Combined-task deep network based on LassoNet feature selection for predicting the comorbidities of acute coronary syndrome

Abstract

Acute coronary syndrome (ACS) is a multifaceted cardiovascular condition frequently accompanied by multiple comorbidities, which can have significant implications for patient outcomes and treatment approaches. Precisely predicting these comorbidities is crucial for providing personalized care and making well-informed clinical decisions. However, there is a shortage of research investigating the identification of risk factors associated with ACS comorbidities and accurately predicting their likelihood of occurrence beyond heart failure. In this study, an approach called Combined-task Deep Network based on LassoNet feature selection (CDNL) is presented for predicting ACS comorbidities, including hypertension, diabetes, hyperlipidemia, and heart failure. In order to identify crucial biomarkers associated with ACS comorbidities, the proposed framework first incorporates LassoNet, which extends Lasso regression to the deep network by adding a skip (residual) layer. Additionally, a correlation score calculation method across tasks is introduced based on measuring the overlap of identified biomarkers and their assigned importance. This method enables the development of an optimal combined-task prediction model for each ACS comorbidity, addressing the challenge of limited representations in traditional multi-task learning. Our evaluation, conducted through a meticulous cross-sectional study at a tertiary hospital in China, involved a dataset of 2941 samples with 42 clinical features. The results demonstrate that CDNL facilitates the identification of significant biomarkers and achieves an average improvement in AUC of 4.93% and 8.58% compared to deep learning multi-layer neural network (DNN) and SVM, respectively. Additionally, it shows an average improvement of 2.64% and 1.92% compared to two state-of-the-art multi-task models.