In a prospective single-center study conducted at Singapore General Hospital, researchers demonstrated that brief facial video recordings can provide modest but statistically significant information about laboratory-measured HbA1c levels, a key marker for type 2 diabetes and glycemic control. The study, published in Frontiers in Digital Health, evaluated a deep learning model named NervoHbA1c, which uses remote photoplethysmography (rPPG) to analyze subtle color changes in facial skin. The central finding is that while the model does not replace laboratory testing, it offers incremental predictive value over simple demographic baselines when the training data matches the specific camera hardware used for acquisition.
Type 2 diabetes remains a major public health challenge, with an estimated 589 million adults living with the condition globally in 2024. A significant portion of these individuals remains undiagnosed, particularly in settings where systematic screening is difficult to deliver at scale. Conventional HbA1c testing, while widely used for diagnosis and monitoring, depends on blood sampling, laboratory infrastructure, and trained personnel. These requirements limit its scalability as a low-friction screening tool in community or resource-constrained environments, creating a need for complementary, noninvasive approaches to identify individuals who may benefit from confirmatory laboratory testing.
Remote photoplethysmography offers a potential solution by detecting subtle color changes in facial video recordings caused by blood volume pulsations in the microvasculature. This technology enables contactless acquisition of physiological signals using standard camera hardware. However, predicting HbA1c is particularly challenging because it is not an instantaneous hemodynamic variable like heart rate. Instead, HbA1c reflects cumulative glycemic exposure over the preceding two to three months. Any relevant camera-derived signal likely reflects complex relationships between chronic glycemic exposure, hemoglobin-related optical properties, and microvascular function, rather than a direct optical measurement of glycated hemoglobin.
The study involved 254 participants who had both a usable facial recording and a laboratory HbA1c value. The participants were recruited across sequential phases at the Pre-Admission Centre and the Diabetes and Metabolism Centre. The development cohort consisted of 58 participants recorded on a laptop webcam, while two independent evaluation cohorts, SGH-5 and SGH-6, comprised 90 and 106 participants respectively, recorded on HONOR 200 smartphones. Each recording was reduced to the mean red, green, and blue values of six facial regions over 8.53 seconds, ensuring that no high-resolution facial images were used as model input.
NervoHbA1c is a convolutional neural network with 4.27 million trainable parameters that processes these RGB time-series data alongside demographic inputs such as age, sex, body mass index, and skin tone. The model was evaluated using rigorous subject-level partitioning to prevent identity leakage, a common issue in rPPG studies where multiple windows from the same individual can inflate apparent performance. The evaluation included pooled 5-fold cross-validation, leave-one-cohort-out testing, and comparisons against simple baselines such as the mean, median, and demographics-only predictors.
In the pooled evaluation of all 254 participants, NervoHbA1c achieved a mean absolute error (MAE) of 1.081% (95% CI 0.973–1.193). This performance was superior to the median baseline (1.230%), the demographics-only model (1.237%), and the mean baseline (1.254%). The model’s predictions correlated with laboratory HbA1c values with a Pearson correlation coefficient of 0.505. Furthermore, the model had the lowest error in every leave-one-cohort-out split, and its error decreased steadily as the training set size grew, indicating that additional data would likely improve performance further.
A critical finding emerged during the independent evaluation on the smartphone cohorts. The published checkpoint of NervoHbA1c, which was trained exclusively on 53 webcam participants, failed to transfer to the smartphone recordings. On the SGH-6 cohort, the model’s MAE was 1.478%, which was significantly worse than the demographics-only comparator (1.205%). This failure was attributed to the change in acquisition device. A logistic classifier could distinguish webcam from smartphone recordings with an AUROC of 1.000, driven mainly by differences in DC color levels and signal quality due to integer precision in the smartphone application.
When the model was retrained to include smartphone recordings, it outperformed the best comparator on both smartphone cohorts in leave-one-cohort-out evaluation. This suggests that the rPPG signal does carry information about HbA1c, but the model must be trained on data from the same type of device intended for deployment. The study also highlighted that the model’s performance was weaker at higher HbA1c values, where the development cohort had limited representation. This underprediction at high HbA1c levels is a problem of data coverage rather than a specific failure of the optical signal, as simple mean-baseline predictors also performed poorly in this range.
The researchers emphasize that the current model is not a replacement for laboratory HbA1c testing or a stand-alone screening tool. The effect size is modest, and the technology is sensitive to acquisition shift, meaning that changes in camera hardware, lighting, or software can significantly impact performance. The study underscores the importance of participant-level evaluation and the inclusion of simple baseline comparators in digital biomarker research. Without these rigorous methodological standards, apparent model performance can be misleadingly inflated by identity leakage or lack of proper baselines.
Future work should prioritize targeted data collection to address the under-representation of high HbA1c values and to standardize acquisition conditions. The researchers plan a prospective cohort of at least 300 additional participants recorded with high-resolution video at full precision on a single smartphone model. This will allow for retraining and a pre-registered external evaluation to establish the true size of the effect. The study provides a methodological test case for camera-based digital biomarkers, showing that a modest signal can be demonstrated only when appropriate baseline comparators and participant-level evaluation are in place.
The biological mechanism behind the association between facial rPPG signals and HbA1c remains a hypothesis. Type 2 diabetes is associated with arterial stiffness and changes in vascular function, which could plausibly influence the recorded signal. However, the correlation is indirect and does not constitute a direct measurement of glycated hemoglobin. The study’s findings support continued research into camera-based cardiometabolic risk stratification, provided that future studies are designed to address the specific failure modes identified here, including high-HbA1c under-representation and sensitivity to acquisition shift.
In conclusion, short facial rPPG recordings carry modest but recoverable information about HbA1c beyond demographics when the training data matches the acquisition device. The study highlights the potential of noninvasive sensing for cardiometabolic risk stratification while clearly delineating the current limitations and the need for rigorous validation. As digital health technologies advance, the integration of such models into clinical workflows will require careful attention to data quality, device consistency, and the clinical context of the predictions.
Subject of Research: Camera-based HbA1c estimation using remote photoplethysmography: development, independent evaluation, and baseline comparator analysis
Article Title: Camera-based HbA1c estimation using remote photoplethysmography: development, independent evaluation, and baseline comparator analysis
Article References: Abdullah, H. R., Tan, S. Y., Kalhara, I., Wang, S., Lee, V. J. Q., Ng, R. R. G., & Bee, Y. M. (2026). Camera-based HbA1c estimation using remote photoplethysmography: development, independent evaluation, and baseline comparator analysis. Frontiers in Digital Health, 8, Article 1978726. https://doi.org/10.3389/fdgth.2026.1978726
Image Credits: AI Generated
DOI: 10.3389/fdgth.2026.1978726
Keywords: Camera-based, HbA1c, estimation, remote, photoplethysmography, development, independent, evaluation, baseline, comparator, analysis, scientific research
News Source: Ophelia Keating. (November 7, 2026). Facial Camera Signals Predict HbA1c Beyond Demographics. Scienmag.
