How an ICU digital twin forecasts blood sugar swings on standard laptops

When a severe illness like sepsis strikes, blood sugar can swing fast between dangerous highs and lows. Both sharp spikes and steep crashes make it much harder for very sick hospital patients to recover. Doctors need an early warning to step in before these crashes happen. Yet every human body reacts in its own way during an illness. Could a smart software tool learn a patient's bodily signals quickly enough to give advance warning right at the bedside?

Building a bedside twin

To tackle this challenge, researchers created an ICU digital twin, which is a virtual computer model that updates using live patient data to mimic bodily changes. The system gets its information from continuous glucose monitoring, a wearable sensor that reads sugar levels under the skin throughout the day and night. Setting up a fresh model from scratch for each person takes far too long in a busy hospital. Instead, the team turned to PatchTST, a neural network architecture that slices continuous time-series data into small chunks or patches to capture long-term trends.

Because the network has already learned broad patterns from past data, it can start working the moment a patient arrives. The tool runs zero-shot inference, which means running a machine learning model directly on new data without prior training on that specific patient. It takes the last 30 minutes of sensor numbers to predict where sugar levels will head over the next 15 to 30 minutes. As fresh sensor readings arrive, the system pools overlapping predictions using a time-decay method and compares them to actual body numbers.

Fast learning on ordinary kit

The team tested the tool on past records from 10 patients with sepsis and diabetes. Models trained completely from scratch struggled to keep up, but the pre-trained setup adapted with speed. To tune itself, the system used linear probing, a fine-tuning method where most of a pretrained model is frozen and only the final decision layer is retrained. This step updated just 0.36% of the network's settings—23,055 out of more than 6.3 million values. Even full model fine-tuning took under 48 seconds on a standard laptop.

Looking 30 minutes ahead instead of 15 roughly doubled the prediction errors, but the worst-case relative error stayed below 9.5%. The training time grew by only about one second. If the tracking error passed 5%, the setup ran a rapid update to fix its predictions. If the error ever reached 10%, a safety cutoff would pause forecasts to avoid bad guidance. Across all 10 patient test cases, that 10% safety line was never breached.

What we still do not know

The trial used historical records rather than live bedside tests. The wearable sensors check fluid under the skin rather than blood, which causes a slight delay that needs real-time hospital checks. The model also works only from sugar monitor numbers, leaving out insulin doses and meal times. Doctors often plan insulin one to three hours ahead, so whether this tool can reliably look past 30 minutes remains untested.