Prediction of Heart Failure in Type 2 Diabetes Mellitus Subjects Using Machine Learning: A Cross-Sectional Study

Abstract

Abstract The concomitance of Type 2 Diabetes Mellitus (T2DM) and heart failure has made scientists investigate ways the onset of heart failure in T2DM can be predicted. Machine learning techniques have been shown to help with the prediction of heart disease and several model algorithms have been affirmed as good. This study aimed at predicting heart failure in T2DM subjects using machine learning techniques. A total of 123 blood samples from 59 healthy subjects without T2DM (controls) and 63 T2DM subjects (tests) were analyzed for biochemical parameters [troponin (TnI), electrolytes, Lactate dehydrogenase (LDH), Aspartate aminotransferase (AST), Alanine transaminase (ALT), AST/ALT ratio, Creatinine phosphokinase (CK-MB), Fasting Blood Sugar (FBS), Cholesterol, Triglyceride, B-Natriuretic peptide (BNP)] using standard procedures. Demographic data and biochemical results were all subjected to machine learning algorithms. The results of ML showed that the Random Forest algorithm is the best model for heart failure prediction with 87% accuracy. SHAP value (impact on model output) among all possible combinations identified glucose (FBG), BNP, Systolic and diastolic blood pressure, and waist circumference as important features in the prediction of heart failure in T2DM. The permutation importance score of the features studied showed systolic BP, BNP, MUAC and troponin I in this order to have the highest positive importance to the prediction of heart failure in T2DM. Height, weight, and waist circumference have small negative importance values meaning they slightly decrease model performance. The study concluded that CK-MB, BNP, and troponin I alone may not be early indicators of heart failure in T2DM subjects. However, subjecting them to ML and combining them with the key features identified would make prediction better.