The accuracy of survival prognosis for patients with diabetes improved by UNN researchers
Scientists at the UNN Research Institute of Biogerontology and the Research Centre for Artificial Intelligence have developed an AI-based model capable of predicting the risk of death from all possible causes in patients with diabetes and explaining its conclusions to medical specialists.
The model was trained using data on the health status of over 550 patients with diabetes who were monitored for 17 years. From hundreds of clinical and laboratory parameters, the artificial intelligence identified ten key biomarkers that shape the long-term prognosis. The main innovation of the study is the interpretation of neural network analysis using the Shapley additive explanations (SHAP) method, which reveals which data played a decisive role in the forecast. The accuracy of the survival prediction over a 17-year period throughout the entire observation period reaches 84%.
"The novelty of this research lies in creating not only an accurate, but also an explainable predictive tool. Using the method for interpreting AI forecasts, we can identify the relationships between dozens of parameters of the patient's condition. For instance, we have found that age, the duration of illness, and the number of complications are some of the most significant risk factors for mortality in diabetes. Moreover, our method enables the creation of an individual risk profile for each patient. For example, the model can demonstrate that a patient's 68% high risk of death is primarily attributable to elevated creatinine levels, age, and four diabetes-related complications," explained Mikhail Ivanchenko, the study's author and Director of the UNN Research Institute of Biogerontology.
While pointing to known basic parameters, the study highlights the importance of certain indicators whose biological role scientists have yet to determine. For example, along with brain natriuretic hormone (NT-proBNP), which reflects the latent stress of the heart muscle, or creatinine, which indicates kidney function, a particular structure of N-glycan in blood serum emerges as a biomarker for immune regulation and aging processes.
"By identifying these hidden patterns, we are shifting the focus from general disease management to targeted individual risk management. If the model shows that systemic inflammation is the primary contributor to the risk, we can consider anti-inflammatory therapy. When the lipid profile is the key factor, this allows us to optimise the prescription of relevant medications. Thus, explainable artificial intelligence transforms from an abstract algorithm into a practical assistant that doesn't replace doctors but enhances their clinical thinking. This is an important step towards prolonging and improving the lives of millions of people living with diabetes," said Mikhail Ivanchenko.
The research was conducted as part of a project by the UNN Research Centre for Artificial Intelligence. The findings were published in the journal Frontiers in Endocrinology.
