Deployment-Oriented Evaluation of Healthcare Reimbursement Cost Prediction

Abstract

The use of healthcare cost prediction models to assist with resource allocation and risk management is on the rise, but their use is limited by the requirement for reliable, easily interpretable explanations. Although there are multiple methods of providing these explanations, their operational use in healthcare settings remains inadequately evaluated. This research analyses four of the available explainable models (i.e., Permutation Feature Importance, Tree Importance, Local Interpretable Model Explanations, and Shapley-based explanations) using a Random Forest healthcare cost prediction model developed on a real-world dataset (comprising 2,302 aggregated patient segments and 54 features) with a coefficient of determination of 0.9957. The analysis used three criteria that are critical to the deployment of these models: steadiness of explanation under data perturbation, correspondence of feature importance across methods, and computational latency. Results show that each of the global explainable models has relatively high steadiness and relatively low latency, making them applicable for real time and regulatory use; however, the local explainable models provide more intuitive instance-level explanations at the expense of lower steadiness and greater computational demands. A tiered deployment framework for method selection is proposed based on these results as guidance for selecting methods based on clinical, regulatory, and operational needs. This research provides insight into the practical use of explainable healthcare cost prediction systems within real-world environments.

Ähnliche Arbeiten