Deep learning models can forecast a person's blood glucose with impressive accuracy, but the authors of a new study argue that insufficient consideration of physiological consistency and physiological explainability has led to a lack of clinical confidence in them. The authors point to two challenges: deep learning models may learn relationship patterns that violate fundamental physiological principles, and their black-box character leaves results lacking physiological explainability. A neural network trained on continuous glucose monitor readings, meal logs and insulin pump records may quietly learn relationships that violate basic physiology, such as insulin raising blood glucose or carbohydrates lowering it. The authors warn that further decisions based on such incorrect relationships may lead to severe clinical consequences.

In an original article published on 29 October 2026 in Medical & Biological Engineering & Computing, the researchers introduce PhysioXGluNet, a physiology-informed network architecture with attributable branches designed for explainable blood glucose prediction. Rather than feeding all variables into one undifferentiated network, the model splits its work into separate branches, each responsible for a distinct physiological driver of blood glucose, and then reassembles the pieces into a forecast.

The architecture is grounded in the Hovorka physiological model, a compartmental description of glucose and insulin dynamics. PhysioXGluNet contains a basic regulation branch covering endogenous glucose production and glucose consumption by the central nervous system, a carbs branch modeling the dynamic response to carbohydrate intake, a bolus branch modeling the response to bolus insulin, and a glucose branch that captures residual changes caused by other unknown factors. Each branch outputs the blood glucose changes induced by its own factor. Combined with the current blood glucose value, these contributions yield the predicted blood glucose levels across the prediction horizon.

Keeping the branches separate does two things at once. It preserves the original physiological scale of each variable, so learning is not distorted by scale differences between, say, grams of carbohydrate and units of insulin. It also makes the model’s internal reasoning visible: the net change in blood glucose attributable to each input variable can be read directly from the corresponding branch, allowing clear traceability of the dominant factors at each time step driving future blood glucose changes. That is a structural property of the network, not a post hoc explanation bolted on after training.

The bolus and carbs branches rely on a novel component the authors call the physiology-informed recurrent neural network, or PIRNN. Instead of a generic recurrent cell, the PIRNN is explicitly derived from the Hovorka model’s ordinary differential equations for insulin and glucose absorption, discretized with the forward Euler method. The learnable parameters inside it carry physiological meanings, such as the time-to-maximum insulin absorption, the insulin distribution volume and the insulin elimination rate. Because these parameters are trained on individual data, the model can reveal personalized, heterogeneous physiological responses of different patients to the same life events.

A practical wrinkle arises because the model processes fixed sliding windows of data. A naive recurrent cell would reset to zero at the start of each window, losing memory of meals and boluses that occurred before the window opened. The team solved this with a state inheritance mechanism: the current state of the first time step in one instance is saved and used as the initial state of the next instance, which shares a physical timeline offset by a single five-minute sampling interval. This preserves the continuity of physiological states, so the model can still perceive the sustained impact of events whose effects generally persist for two to six hours, even when the events themselves fall outside the input window.

Training is steered by a physiology-informed composite loss with three penalty terms. The first penalizes incorrect effect direction, pushing the model to learn that insulin only lowers blood glucose and carbohydrates only raise it. The second penalizes incorrect effect intensity, enforcing that when physiological flux values are very small, the estimated effect should be minimal or even zero; the authors set thresholds of 50 mg/min for glucose appearance and 3 mU/L for plasma insulin to identify weak effects. The third is a smoothness regularization on rhythmic effects. The loss weights were calibrated by a standard balancing practice, scaling the direction penalty by 1000 and the intensity penalty by 100 so that no single term dominates early training simply because of its raw magnitude.

Physiological constraints are also wired directly into the network’s prediction heads. The bolus branch ends in a direction layer that forces its output to be non-positive, and the carbs branch forces its output to be non-negative, implemented through customized rectified linear units. The effect encoder within these branches uses multi-head causal self-attention with a banded causal mask, restricting attention to the twelve time steps prior to the current moment, reflecting that the impacts of bolus and carbohydrates on blood glucose are cumulative and delayed rather than instantaneous. The basic regulation branch, meanwhile, uses factor generators to simulate the inhibitory effect of basal insulin on endogenous glucose production and the ultradian and circadian rhythms of glucose production and central nervous system consumption, with rhythmic effects assumed to fluctuate between 0.8 and 2 times a base value.

