Machine Learning and Deep Learning for Predicting Response and Outcomes After Cardiac Resynchronization Therapy: A Systematic Review and Meta-Analysis of ECG-Based Models

Background/Objectives: Up to 40% of patients receiving cardiac resynchronization therapy (CRT) do not experience significant clinical improvement, highlighting the limitations of conventional selection criteria. Artificial intelligence (AI) is increasingly being applied to electrocardiogram (ECG) and other clinical data to improve patient selection. This systematic review and meta-analysis evaluated the performance and methodological quality of machine learning (ML) and deep learning (DL) models using ECG-derived information to predict CRT response and clinical outcomes. Methods: A systematic literature search was conducted in Scopus and Web of Science for studies published between January 2006 and April 2026. Eligible studies developed or evaluated ML or DL algorithms primarily using pre-implantation ECG data to predict CRT response, including changes in left ventricular ejection fraction (LVEF) or left ventricular end-systolic volume (LVESV), or mortality. Model performance and prognostic associations were synthesized using random-effects meta-analyses of the area under the receiver operating characteristic curve (AUROC) and hazard ratios (HRs), respectively. Methodological quality and risk of bias were assessed using the PROBAST+AI tool. Results: We identified 14 eligible studies containing 34 distinct model evaluations. Most models utilized digital 12-lead ECG waveforms or single representative beats. Across 26 model-outcome estimates, the exploratory pooled AUROC was 0.710 (95% CI: 0.660–0.759), with substantial heterogeneity (I2 = 94.7%; p < 0.001). Across 10 heterogeneous model–outcome estimates, the exploratory pooled HR was 1.75 (95% CI: 1.42–2.15; I2 = 83.3%), although differences in prognostic constructs and endpoints precluded interpretation as a common effect. ML models had a higher pooled AUROC than DL models (0.752 vs. 0.669; exploratory Wald test: difference = 0.084; 95% CI: 0.002–0.165; p = 0.044). Pooled HR estimates also differed (ML: 2.22; DL: 1.46; ratio = 1.51; p < 0.001). These indirect subgroup comparisons were limited by clinical and methodological heterogeneity and potential dependence among estimates. PROBAST+AI assessment identified a high overall risk of bias in 32 of the 34 models, with concerns concentrated primarily in the participants and data sources and analysis domains. Independent geographic external evaluation was reported in only two studies, and these evaluations did not both constitute independent validation of fully specified CRT prediction models. Conclusions: ECG-based ML and DL models demonstrate moderate discrimination for predicting CRT outcomes. However, substantial heterogeneity, high risk of bias and limited independent external validation currently preclude routine clinical implementation. Future research should prioritize standardized outcome definitions, independent external validation, and head-to-head comparisons within the same cohorts to establish incremental predictive and clinical utility. Registration: The review protocol was prospectively registered in PROSPERO (CRD420261448863).