Explainable Machine Learning for Diabetes Risk Prediction: Development of a Baseline Predictive Engine for the Wo’tosuga Digital Health Platform Targeting African Populations

Abstract Background Diabetes mellitus poses a growing public health burden across sub-Saharan Africa, with an estimated 54% of affected individuals undiagnosed. Early identification of at-risk individuals is critical to reducing complications, yet laboratory-based screening remains inaccessible to much of the population. Machine learning approaches offer a scalable, non-invasive alternative for population-level risk stratification; however, existing models are predominantly trained on non-African datasets, limiting their direct applicability to African healthcare contexts. This study aimed to develop and evaluate an explainable machine learning model for binary diabetes risk prediction as the foundational predictive engine of the Wo’tosuga digital health platform. Methods Data were obtained from the United States Behavioral Risk Factor Surveillance System (BRFSS) across three survey cycles (2019, 2021, 2023), yielding an analytical dataset of 869,319 respondents across 21 predictive features following preprocessing. Pre-diabetic cases were reclassified as diabetic to create a binary outcome reflecting the screening priority of identifying all at-risk individuals. Eight machine learning algorithms were evaluated with sensitivity designated as the primary evaluation criterion. Hyperparameter optimisation was conducted via GridSearchCV with 3-fold cross-validation, with recall as the optimisation metric. SHAP (SHapley Additive exPlanations) values were computed to characterise feature importance. Results LightGBM demonstrated the strongest overall performance, achieving a sensitivity of 0.807, specificity of 0.694, AUC-ROC of 0.824, and negative predictive value (NPV) of 0.912. Following hyperparameter optimisation and full-data retraining, the final LightGBM model achieved sensitivity of 0.809, AUC-ROC of 0.828, and NPV of 0.952, representing marginal but consistent improvement. XGBoost (tuned: sensitivity 0.806, AUC 0.824) and logistic regression (sensitivity 0.779, AUC 0.820) served as comparators. SHAP analysis identified age, self-reported general health, BMI, blood pressure status, and time since last medical checkup as the five most influential predictors, consistent across both the primary and baseline models. Conclusions A LightGBM model trained on population-level survey data demonstrates promising discrimination performance for diabetes risk screening, with a strong NPV supporting its use as a triage instrument. The exclusive use of non-African training data is the primary limitation and the central motivation for the Wo’tosuga platform design, which incorporates an anonymous data donation mechanism to collect African-specific data for model retraining. These results establish the Phase 1 performance baseline for a planned programme of African-context model development and validation.