Developing An Artificial Intelligence – Machine Learning Model To Identify Patients Who Have Or Are At Risk Of Type 2 Diabetes Using Their Clinical Data And Retinopathy Grade In The Absence Of A Blood Test In Kenya. .

Diabetes and its complications are a global burden due to their rising prevalence. Lack of awareness, especially in low- and middle-income countries (LMICs), leads to many people with undetected diabetes, which can cause irreversible health issues like retinopathy. AI in healthcare can cut diagnostic costs by 50%, improve health outcomes by 40%, and speed up diagnosis. Inequity in healthcare technology is shown by the lack of Black African data training in most AI diabetes detection tools. This study uses AI and machine learning to predict high-risk Type 2 Diabetes patients in LMICs for timely treatment to prevent serious health issues and sight loss.

A cross-sectional analytical study was conducted at Lions Sightfirst Eye Hospital (Nairobi, Kenya).

Data from eligible, consenting subjects (n=599) was collected using a form and then anonymised. Data collected included blood pressure, glucose, BMI, waist-hip ratio, pulse, demographics, family history, and physical activity, as well as fundus photography. A total of 1530 fundus images were collected and encrypted. Images were preprocessed and features extracted with RETFound (ViT, Nature 2023). Four models—clinical-only, image-only, fusion XGBoost, and fusion neural network—were trained with strict patient-level splits (n=90 test). An AI review layer validated predictions to identify associations and predict Type 2 Diabetes risk.

 The fusion XGBoost model achieved AUC 0.885, 83.3% accuracy, 88.4% sensitivity, and 78.7% specificity on 90 held-out test patients. In a sample of 43 diabetic patients, it detected 38 diabetics, missing only 5.

Based on clinical history alone, the model achieved AUC 0.860 (sensitivity 65.1%), and using fundus images alone, it achieved AUC 0.820 (sensitivity 74.4%) thus confirming that retinal and clinical data are complementary. The Claude AI clinical review agreed with the model in 86.7% of cases (78/90) and independently flagged 12 patients for further review, including one case where it correctly identified a model error.

Based on the model results, non-invasive diabetes screening using fundus photography and patient history is feasible and effective in a Black African population. The use of a multimodal fusion approach outperformed either data source alone, showing that retinal vascular changes complement clinical risk factors.

With 88.4% sensitivity, the system could identify nearly 9 in 10 diabetic patients for confirmatory testing without a blood draw. Further efforts will include fine-tuning the retinal AI backbone for the general population to improve image-based detection, prospective validation, and multi-site expansion across East Africa to establish generalisability for community-level deployment.