A critical systematic review of ensemble learning methods for predictive healthcare

Abstract Ensemble learning has emerged as a significant paradigm in predictive healthcare, offering enhanced accuracy, robustness, and generalization relative to single-model methods. The systematic review critically examines current advancements in ensemble-based illness prediction across five major disease categories: heart disease, diabetes, liver disease, kidney disease, and skin disease. Applying the PRISMA method, 287 studies published from 2020 to 2024 were found in major scientific databases, establishing the first group. Subsequent to the comprehensive screening and eligibility assessment, 91 excellent papers were selected for careful examination. Statistical analysis of these studies reveals that approximately 40% employ stacking or hybrid ensemble models with feature selection. These models consistently demonstrate the highest performance, achieving accuracy in the range of 94–97%. 20% of studies are based on simple Bagging or Voting based ensemble methods. In contrast, about 20% of the studies adopt bagging or voting-based ensemble methods, which generally result in comparatively lower accuracy. Deep learning–based ensemble approaches account for approximately 12% of the reviewed studies. Overall, this review provides a comprehensive quantitative assessment of ensemble learning techniques in predictive healthcare. It highlights dominant methodological trends, comparative performance outcomes, and various practical challenges related to computational overhead, limited explainability, and dataset biasness.