The evaluation used the OhioT1DM 2018 dataset, which contains six subjects with type 1 diabetes on insulin pump therapy with continuous glucose monitoring throughout an eight-week collection period, including CGM readings every five minutes, insulin doses and self-reported meal times with carbohydrate estimates. The prediction horizon was set at 60 minutes, twelve five-minute steps ahead; the authors note that a 30-minute interval is the minimum effective duration for proactively taking actions to prevent abnormal blood glucose, and that appropriately increasing the prediction horizon improves clinical usability but at the cost of clinical accuracy. PhysioXGluNet was compared against a naive prediction baseline, the CRNN model that won the 2020 blood glucose level prediction challenge, the hybrid-driven LMU and SIV models, and the graph-attention GARNN model, with results averaged over twenty experiments with different random seeds.

Across all subjects, PhysioXGluNet outperformed the baselines on both numerical and clinical metrics. Compared with naive prediction, it improved mean absolute error by 32.44 percent, root mean square error by 30.51 percent, the average Clarke error within Zones A and B by 0.43 percent, and average surveillance error by 31.66 percent. The authors note that the naive predictor already performs strongly on the clinical accuracy metric, so even modest gains there are meaningful but challenging. Ablation experiments confirmed that both the directional prediction head and the physiological loss are needed: models lacking either component learned incorrect action mechanisms or misestimated effect values, while the complete version achieved the best results. The model has 150,798 parameters and an inference time of approximately 0.9 seconds per prediction, well below the five-minute CGM sampling interval.

The explainability analysis goes beyond a single illustrative day. A systematic quantitative analysis of branch output curves across all six subjects found that the mean carbohydrate absorption peak occurred at 45 plus or minus 12 minutes, within the known 40 to 60 minute range for rapid-acting carbohydrate absorption, and the mean carbohydrate effect duration was 2.7 plus or minus 0.5 hours, consistent with the typical 2 to 3 hour absorption window for mixed meals. The mean insulin bolus effect duration was 4.2 plus or minus 0.6 hours with a time to peak effect of 85 plus or minus 18 minutes, matching the established pharmacodynamic profile of rapid-acting insulins such as Humalog and Novolog. The authors caution that this is not a direct validation against ground truth of individual metabolic dynamics in vivo, but it provides stronger evidence than a single-case illustration.

Learned physiological parameters also varied across the six subjects in ways consistent with inter-subject variability, and an ablation with population-unified parameters confirmed that personalization is beneficial. The authors emphasize, however, that this observation is purely methodological: the dataset lacks individual-level clinical metadata such as body weight or insulin sensitivity indices, so no clinically meaningful phenotype should be inferred from parameter values alone. Temporal attention weights in the bolus and carbs branches further showed how the model shifts its focus across past time steps as insulin and glucose concentrations rise and fall, offering instance-level insight into which moments dominate current blood glucose changes.

The team is candid about the limits. The study relies on a single small dataset of six adult subjects aged 40 to 60, all using insulin pump therapy, so the learned parameters and attribution patterns cannot be assumed to transfer to adolescents, elderly patients, pregnant women or people on multiple daily injections. The per-subject training paradigm creates a cold-start problem for new users, leaves the model vulnerable to physiological drift from illness, stress, puberty, pregnancy or medication changes, and requires retraining after therapy changes. The authors position PhysioXGluNet as best suited for retrospective analysis or decision support for patients with long-term stable therapy, and they stress that prospective clinical validation, regulatory considerations and workflow integration would be required before any clinical deployment.

Subject of Research: A physiology-informed deep learning architecture with attributable branches for explainable blood glucose prediction in type 1 diabetes

Article Title: PhysioXGluNet: a physiology-informed network architecture with attributable branches for explainable blood glucose prediction

Article References: Wen, S., Li, H., & Sun, X. (2026). PhysioXGluNet: a physiology-informed network architecture with attributable branches for explainable blood glucose prediction. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03693-y

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03693-y

Keywords: type 1 diabetes, blood glucose prediction, deep learning, physiology-informed neural network, explainable AI, Hovorka model, continuous glucose monitoring, OhioT1DM dataset, insulin bolus, carbohydrate absorption, clinical decision support, personalized medicine