Related Experiment Videos
To construct a risk prediction model for poor discharge readiness in patients undergoing elective unilateral lower
Dan-Dan Lin1, Yu-Cai Jiang2, Lin-Lin Zheng3
1Department of Research, Affiliated Hospital of Putian University, Putian, Fujian, China.
Objective:
To construct and validate a machine learning model integrating clinical and psychosocial factors for predicting the risk of poor discharge preparation (PDP) in patients undergoing elective unilateral lower extremity joint replacement surgery.
Methods:
This study adopted a retrospective observational cohort design, enrolling 265 adult patients who underwent elective unilateral total hip or knee arthroplasty at a tertiary grade A hospital in Fujian Province from December 2022 to April 2023. Patients were divided into the PDP group and the well-prepared group based on the median total score of the Readiness for Hospital Discharge Scale (RHDS) (97.0 points). A total of 25 variables, including demographic, clinical, fear of movement (TSK-13), and socioeconomic factors, were collected. Variable selection was performed using strongly regularized Elastic Net logistic regression, and XGBoost and Elastic Net logistic regression models were constructed, respectively. Model performance was evaluated using 5 × 5 repeated stratified cross-validation, with metrics including AUC, PR-AUC, sensitivity, specificity, calibration, and decision curve analysis (DCA). The SHAP method was used to explain the prediction mechanism of the XGBoost model.
Results:
Elastic Net identified eight key predictors: length of hospital stay, postoperative kinesiophobia, monthly household income, educational level, preoperative kinesiophobia, age, and occupation. The XGBoost model performed the best, with an AUC of 0.833 (95% CI: 0.811-0.855) and a PR-AUC of 0.750 (95% CI: 0.708-0.790), significantly outperforming the logistic regression model (DeLong test p = 0.0035). The calibration curve showed good consistency, and DCA confirmed that it had the highest clinical net benefit within the threshold probability range of 5-40%. SHAP analysis indicated that the shortest length of hospital stay, high preoperative kinesiophobia level, and low household income were the three core risk drivers.
Conclusion:
This study successfully developed a high-precision and interpretable machine learning prediction tool that can effectively identify high-risk patients with poor discharge readiness. The model highlights the critical impact of shortening hospital stay, preoperative psychological state, and socioeconomic status on rehabilitation outcomes, providing a scientific basis for implementing early and personalized interventions.
