Transparent Prediction of Diabetic Retinopathy Using Machine Learning and Fuzzy Logic

Abstract

Diabetic retinopathy (DR) is a major cause of preventable blindness, and with its global prevalence rising rapidly, automated grading of fundus images has become an integral part of screening programs to enhance timeliness and coverage. However, the clinical implementation of machine learning (ML) and deep learning (DL) systems is often limited by insufficient transparency, as it is crucial for ophthalmologists and healthcare stakeholders to have an understandable rationale linked to retinal findings before they can trust algorithmic predictions in safety-critical situations. The objective of the present review is to summarize and organize the state of the art in transparent prediction of DR from fundus photographs, pursuing two complementary pathways: (i) explainable artificial intelligence (XAI) applied to ML/DL models, and (ii) fuzzy logic (including fuzzy inference and hybrid neuro-fuzzy schemes) enabling rule-based linguistic decision-making. The novelty of this survey lies in bringing together these different strands under a common transparency umbrella and deriving a practical taxonomy for fundus-based DR detection/grading. A systematic literature review methodology based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) was adopted. Relevant studies were obtained from major scholarly databases, screened using predefined inclusion/exclusion criteria (fundus modality, DR grading/detection focus, transparency via XAI or fuzzy rules), and synthesized using qualitative data extraction fields covering dataset usage, modeling choices, explainability mechanisms, and reported results. The reviewed evidence suggests a prevailing tendency towards post-hoc explanations for high-performing DL models (e.g., heatmap and attribution-based methods), alongside lesion-centered and rule-extraction strategies that align better with clinical reasoning. Fuzzy and hybrid fuzzy-DL approaches offer inherently interpretable rule bases but face challenges in feature design, scalability, and standardized benchmarking. Key gaps include inconsistent reporting, limited clinical validation, and inadequate integration of domain knowledge into transparent predictive pipelines.