Predicting One-Year Psychiatric Rehospitalization Using Natural Language Processing and Machine Learning Methods Based on Clinical Texts.

Background/

Objectives

Psychiatric readmissions are an important indicator for assessing disease progression and the effectiveness of healthcare services. The aim of this study was to explore factors associated with one-year readmission and to evaluate the performance of machine learning models using clinical information extracted from discharge summaries of patients receiving inpatient treatment in a psychiatric ward.

Methods

This retrospective study included 622 patients hospitalized in the psychiatry department of a university hospital between 2020 and 2024. Clinical variables were extracted from discharge summaries using a theme-based structured coding approach. Factors associated with 1-year readmission were evaluated using univariate and multivariable logistic regression analyses, while length of hospital stay was analyzed using log-transformed linear regression. The predictive performance of logistic regression, Random Forest, HistGradientBoosting, and BERTurk-based machine learning models was also compared.

Results

The 1-year psychiatric readmission rate was 17.0%. In univariate analyses, previous psychiatric hospitalization, longer index length of stay, and manic symptoms were associated with readmission. In multivariable analysis, only previous psychiatric hospitalization remained an independent predictor of readmission (adjusted OR = 2.14, 95% CI: 1.34-3.42; p = 0.001). Previous psychiatric hospitalization, unemployment, and depressive symptoms were independently associated with longer hospital stay. Among the machine learning models, logistic regression using structured clinical variables and theme-based features achieved the best performance (AUC = 0.602).

Conclusions

Previous psychiatric hospitalization was the strongest independent factor associated with 1-year psychiatric readmission. These findings suggest that routinely collected clinical records may contain potentially useful information for understanding psychiatric readmission risk and support further investigation in larger, multicenter studies.