Predicting hospital readmission in patients with mental or substance use disorders: A machine learning approach

• Published in: International journal of medical informatics (Shannon, Ireland) Ireland), 2020-07, Vol.139, p.104136-104136, Article 104136
• Main Authors: Morel, Didier, Yu, Kalvin C., Liu-Ferrara, Ann, Caceres-Suriel, Ambiorix J., Kurtz, Stephan G., Tabak, Ying P.
• Format: Article
• Language: English
• Subjects: Machine learning; Mental disorder; Mental health; Readmission; Substance use disorders
• ISSN: 1386-5056; 1872-8243
• DOI: 10.1016/j.ijmedinf.2020.104136

•Mental or substance use disorders are major contributors of disease burden with high hospital readmission rates.•Previous logistic regression predictive model for readmission in mental health patients shows low predictive ability.•Using ML, our predictive model for readmission in mental health patients demonstrated higher predictive ability.•We illustrated potential applications of our ML model for benchmarking of standardized readmission rates.•Our model may be further validated to aid targeted demographic initiatives to reduce readmissions and benchmarking. Mental or substance use disorders (M/SUD) are major contributors of disease burden with high risk for hospital readmissions. We sought to develop and evaluate a readmission model using a machine learning (ML) approach. We analyzed patients with continuous enrollment for three years and at least one episode of M/SUD as the primary reason for hospital admission. The outcome was readmission within 30-days from discharge. Model performance was evaluated using the Area under the Receiver Operating Characteristic (AUROC). We compared the AUROCs of an extreme gradient boosted tree (XGBoost) model to generalized linear model with elastic net regularization (GLMNet). We analyzed 65,426 unique patients and 97,688 admissions. Patients with mental disorders accounted for 66 % (13.2 % readmission rate) and substance use disorders, 34 % (22.3 % readmission rate). Among all those who had readmissions, 70.7 %, 17.0 %, and 12.4 % had 1, 2, or 3+ readmissions, respectively. Previous hospitalizations, hospital utilization, discharge disposition, diagnosis category, and comorbidity were among the highest important features in the XGBoost model. The XGBoost model AUROC was 0.737 (95 % CI: 0.732 to 0.742) versus the GLMNet 0.697 (95 % CI: 0.690 to 0.703). The AUROC of the final XGBoost model on the testing set was 0.738 (95 % CI: 0.730 to 0.748), higher than published readmission models for mental health patients. The XGBoost model has a better performance than GLMNet and previously published models in predicting readmissions in mental health patients. Our model may be further tested to aid targeted demographic initiatives to reduce M/SUDs readmissions and benchmarking.