Selection of features for modeling the risk of fatal outcomes in patients after myocardial infarction or unstable angina

Aim. To identify main predictors of fatal outcome based on the regional database of patients with myocardial infarction (MI) or unstable angina (UA). Material and methods. The retrospective study included 1515 patients with UA and MI, which accounted for 55,3% of patients treated in the period 2012-2016. The median follow-up was 62 [36; 71] months. The criterion for a complicated course of coronary artery disease (CAD) is fatal outcome. Group 1 included 238 patients with fatal outcome of CAD, while group 2 — 1277 patients with favorable outcome. All obtained data were structured as an Excel file. Variables available to most medical institutions (except coronary angiography data) were extracted from medical records. For automatic feature extraction, we used an ensemble machine learning algorithm developed by Yandex — CatBoost (Categorical Boosting). Results. Mortality over 62-month follow-up was 15,4%. The study used 47 quantitative and qualitative (categorical) features. The filter feature selection identified significant quantitative characteristics, including age, left ventricular (LV) ejection fraction (EF), glomerular filtration rate, creatinine, body mass index, height, weight, body surface area (BSA), red blood cells, hemoglobin, glucose, total cholesterol (TC), lowdensity lipoprotein cholesterol, high-density lipoprotein cholesterol, heart rate, LV end-diastolic volume index, LV end-systolic volume index, pulmonary artery systolic pressure. There were following categorical variables: heart failure (HF), heart failure Killip class, old myocardial infarction, comorbidity, chronic kidney disease, angina pectoris, diabetes, atrial fibrillation, positive troponins, S-T deviation, coronary angiography, percutaneous coronary intervention, nosological unit (UA, anterior or inferior MI). An automatic feature selection using a machine learning algorithm identified the following most significant features determining the probability of death: age, LVEF, BSA, creatinine level, systolic blood pressure, HF, comorbidity, nosological unit. Conclusion. Forty-seven available clinical features were selected from the medical records of patients with MI and UA. There were 8 following most significant parameters for predicting a fatal outcome according to machine selection results: age, LVEF, BSA, creatinine level, systolic blood pressure, HF, comorbidity, nosological unit.