Machine learning in the analysis of medical profiles of patients with coronary heart disease in the context of unbalanced data

Abstract

Background. Using machine learning (ML) techniques in addition to traditional methods can make disease diagnosis more reliable and unbiased. The aim of the study is to compare the efficacy of machine learning methods on imbalanced data for identifying dominant diagnostic features that form the clinical profile of a patient with coronary artery disease (CAD). Materials and Methods. A retrospective, single-center, continuous comparative study was performed with a predominance of patients without CAD (n2 = 3023) over patients with CAD (n1 = 446) in the sample (n = 3469) using the classical case-control method and ML methods with metric evaluation. Within the chosen ML method with optimal metrics, the contribution of each feature to the medical profile of a patient with CAD was assessed based on the calculation of odds ratios. Results. Due to the differences in methodologies and goals, the case-control analysis and ML methods showed differences in the selection of significant features. A comparative analysis of ML algorithms determined weighted logistic regression as the optimal model, demonstrating high predictive value based on ROC analysis: 71 out of 93 patients with CAD and 381 out of 601 patients without CAD were correctly identified. The dominant features of the patient portrait with CAD include advanced age, excess body weight, and elevated levels of lymphocytes and CRP. Female gender, increased levels of hemoglobin and platelets were associated with a lower probability of CAD. Systemic inflammation indices SII, SIRI, AISI have limited diagnostic value. Conclusion. ML methods successfully find patterns and build predictive models in the context of data imbalance.