A Machine Learning–Enhanced Calibration Framework for Green-Light Based Non-Invasive HbA1c Measurement

Abstract

Hemoglobin A1c (HbA1c) is a key biomarker for diagnosing and monitoring diabetes. Although the gold standard of monitoring in diabetes is lab HbA1c testing, invasive sampling means that frequent monitoring is not feasible. The current research paper suggests and tests a machine-learning-based system of calibration to determine HbA1c based on non-invasive green-light reflectance signals. A 90 adult sample was enrolled with laboratory HbA1c comprising of a standardized NGSP certified test. Measuring green-channel reflectance concurrently on the device were taken. Ordinary least squares (OLS), Deming regression with λ-sensitivity, Passing regression and Huber regression and machine learning algorithms, such as Random Forest (RF), LightGBM, and XGBoost, were used to develop calibration models. The evaluation of performance was based on the nested cross-validation, where the measurements encompassed the mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), concordance correlation coefficient (CCC) by Lin, the BlandMandel analysis, and the receiver operating characteristic (ROC) curve at clinical thresholds (5.7% and 6.5%). Optimistic regression approaches (Huber and PassingBablok) had the lowest calibration error (MAE 1.16 HbA1c percentage points), which was better than OLS and Deming (MAE 1.29). The calibration of machine learning models did not improve: LightGBM was no worse than OLS, RF marginally worse, and XGBoost was significantly worse (MAE = 1.61). All the models had low explained variance (R2<0), weak agreement (CCC < 0.1) and ROC AUCs of almost random (0.53 -0.62). The offered framework offers a replicable regression and ML model comparison protocol as a non-invasive HbA1c calibration. Although the present device signals are not adequate to actually implement clinically, the paper shows the significance of intense calibration, sound regression as well as a methodological clear and open approach. These results provide a methodological framework on the future applications of multimodal sensing and machine learning methods in non-invasive diabetes monitoring.