Early prediction of high-cost inpatients with ischemic heart disease using network analytics and machine learning

摘要

Although identifying high-cost inpatients with ischemic heart disease (IHD) at the point of admission is helpful for timely intervention and reducing costs, it is a difficult task due to the limited data available at such an early stage of treatment. The aim of this study was to tackle this challenge by integrating network analytics with machine learning. In this study, we first constructed two networks including Phenotypic Comorbidity Network and Distance-based Disease-Cost Network based on the hospital discharge records (HDR) of Chengdu, China during 2015–2017. Based on the two networks, three novel network features were generated to capture the potential relationships between comorbidities and high healthcare costs. Six machine learning models (Logistic Regression, Decision Tree, Neural Network, Random Forest, XGBoost, and LightGBM) with different input features (network and non-network features) were developed to compare their performance using HDR of 323,907 admissions for IHD between 2018 and 2019. Our results showed that LightGBM with the input feature set including network and non-network features achieved the best area under the receiver-operating characteristic curve (AUC) of 0.900. Network features were the most significant features accounting for the majority (47.4%) of feature importance. Models using network features provided more accurate prediction compared with those using only non-network features. In conclusion, the proposed approach combining network analysis and machine learning is useful for predicting high-cost inpatients with IHD at an early stage of admission, which highlights the potential value of developing and disseminating network-based machine learning in healthcare field.