Predicting Future High-Cost Schizophrenia Patients Using High-Dimensional Administrative Data.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Psychiatric Epidemiology

Background:

• Serious and persistent mental illnesses like schizophrenia impose a significant burden on healthcare systems.
• Existing risk adjustment models often underestimate healthcare costs for patients with mental and behavioral health conditions.
• Accurate prediction of high-cost patients is crucial for effective resource allocation in healthcare organizations.

Purpose of the Study:

• To develop and evaluate advanced supervised machine learning models for identifying future high-cost schizophrenia patients.
• To compare the performance of developed models against the Centers for Medicare & Medicaid Services Hierarchical Condition Categories (CMS-HCC) model.

Main Methods:

• Retrospective analysis of a payer administrative database including 97,862 schizophrenia patients (ICD9 code 295.*) from 2009-2014.
• Development of three predictive models (baseline, intermediate, final) using scalable orthogonal regression, attribute selection, and random forests.
• Model evaluation using R-squared, patient classification accuracy (PCA), and cost accuracy (CA) in a dedicated evaluation cohort.

Main Results:

• The final advanced model achieved a 0.23 R-squared, 43% PCA, and 63% CA at the top 10% cost cutoff.
• The CMS-HCC model achieved a 0.09 R-squared, 27% PCA, and 45% CA for the same cost cutoff.
• The advanced model identified 33% of total costs in the top 10% of patients, compared to 22% for the CMS-HCC model.

Conclusions:

• Advanced supervised machine learning methods, incorporating detailed healthcare and medication data, significantly enhance the prediction of high-cost schizophrenia patients.
• The developed models demonstrate superior performance compared to the established CMS-HCC model for identifying high-cost individuals.
• Improved predictive accuracy facilitates better resource allocation and management strategies for schizophrenia patients.