Few-Shot Transfer Learning for Diabetes Risk Prediction Across Global Populations.

Area of Science:

• Computational biology
• Machine learning
• Public health

Background:

• Type 2 diabetes mellitus (T2DM) affects over 537 million adults globally, with a significant impact on low- and middle-income countries.
• Predictive models for T2DM often fail to generalize across different populations due to variations in data and measurement practices.
• Limited diagnostic resources in many regions hinder effective T2DM screening and management.

Purpose of the Study:

• To evaluate a few-shot domain adaptation framework for improving T2DM risk prediction across diverse populations.
• To assess the effectiveness of fine-tuning models with small percentages of local data for cross-population generalization.
• To investigate population-specific feature importance and decision thresholds in T2DM prediction.

Main Methods:

• A few-shot domain adaptation framework using a multilayer perceptron with four clinical features (age, BMI, MAP, plasma glucose) was employed.
• The framework was tested across three adult cohorts: Bangladesh, Iraq, and the Pima Indian dataset.
• Models were pre-trained on a source cohort and fine-tuned on 1-20% of labeled target data, with performance compared to threshold tuning.

Main Results:

• Few-shot fine-tuning restored predictive capability in cases where source-only models failed, significantly improving F1 scores.
• Adaptation enhanced AUROC by up to 0.14 and accuracy by 17 percentage points in moderate-performance transfers.
• SHAP analyses revealed population-specific shifts in feature importance (e.g., BMI in Pima) and decision thresholds for glucose.

Conclusions:

• Few-shot domain adaptation, using as little as 5-20% of local labels, effectively enhances cross-population T2DM risk prediction.
• The approach utilizes routinely available clinical features and offers a scalable, interpretable solution for preventive screening.
• This method can democratize T2DM screening in diverse and resource-constrained settings.