A Systematic Literature Review of Machine Learning and Explainable AI Approaches for Diabetes Prediction: Taxonomy, Gaps, and Future Directions (2020–2025)

Diabetes mellitus is a rapidly escalating global public health crisis affecting over 537 million adults worldwide, with projections suggesting 783 million by 2045. The exponential growth of health data repositories and machine learning research has produced a vast body of literature on automated diabetes risk prediction. However, this literature suffers from recurring methodological inconsistencies — particularly over-reliance on a single benchmark dataset, data leakage from improper SMOTE application, absence of threshold optimisation for real-world imbalanced distributions, and shallow explainability that hinders clinical adoption. This paper presents a systematic literature review of 34 studies published between 2020 and 2025, covering classical machine learning, ensemble methods, deep learning architectures, and explainable AI frameworks applied to diabetes prediction. We propose a six-category taxonomy of approaches, analyse comparative performance across datasets, and identify eight specific research gaps that remain unaddressed in the existing literature. We also present our own empirical work — a SMOTE-enhanced hybrid RF+MLP stacking framework validated across Pima Indian and BRFSS 2015 datasets achieving consistent AUC of 0.808 and 0.816 respectively with 90.7% clinical recall — as a representative study addressing several identified gaps. Based on the review, we outline seven concrete future research directions essential for bridging the gap between research prototypes and clinically deployable diabetes screening tools.