A person with type 1 diabetes drives a car in thick fog, seeing only a few meters ahead. Blood sugar can plummet or spike, and the insulin pump must react in advance. A four-hour forecast is an attempt to glimpse the road's turn through the haze.
Standard neural networks predict step by step, like a driver who recalculates the route every second based on the previous command. Errors accumulate, and after a couple of hours the forecast drifts downward — the network systematically underestimates. A physiological model relies on equations of body function, but it's an averaged portrait: it doesn't account for individual metabolism.
The new approach, PhysioSeq2Seq, works like an experienced navigator. First, over three hours of observations, it finds a digital twin in the database whose sugar levels match the patient's current data. Then it takes 10 internal variables from that twin (such as glucose in blood and tissues, and how quickly it's consumed) and feeds them to the neural network. The network then outputs a forecast for 48 steps ahead at once, without looking back at its own previous answers. It's like plotting the entire route to the destination in one go, rather than recalculating every meter.
This 'hybrid' keeps the forecast from veering off from reality: physiology holds it within reasonable bounds, while machine learning adapts to the individual. As a result, the system helps the insulin pump work proactively, as if the fog lifted and the driver could see not just the hood but the entire curve ahead.
as written by the authors arXiv:2605.16860
Accurate long-horizon glucose forecasting is critical for automated insulin delivery systems, which help people with type 1 diabetes (T1D) manage their glucose and avoid dangerous hypoglycemia. However, standard recursive long short-term memory (LSTM) networks suffer from systematic negative bias at longer horizons due to error compounding, while purely mechanistic ordinary differential equation (ODE) models fail to generalize across individuals when parameterized at the population level. We propose PhysioSeq2Seq, a hybrid architecture that combines patient-specific physiological modeling with a sequence-to-sequence (Seq2Seq) LSTM. For each glucose segment, twin matching searches a population of 300 parameterized digital twins to identify the best-fitting physiological match from a 3-hour continuous glucose monitoring (CGM) history. The 10 internal ODE state variables of the matched twin are injected as exogenous covariates into both the encoder and decoder of the Seq2Seq LSTM. This simultaneous 48-step prediction strategy eliminates recursive error compounding, while the ODE features provide a physics-grounded constraint that bounds long-horizon drift within physiologically plausible ranges. PhysioSeq2Seq was trained on CGM and insulin data from 348 participants in the Type 1 Diabetes Exercise Initiative (T1DEXI) dataset and evaluated on 74 held-out participants. At the 240-minute horizon, PhysioSeq2Seq achieves a mean absolute error of 39.28 mg/dL and a mean error of -10.62 mg/dL, reducing bias by 13.89 mg/dL over the recursive LSTM and reducing mean absolute error by 28.62 mg/dL over the ODE-based digital twin. These results show that eliminating architectural feedback and injecting patient-matched physiological states is an effective and clinically meaningful strategy for long-horizon glucose forecasting in T1D.
A sugar level of 39.28 mg/dL is roughly two teaspoons of glucose dissolved in all the blood of an adult.