Scientists from Pitirim Sorokin Syktyvkar State University and their colleagues have developed an AI system to predict in-hospital mortality in patients with acute coronary syndrome. According to the university, the model proved to be more accurate than the international GRACE score, which is widely used in clinical practice.
To train the system, specialists analyzed data from over 14,000 patients from Komi, St. Petersburg, and the Leningrad Region. After data preparation, information on 13.3 thousand people was included in the final analysis. For each patient, 28 clinical parameters were assessed, including age, hemodynamic parameters, laboratory data, and instrumental examination results.
The CatBoost gradient boosting model showed the best result: its AUC-ROC reached 0.961 compared to 0.919 for the GRACE score. The developers believe that machine learning helps to find complex relationships between indicators that traditional clinical scales do not always take into account.
This allows for the formation of a more accurate individual patient risk profile already at the hospitalization stage, before percutaneous intervention, which in some cases can be dangerous for older patients, has been performed.
The researchers also checked which factors most strongly influence the prognosis. These included left ventricular ejection fraction, Killip class heart failure, age, systolic blood pressure, and dyslipidemia. The system is not yet ready for widespread implementation: the next step should be multicenter clinical trials in various medical institutions.