This allows the model to learn not only from the direct sequence of events, but also from complex relationships between different physiological states,” explains Maskeliūnas.
An important component of the model is what is known as an attention mechanism. It allows the algorithm to identify which earlier data points are most relevant to the outcome.
“It is similar to an experienced doctor assessing a patient’s condition: rather than mechanically reviewing all the available data, they identify the factors that matter most in that particular situation,” explains Maskeliūnas.
This principle also addresses another important challenge surrounding the use of AI in medicine: explainability. Even a highly accurate algorithm has limited value in healthcare if clinicians cannot understand how it arrived at a particular result. The researchers’ approach therefore makes it possible to identify which features and earlier points in time had the greatest influence on the model’s output.
Model Could Support Clinical Decision-Making
The model was tested using international type 1 diabetes datasets and demonstrated high accuracy in both glucose forecasting and hypoglycaemia risk assessment. The researchers also found that performance improved when the algorithm was trained to carry out both tasks simultaneously rather than separately.
“The results showed that the new models achieve very high forecasting accuracy and perform well even with patients whose data the algorithm has not previously encountered,” says KTU PhD student Muhammad Abdullah Sarwar, who contributed to the development of the model.
The personalised insulin adjustments evaluated in the study are not intended to allow patients to alter their treatment independently. Instead, they were investigated as a potential decision-support tool for clinical analysis.
“This is particularly important when developing widely applicable clinical decision-support tools, as such systems need to perform reliably across different hospitals and countries, and with different patient populations,” emphasises Sarwar.
High accuracy, however, does not mean that such algorithms are ready to become part of patients’ everyday lives. “Further clinical studies are needed before the technology can be widely adopted in clinical practice. The models were evaluated using historical patient data, so it is essential to demonstrate that they remain equally accurate under real-world clinical conditions. Our medical partners will be responsible for this next stage,” says Maskeliūnas.
Alongside clinical testing, questions remain around data security, patient privacy and the seamless integration of algorithms into clinicians’ everyday work. The researchers see AI not as a technology that could replace doctors, but as a way to make better use of the vast amounts of data generated by patients and, in the future, provide more personalised support.
“Advanced, explainable AI models will not only allow us to predict glucose changes more accurately, but will also provide a foundation for developing safer, more personalised and more effective decision-support systems that can help reduce the heavy workload faced by clinicians,” says the KTU professor.
The article GAT-BiGRU: explainable multi-task temporal graph learning for glucose forecasting, hypoglycemia risk, and counterfactual insulin adjustment is available here.
