Nine-feature model predicts visual impairment in Saudi diabetes cohort

A machine-learning model based on nine selected health, demographic and lifestyle indicators achieved 95% accuracy and a 0.96 area under the curve for predicting visual impairment among people with type 2 diabetes at King Abdulaziz University Hospital in Saudi Arabia. The retrospective cross-sectional study analyzed medical records from 1,000 patients and found that a feature-selected CatBoost model performed better than comparison models, although the results do not establish how the tool would perform outside this hospital or in routine clinical care.

Researchers divided the records into two analytical groups: a full dataset containing patients with hemoglobin A1c, or HbA1c, levels of at least 6.5%, and a clinically relevant high-risk subgroup with HbA1c above 8%. The reduced subgroup incorporated 19 demographic, lifestyle and lipid-profile features. Models developed with this reduced dataset consistently performed better across the algorithms evaluated. The strongest result came from a CatBoost model whose features were selected with the Whale Optimization Algorithm, abbreviated WOA, and used nine indicators.

Within the high-risk subgroup, WOA-CatBoost achieved an accuracy of 0.95 and an area under the receiver operating characteristic curve, or AUC, of 0.96, with a 95% confidence interval of 0.94 to 0.98. A Random Forest model using features selected through the Black Widow Optimization Algorithm, or BWO, reached 0.92 accuracy and an AUC of 0.90, with a 95% confidence interval of 0.87 to 0.93, using 10 features. The baseline multilayer perceptron reached 0.91 accuracy and an AUC of 0.94, with a 95% confidence interval of 0.91 to 0.97, using 13 features. Models were developed with an 80/20 stratified train-test split, and five-fold cross-validation was conducted within the training set for internal validation.

Paired comparisons of the five-fold cross-validation accuracies further favored WOA-CatBoost. Bonferroni-corrected paired t-tests, using a significance threshold of 0.017, identified a statistically significant performance advantage over BWO-Random Forest, with a t value of 10.17 and p below 0.001. Predictors recurring across the three best-performing models were HbA1c, body mass index, age, occupation, triglycerides, type of food and smoking. The supplied findings do not report individual feature-importance scores or coefficients, so the relative contribution of each predictor cannot be determined from the available evidence.

Because the research was retrospective and cross-sectional, it evaluated model performance using existing medical records rather than testing the system prospectively or measuring effects on patient outcomes. The single-hospital sample also limits how directly the findings can be generalized to other Saudi healthcare settings or populations. Within the studied cohort and analytical subgroup, however, the results show that WOA-CatBoost combined strong measured discrimination with a smaller feature set than the comparison models. The authors described the approach as a parsimonious predictive tool for visual-impairment risk stratification in type 2 diabetes within the Saudi healthcare context, while further validation would be needed to assess its usefulness beyond the analyzed records.

Subject of Research: Machine-learning prediction of visual impairment in a Saudi type 2 diabetes cohort

Article Title: Predicting visual impairment in type 2 diabetes using key performance indicators identified by the Black Widow and Whale Optimization Algorithms: A case study in Saudi Arabia

Article References: Al-hussein, F., Tafakori, L., Abdollahian, M., & Al-Shali, K. (2026). Predicting visual impairment in type 2 diabetes using key performance indicators identified by the Black Widow and Whale Optimization Algorithms: A case study in Saudi Arabia. PLOS One, 21(10), e0361022. https://doi.org/10.1371/journal.pone.0361022

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361022

Keywords: visual impairment, type 2 diabetes, machine learning, CatBoost, feature selection, HbA1c, risk prediction, Saudi Arabia, medical records, retrospective study, Predicting, visual