Objective:

The primary aim of this study is to predict the risk of 30-day behavioral health (BH) readmission in a Medicaid population. The secondary aim is to identify risk factors of 30-day BH readmission.

Methods:

Highmark Health Options members aged 18 or older in Delaware were included if they had any behavioral health disorder diagnosis coded as part of any inpatient hospital stay between 2015-01-01 and 2020-12-31. BH admissions were defined as admissions with BH International Classification of Diseases (ICD)-10 Codes listed in any diagnosis field of the claim. The index admission was established as any admission of BH treatment. A 30-day readmission event was defined as any BH admission occurred 1-30 days after the discharge date of the index admission. We used Random Forest (RF), XGBoost (XGB) and Google TabNet. This prediction was based on claims, social determinants of health, and demographic data in the past 365 days of index admission. The imbalance of outcome was handled by Synthetic Minority Oversampling Technique and Tomek Links. A 10-fold repeated stratified cross-validation was used for validation.

Results:

There were 26,786 index admissions from 8,710 unique members with 6,099 30-day BH readmission events. XGBoost had the higher score of Area Under the Curve (AUC) and precision than RF and TabNET (AUC: XGB-0.7388 , RF-0.6948, TabNet-0.6531; precision: XGB-0.6715, RF-0.6104 , TabNet-0.4125), while TabNet had the highest recall (TabNet - 0.5901, XGB - 0.5721 , RF - 0.4967). The risk factors with the highest impact were gender, age, serious and persistent mental illness issues (SPMI), and low chronicity (i.e., having been diagnosed with BH issues for less than 3 times).

Conclusions:

Using machine learning models, we were able to identify members with a high likelihood of 30-day BH readmission events. Gender, age, SPMI and chronicity were the high-impact risk factors of 30-day readmission events.