the best long-term survival in cluster 2, intermediate survival in cluster 3, and the worst survival in cluster 1 (p<0.0001). ROC curve comparisons for 4-year survival based on pre-CRT findings with or without the 6-month response cluster showed that the cluster increased the AUC from 0.818 to 0.870 (p=0.069).Conclusions: Response clusters based on 6-month parameters were strongly associated with long-term survival and improved prognostication compared with just pre-CRT predictors alone. This response clustering approach based on machine learning promises to be very useful for clinical risk stratification in heart failure after CRT.