MLNet: Enhancing Joint Predictive Modeling of Chronic Diseases Using Deep Learning

Changjing Song, Zhanpeng Luo, Li Lu, Qian Su

Abstract: Chronic diseases, such as CKD, diabetes, and hypertension, are widely recognized as major challenges in healthcare, but timely diagnosis of these conditions remains challenging. In this regard, employing machine learning, particularly deep learning, to process EHR data for disease prediction presents a viable solution. Furthermore, leveraging the correlations among different chronic diseases for joint prediction might enhance effectiveness and reduce time and economic costs. In this work, we conduct research on the joint prediction of CKD, diabetes, and hypertension - three chronic diseases affecting the 18-65 age group - based on the MIMIC-IV dataset. We propose the MLNet and develop a comprehensive data preprocessing pipeline. This includes using a multi-label feature selection method based on mutual information to select optimal feature subsets and employing the Focal Loss to address data imbalance. For the task of jointly predicting the three chronic diseases, MLNet outperforms six other predictive algorithms, achieving a best Micro AUC of 90.1%. Among the three chronic diseases, MLNet demonstrates the best predictive performance for CKD, with an AUC of 95.9% and a recall rate of 95.8%.