AI Model Reads Glucose from Wearable Signals, but Struggles with Unseen People

Researchers have developed an uncertainty-aware machine learning framework that estimates blood glucose non-invasively from wearable physiological signals, combining photoplethysmography (PPG) and galvanic skin response (GSR) measurements. The study, published in Frontiers in Medical Technology, reports strong accuracy when tested on data from the same individuals it was trained on, but substantially weaker performance when applied to people the model had never seen—a limitation the authors say must be overcome before any clinical or wearable deployment.

The framework processes synchronized segments of PPG and GSR signals using deterministic harmonic Fourier features, which transform the raw physiological recordings into a representation the network can interpret. These features then pass through a dense projection layer, a bidirectional gated recurrent unit (Bi-GRU) equipped with additive attention, and finally a multilayer regressor that produces the glucose estimate. Training relies on a composite objective that combines mean absolute error, a smoothness penalty on input gradients, and a soft constraint on standardized outputs. At inference time, Monte Carlo dropout is used to estimate how uncertain the model is about each prediction—an important addition, because single point estimates alone do not convey reliability.

The evaluation used PhysioCGM recordings from 10 adults with type 1 diabetes, with continuous glucose monitor readings serving as the reference values. Under a random sample-wise 80/20 split—meaning data points from the same individuals appeared in both training and test sets—the model achieved a mean absolute error of 3.09 mg/dL, a root mean-square error of 4.72 mg/dL, and a coefficient of determination of 0.960. Agreement with reference glucose values was also strong in clinical error-grid terms: 96.3 percent of predictions fell in Clarke Error Grid Zone A, and 99.1 percent fell in Zones A and B combined.

However, the picture changed markedly under ten-fold leave-one-subject-out evaluation, in which each fold held out one entire participant. In that setting, the mean absolute error rose to 16.75 mg/dL, with a between-fold standard deviation of 0.44 mg/dL, and agreement deteriorated substantially in the hypoglycemic range. The authors attribute this pattern to the fact that sensor–glucose relationships vary across individuals, so a model trained on nine people does not necessarily transfer to the tenth.

The study’s conclusion is deliberately cautious: the results demonstrate strong within-cohort interpolation—accurate estimation for individuals whose data the model has seen—but limited generalization to unseen individuals and to underrepresented glycemic extremes. The authors state that larger participant-disjoint studies and calibrated uncertainty estimates are required before the approach could be considered for clinical or wearable use. For now, the work establishes a modeling pipeline that pairs multimodal signal fusion with explicit uncertainty quantification, while making clear that the headline accuracy numbers reflect a favorable evaluation split rather than performance on new users.

Subject of Research: Non-invasive blood glucose estimation from wearable PPG and GSR signals using an uncertainty-aware deep learning model

Article Title: Uncertainty-aware fourier-encoded attentive Bi-GRU for non-invasive blood glucose estimation from multimodal PPG and GSR signals

Article References: Prakash, A. J., Jelinek, H. F., & Atef, M. (2026). Uncertainty-aware fourier-encoded attentive Bi-GRU for non-invasive blood glucose estimation from multimodal PPG and GSR signals. Frontiers in Medical Technology, 8, Article 1979468. https://doi.org/10.3389/fmedt.2026.1979468

Image Credits: AI Generated

DOI: 10.3389/fmedt.2026.1979468

Keywords: blood glucose, non-invasive monitoring, photoplethysmography, galvanic skin response, machine learning, uncertainty quantification, Bi-GRU, type 1 diabetes, continuous glucose monitoring, wearable sensors, Uncertainty-aware, fourier-encoded