Machine learning prediction of discharge destination in patients with Parkinson's disease; a nationwide cohort study.

Risk stratification during hospitalization may support real-world discharge planning. We developed and validated machine learning models and an interpretable risk score to predict discharge destination among patients hospitalized with Parkinson's disease using a nationwide administrative claims database. Adults aged ≥50 years hospitalized between November 2017 and June 2023 were included, and the first hospitalization was defined as the index admission. Discharge destination was categorized as home, facility, or in-hospital death. The dataset was randomly divided into training (80%) and testing (20%) cohorts. Random forest models were constructed for all discharge outcomes, and an elastic net logistic regression model was developed for facility discharge. Among 281,664 index admissions, 48.0% were discharged home, 44.8% to a facility, and 7.2% died in hospital. The random forest models achieved AUCs of 0.775 for home discharge, 0.774 for facility discharge, and 0.832 for mortality. The elastic net model demonstrated an AUC of 0.752. A seven-item risk score identified a high-risk group with a 73.8% facility discharge rate compared with 40.6% in the low-risk group. These models provide clinically interpretable risk stratification to support multidisciplinary discharge planning.