Modeling Mental Health Case-Mix for Quality Improvement—A Comparison of Statistical and AI Models

Abstract

Background/Objectives: With the rising prevalence of mental health (MH) disorders, improving the effectiveness and quality of MH care has become increasingly imperative. To improve patient care outcomes, it is essential to accurately assess staffing needs and compare outcomes across providers to identify best practices. However, without a robust case-mix adjustment system that accounts for disease severity, efforts to measure staffing requirements and evaluate patient outcomes are of limited value. This study aimed to develop such a system by leveraging a large study population, more clinically homogeneous groups, and advanced modeling techniques. Methods: In this retrospective population-based study, over two million MH patients (n = 2,088,174) were grouped into 162 clinically homogeneous categories using Clinical Classifications Software Refined (CCSR) to enhance predictive accuracy. We evaluated the performance of four statistical models and four artificial intelligence (AI) models to identify the model that delivered the highest predictive power. Results: Among the statistical models, the Box–Cox regression yielded the highest predictive power (R2 = 0.42; percent of variation explained [PVE] = 0.300). Among the AI models, CatBoost performed best (R2 = 0.458; PVE = 0.311). While the AI models outperformed traditional statistical models, the improvements were modest. Sensitivity analyses confirmed the robustness of these models. Conclusions: Both the Box–Cox and CatBoost models demonstrated superior predictive performance compared to those reported in the literature. These findings suggest that a case-mix system based on either model can be used for risk adjustment to optimize staffing levels and benchmark patient outcomes for quality improvement.