Machine Learning and Interpretability Study for Predicting 30-Day Unplanned Readmission Risk of Schizophrenia: A Retrospective Study.

PURPOSE

To build a 30-day unplanned readmission (UPR) risk prediction model based on machine learning (ML) and SHapley Additive exPlanation (SHAP) with data obtained from the electronic medical records (EMRs) of patients with schizophrenia, so as to provide support for early intervention in clinical treatment.

PATIENTS AND METHODS

This retrospective study selected 1,123 patients with schizophrenia who were hospitalized at least once from January 1, 2021 to June 30, 2024 according to their EMRs. Models were constructed after screening variables using the multiple linear regression and feature importance methods. The model was constructed using five ML algorithms: logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGB). The area under the receiver operating characteristic curve (AUC) and SHAP were applied to verify the predictive power and interpretability, respectively, of the five models.

RESULTS

The 30-day UPR rate was 30.54% (343/1,123). The important risk factors were number of somatic comorbid diseases, duration of the disease course, length of the latest hospital stay, drug withdrawal history, and sex. The AUC values of the LR, DT, RF, XGB, and SVM models for predicting the 30-day UPR in the testing set were 0.794, 0.717, 0.823, 0.830, and 0.810, respectively.

CONCLUSION

An XGB risk prediction model can accurately evaluate the 30-day UPR of patients with schizophrenia. Combined with SHAP, it can provide patients with personalized risk predictions, thereby assisting medical staff in achieving early discharge plans and transitional care.