An efficient model for predicting high-cost members
Healthcare costs have been a significant issue in the United States. A large proportion of the cost is potentially avoidable if high-quality care is provided to the appropriate high-cost members. Moreover, only a small proportion of members consume the majority of healthcare costs and medical resources. By predicting more accurately those likely to become high-cost members and providing proactive care to those members, a large amount of the medical cost and resources can be diverted with effective prevention and outreach programs. Accurately identifying high-cost members in the future is critical for proactive care. With the rapid advancements of big data platforms, as well as data science tools and techniques, machine-learning algorithms have been successfully applied for predicting different types of outcomes within the healthcare community. In this paper, we propose an efficient and effective predictive model based on the recent open-source algorithm LightGBM to predict various high-cost members based on different types of features. The predictive model was compared to several current “state-of-the-art” and well-known machine-learning algorithms. Additionally, different feature categories were extracted from the claims data, social determinants of health data, access to care information, and the proposed algorithm was evaluated on different feature groups. The experimental results show that the proposed predictive model combined with a specific feature group with more robust features can yield better performance in terms of different metrics while retaining operational efficiency.