Machine Learning Framework for HbA1c Prediction: Data Enrichment, Cost Optimization, and Interpretability Through Stratified Regression and Multi-Stage Feature Selection

TL;DR

This paper presents a machine learning model that predicts HbA1c levels using a small set of clinical features, offering a cost-effective and interpretable solution for large-scale health assessments.

Contribution

The study introduces a unified framework for continuous HbA1c prediction that integrates cost-efficient feature selection, stratified regression, and model explainability.

Findings

The optimal model achieved R2 = 0.7161 using only 40 selected features from 224 candidates.

Interpretability analysis showed clinically coherent relationships aligned with physiological expectations.

The framework reduces feature dependency and enables cost-efficient HbA1c estimation in resource-limited settings.

Abstract

Background: Measuring glycated hemoglobin (HbA1c) is essential for assessing long-term glycemic control, yet direct testing remains expensive and underutilized in many large-scale health surveys and resource-constrained settings. This study aims to (i) deliver a highly accurate and interpretable ML model for predicting HbA1c from routinely collected clinical, biochemical, and demographic data, (ii) reduce dependency on extensive laboratory panels by identifying a compact, cost-efficient subset of key predictors, and (iii) establish a transferable, explainable modeling framework applicable across chronic disease biomarkers. Unlike prior HbA1c prediction studies that focus primarily on classification or accuracy-driven models, this work introduces a unified framework for continuous HbA1c regression that jointly integrates cost-oriented feature parsimony, stratified regression validation, and explainability by design. Methods: We aggregated data from the National Health and Nutrition Examination Survey (NHANES) cycles 2007–2020, encompassing 66,148 records and 224 candidate features. We implemented a two-stage feature selection pipeline: Incremental Correlation Selection (ICS) to narrow the variable space, followed by Recursive Feature Elimination with Cross-Validation (RFECV) to isolate the most informative features. Model interpretability was assessed using partial dependence plots and feature importance analysis. Results: The optimal model, LightGBMRegressor with most-frequent imputation, achieved R2 = 0.7161, MAE = 0.334, MSE = 0.304, and MAPE = 5.56%, while using only 40 selected features. Interpretability analysis revealed clinically coherent relationships that align with physiological expectations. Discussion: The proposed framework maintains robust predictive performance while substantially reducing the number of required input features, enabling cost-efficient HbA1c estimation together with transparent, physiologically coherent model insights. By consolidating continuous HbA1c prediction, cost-aware feature selection, stratified evaluation, and explainability within a single pipeline are enhanced. Conclusions: This study advances beyond existing approaches and offers a practical blueprint for scalable biomarker estimation in population health and clinical decision-support applications. Its explainable, efficient, and generalizable design positions it as a strong candidate for clinical decision-support and population-health applications.

Figures