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

BACKGROUND

The burden of serious and persistent mental illness such as schizophrenia is substantial and requires health-care organizations to have adequate risk adjustment models to effectively allocate their resources to managing patients who are at the greatest risk. Currently available models underestimate health-care costs for those with mental or behavioral health conditions.

OBJECTIVES

The study aimed to develop and evaluate predictive models for identification of future high-cost schizophrenia patients using advanced supervised machine learning methods.

METHODS

This was a retrospective study using a payer administrative database. The study cohort consisted of 97,862 patients diagnosed with schizophrenia (ICD9 code 295.*) from January 2009 to June 2014. Training ( = 34,510) and study evaluation ( = 30,077) cohorts were derived based on 12-month observation and prediction windows (PWs). The target was average total cost/patient/month in the PW. Three models (baseline, intermediate, final) were developed to assess the value of different variable categories for cost prediction (demographics, coverage, cost, health-care utilization, antipsychotic medication usage, and clinical conditions). Scalable orthogonal regression, significant attribute selection in high dimensions method, and random forests regression were used to develop the models. The trained models were assessed in the evaluation cohort using the regression , patient classification accuracy (PCA), and cost accuracy (CA). The model performance was compared to the Centers for Medicare & Medicaid Services Hierarchical Condition Categories (CMS-HCC) model.

RESULTS

At top 10% cost cutoff, the final model achieved 0.23 , 43% PCA, and 63% CA; in contrast, the CMS-HCC model achieved 0.09 , 27% PCA with 45% CA. The final model and the CMS-HCC model identified 33 and 22%, respectively, of total cost at the top 10% cost cutoff.

CONCLUSION

Using advanced feature selection leveraging detailed health care, medication utilization features, and supervised machine learning methods improved the ability to predict and identify future high-cost patients with schizophrenia when compared with the CMS-HCC model.