A Review of Artificial Intelligence-Based Diabetes Prediction

Using Machine Learning, Deep Learning, and

Optimizer-Assisted Decision Support

Ancy Cheriyan 1,*

1 School of Artificial Intelligence, Bahrain Polytechnic ,PO Box 33349, Isa Town, Bahrain

Received: May 14, 2026 Revised: June 29, 2026 Accepted: August 22, 2026 ⋆ Corresponding author

ABSTRACT

Diabetes prediction has become an important direction in clinical artificial intelligence because early identification of

high-risk individuals can support preventive intervention, personalized monitoring, and improved long-term healthcare

outcomes. This review examines recent machine learning, deep learning, and optimizer-assisted methodologies

for diabetes prediction and related diagnostic decision support. It discusses how clinical variables, physiological

measurements, biomedical signals, retinal images, population-specific attributes, and healthcare monitoring data can

be transformed into predictive evidence through preprocessing, feature engineering, feature selection, model training,

hyperparameter optimization, and risk interpretation. The review also emphasizes major methodological concerns,

including class imbalance, model robustness, interpretability, subgroup variation, complication-oriented prediction,

and deployment within connected healthcare environments. Overall, the reviewed literature indicates that effective

diabetes prediction should be developed as an integrated clinical decision-support pipeline rather than as a single

classifier comparison, with careful attention to data quality, feature relevance, model transparency, and practical

healthcare usability.

Keywords: Diabetes prediction Machine learning Deep learning Metaheuristic optimization Clinical decision

support

1. INTRODUCTION

Diabetes mellitus has become a major clinical and computational

challenge because its early identification depends

on the accurate interpretation of heterogeneous biomedical

indicators, lifestyle-related attributes, physiological measurements,

and patient-specific risk patterns. In this context, artificial

intelligence has increasingly been adopted as a decisionsupport

direction for improving early diabetes prediction, reducing

diagnostic uncertainty, and enabling more responsive

screening strategies before severe complications emerge. Recent

studies have shown that diabetes prediction is no longer

limited to conventional tabular classification, but is expanding

toward hybrid machine learning frameworks that integrate

optimized classifiers, signal-derived biomarkers, and

population-specific modeling strategies to improve diagnostic

reliability across different patient groups [1]. This direction is

particularly important because type 2 diabetes classification

may be affected by demographic variation, sex-related differences,

and population-specific clinical distributions, which

makes generalized prediction models insufficient when they

are not carefully adapted to the characteristics of the target

cohort [2].

A central issue in diabetes prediction is the selection of clinically

informative features from datasets that may contain