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

Technical Summary: Explainable Machine Learning for Diabetes Risk Prediction on the Wo'tosuga Platform

Problem Statement
Diabetes mellitus represents a rapidly escalating public health crisis in sub-Saharan Africa, with an estimated 54% of affected individuals remaining undiagnosed. Conventional laboratory-based screening (e.g., fasting plasma glucose, HbA1c) is often inaccessible due to infrastructural, financial, and geographic constraints. While machine learning (ML) offers a scalable, non-invasive alternative using self-reported variables, existing predictive models suffer from two critical limitations:

1. Data Generalizability: Most models are trained on North American or European datasets (e.g., Pima Indian, US BRFSS), limiting their applicability to African populations due to differences in risk factor distributions and phenotypic characteristics.
2. Evaluation Metrics: The literature predominantly relies on overall accuracy or AUC-ROC, which are inappropriate for imbalanced screening tasks. In low-resource settings, the clinical priority is maximizing sensitivity (identifying true at-risk individuals) to prevent missed diagnoses, rather than maximizing accuracy which often favors the majority class.

Methodology
This study establishes a Phase 1 baseline predictive engine for the Wo'tosuga digital health platform using a rigorous, reproducible pipeline:

• Data Source: The study utilized three cycles (2019, 2021, 2023) of the US Behavioral Risk Factor Surveillance System (BRFSS), yielding an analytical dataset of 869,319 respondents.
• Preprocessing:
  • Feature Selection: 21 predictive features were retained after removing variables with >50% missingness.
  • Target Recoding: Pre-diabetic cases were reclassified as diabetic (positive class) to align with the clinical goal of flagging all at-risk individuals for intervention.
  • Class Imbalance: The dataset exhibited a ~5.4:1 imbalance (84.4% non-diabetic vs. 15.6% diabetic). Instead of synthetic oversampling (SMOTE), the study employed cost-sensitive learning (e.g., class_weight='balanced', scale_pos_weight) to preserve natural probability distributions.
  • Feature Engineering: Categorical variables (marital/employment status) were collapsed for interpretability; a socioeconomic tier variable was derived to be transferable to non-US contexts.
• Model Development:
  • Eight algorithms were evaluated: Logistic Regression, Decision Tree, Extra Trees, Gradient Boosting, AdaBoost, XGBoost, LightGBM, and K-Nearest Neighbors.
  • Optimization: Hyperparameter tuning was conducted via GridSearchCV with 3-fold cross-validation, explicitly optimizing for recall (sensitivity).
  • Evaluation: Sensitivity was designated as the primary metric, alongside specificity, PPV, NPV, F1-score, AUC-ROC, and Brier score.
• Explainability: SHAP (SHapley Additive exPlanations) values were computed using the TreeExplainer algorithm to determine global feature importance and local prediction contributions.

Key Results

• Model Performance: LightGBM emerged as the superior algorithm. Following hyperparameter tuning and retraining on the full dataset, the final model achieved:
  • Sensitivity: 0.809 (80.9% of true diabetics correctly identified).
  • Specificity: 0.701.
  • AUC-ROC: 0.828.
  • Negative Predictive Value (NPV): 0.952.
  • Note: While Gradient Boosting achieved a similar AUC-ROC (0.826), its sensitivity was critically low (0.190), rendering it clinically unsuitable for screening despite high accuracy.
• Comparators: XGBoost (Sensitivity: 0.806) and Logistic Regression (Sensitivity: 0.779) served as strong comparators, suggesting the predictive signal in this dataset is largely captured by a small set of linearly separable risk factors.
• Feature Importance: SHAP analysis identified the top five predictors, consistent across both LightGBM and Logistic Regression:
  1. Age
  2. Self-reported general health
  3. BMI
  4. Blood pressure status
  5. Time since last medical checkup
• Interpretability: The model demonstrated clinically coherent patterns (e.g., monotonic risk increase with age and BMI; inverse relationship between general health rating and risk). Notably, the socioeconomic tier variable showed negligible importance, likely due to the coarseness of the derived metric.

Significance and Claims
The paper positions this work not merely as a predictive modeling exercise, but as the foundational infrastructure for a translational research loop designed to address the "African medical data gap."

1. Clinical Utility as a Triage Tool: The model is explicitly framed as a screening and triage instrument, not a diagnostic tool. The high NPV (0.952) is highlighted as the most clinically significant metric for low-access settings, providing reliable reassurance for negative results and enabling the prioritization of positive cases for scarce laboratory resources.
2. The "Wo'tosuga Research Loop": The authors acknowledge that training exclusively on US data is a primary limitation. However, they reframe this as the central motivation for the platform's architecture. The Wo'tosuga platform integrates an anonymous data donation mechanism, allowing users to contribute their risk factors and predictions. This creates a self-improving cycle where the US-trained baseline enables immediate deployment, while the resulting user interactions generate the African-specific data required to retrain and validate future model iterations.
3. Methodological Rigor: The study emphasizes the critical importance of metric selection in clinical AI. By prioritizing sensitivity over accuracy, the research demonstrates that models appearing "competitive" under standard benchmarks can be clinically dangerous if they fail to detect the majority of at-risk individuals.

In conclusion, the paper establishes a reproducible, explainable baseline (LightGBM, Sensitivity ~0.81, NPV ~0.95) for diabetes risk stratification. Its primary contribution is the architectural design of a digital health platform that simultaneously functions as a clinical screening tool and a data collection engine to progressively close the gap in African-specific health data.