Comparison of Temporal and Non-Temporal Features Effect on Machine Learning Models Quality and Interpretability for Chronic Heart Failure Patients

Abstract

Chronic diseases are complex systems that can be described by various heteroscedastic data that varies in time. The goal of this work is to determine whether historical data helps to improve machine learning predictive models or is it more efficient to use the latest data describing the disease in particular moment in time. For simplicity we call features from the first group dynamic and features from the second one – static. We study the way both groups affect predictions quality and its interpretation. We set the experiments on data of chronic heart patients from Almazov Medical Research Center. From this data we extracted more than 300 features from patient comorbidity, anamnesis, analysis, etc. In terms of Chronic Heart Failure (CHF) modelling three different tasks have been selected: CHF identification as main diagnosis, CHF stage classification and diastolic blood pressure prediction. For each task several machine learning algorithms on three groups of features: static, dynamic and the whole feature set. The results show that, in general, models perform better on combination of temporal and non-temporal features.