Objective To construct an explainable machine learning （XML）-based model for predicting rehabilitation-related discharge in stroke patients， to identify the key influencing factors for rehabilitation-related discharge， and to provide data support for rehabilitation assessment and healthcare resource allocation.Methods Data were extracted from the MIMIC-IV v3.1 database， and a total of 14 824 stroke patients were identified based on the ICD-9/10 codes. Structured variables including demographic data， admission features， and hospitalization data were extracted to construct predictive models for rehabilitation-related discharge. Logistic regression and extreme gradient boosting （XGBoost） algorithms were used to construct models， and these models were compared in terms of discriminatory ability， calibration performance， and net clinical benefit. The SHapley Additive exPlanations （SHAP） algorithm was used to evaluate the contribution of each feature to the results of prediction， and partial dependence and individual conditional expectation plots were used for interpretability analysis.Results In the overall samples， the rate of rehabilitation-related discharge was 36.9%. The logistic and XGBoost models had an area under the curve of 0.637 （95%CI： 0.620-0.653） and 0.630 （95%CI： 0.613-0.647）， respectively， indicating a level of moderate discriminatory performance. Both models achieved an average precision of 0.473， with a Brier score of 0.220 and 0.223， respectively， suggesting that both models had good calibration. The decision curve analysis showed that the logistic model provided the greatest net benefit within the threshold range of 30%- 40% （P<0.05）. The SHAP analysis showed that age （mean SHAP=0.280）， insurance type （mean SHAP=0.266）， and admission route （mean SHAP=0.237） were the main influencing factors. The partial dependence analysis showed the highest probability of rehabilitation-related discharge around the age of 40 years， which then decreased with the increase in age （P<0.001）； the patients admitted through emergency or transfer had a significantly higher probability of rehabilitation-related discharge than those admitted through outpatient service or self-admissions （P<0.001）； there was no significant difference between the patients with different subtypes of stroke （P=0.236）. The AUC of the model remained stable （0.60-0.70） across 2008-2019， suggesting that the model had good robustness. The high predicted risk group had a significantly higher rate of rehabilitation-related discharge than the low predicted risk group （P<0.001）， and the predicted probability was weakly negatively correlated with the length of hospital stay.Conclusions The predictive model for rehabilitation-related discharge in stroke patients is constructed based on XML and SHAP， and the results show that age， insurance type， and admission route are the main characteristic variables affecting rehabilitation-related discharge. The model shows favorable calibration and robustness in real-world data， providing a feasible quantitative tool for early identification of rehabilitation in stroke patients and optimization of resource allocation.