Development and validation of a machine-learning model for the risk of potentially inappropriate medications in elderly stroke patients

Objective To construct a risk prediction model for potentially inappropriate medications (PIM) in elderly stroke patients based on multiple machine-learning algorithms, providing decision support to identify high-risk patients and ensure rational clinical medication use. Methods A total of 1,252 discharged stroke patients from a tertiary hospital in Anhui Province, China, were included from January 2023 to December 2024. PIM was assessed using the American Geriatrics Society 2023 Updated Beers Criteria ® . Univariate analysis identified factors potentially associated with PIM, and the least absolute shrinkage and selection operator regression analysis was applied to select variables. The dataset was randomly split into training and internal validations sets in a 7:3 ratio. Additionally, a dataset independent of the training set in terms of time was selected, consisting of 240 stroke patients diagnosed at the same hospital from January to February 2025, to serve as an external validation cohort. Four machine-learning models, Random Forest, Elastic Net (Enet), Support Vector Machine Classifier, and Extreme Gradient Boosting were built using the meaningful variables identified after selection. The evaluation of machine-learning models was carried out through the discrimination, calibration, and clinical utility. SHapley Additive exPlanation (SHAP) values were utilized to rank the importance of features and to interpret the best-performing model. Results Among 1,252 patients, 675 (53.91%) had PIM, with 107 types and 1,140 occurrences of PIM. Both in internal and external validation cohort, Enet performed the best. The area under the curve (AUC) of Receiver Operating Characteristic (ROC) curve of Enet in external validation set was 0.894 (0.854, 0.933). The model’s calibration curve closely followed the ideal curve, and the clinical decision curve showed high net benefit within a threshold probability range of 15%–97%. The results indicate that the Enet prediction model exhibits good accuracy and generalizability, offering a basis for guiding clinical treatment. Conclusion The PIM risk prediction model developed using machine-learning can effectively identify PIM, aiding in the implementation of targeted interventions to prevent and reduce the risk of PIM in elderly stroke patients.