Healthcare Cost Prediction for Heterogeneous Patient Profiles Using Deep Learning Models with Administrative Claims Data

• 成果类型: Article; Early Access
• 署名作者: Morid, Mohammad Amin; Sheng, Olivia R. Liu
• 署名单位: Santa Clara University; Arizona State University; Arizona State University-Tempe
• 刊物名称: INFORMATION SYSTEMS RESEARCH
• ISSN/ISSBN: 1047-7047
• DOI: 10.1287/isre.2021.0643
• 发表日期: 2025
• 关键词: RISK-ADJUSTMENT; rheumatoid-arthritis; identifying patients; analytics; selection; capitation; records

摘要

Accurate and fair patient cost predictions, which can lead to healthcare payer cost savings, are essential to support effective decision making regarding health management policies and resource allocations. Patient cost prediction models utilize administrative claims (AC) data collected from multiple healthcare providers, which payers (e.g., government agencies and private insurance companies) rely on for various reimbursement purposes. Both the variety of patient clinical profiles and the multisource nature of the big data from ACs introduce heterogeneity, which undermines both the generalization power and the algorithmic fairness of cost prediction models. In particular, the prediction performance and economic outcomes-such as both underpayments and overpayments-of these models for high-need (HN) patients with multiple and complex chronic conditions differ from those of healthy patients, as their underlying heterogeneous medical profiles are distinct. This study, grounded in sociotechnical considerations for patient cost prediction, presents two key design insights. First, we designed a channel-wise deep learning framework to reduce AC data heterogeneity through effective representation learning, with a separate channel each type of code as well as each type of cost. Second, we incorporated humanistic outcomes and a multichannel entropy measurement into a flexible evaluation design for patient heterogeneity. We evaluate the effectiveness of the proposed channel-wise framework both internally and externally using two real-world data sets containing approximately 111,000 and 134,000 individuals, respectively. On average, channel-wise models substantially reduce prediction errors by 23% compared with the most competitive single-channel counterparts, leading to respective reductions of 16.4% and 19.3% in overpayments and underpayments for patients. The reduction in bias for predictions involving HN patients is more significant than for other patient groups. Our findings offer important implications for decision makers in healthcare and other fields facing similar sociotechnical challenges related to the interplay between diverse population behaviors and data heterogeneity.