Explainable prediction self-assessment model for potentially inappropriate prescribing risk in older adults

Abstract

Background

Older adults with multiple chronic conditions often face challenges of potentially inappropriate prescribing (PIP). This study aimed to develop and validate an explainable machine learning (ML) model to predict PIP risk, aiding clinicians and patients in improving medication safety and self-management.

Methods

Data from geriatric outpatient prescriptions in six Chinese cities were analyzed using Chinese criteria. LASSO regression identified risk variables. Three machine learning (ML) models—logistic regression (LR), random forest (RF), and neural network (NN)—were training and internal validation (7: 3) and external validation cohort. Model performance was assessed via area under the ROC curve (AUC). SHapley Additive exPlanation (SHAP) values explained variable importance, and risk cutoff points were determined using the Youden index and prevalence data.

Results

Among 131,894 prescriptions, 29.00% (38,245) were PIP. The NN model with nonsampling performed best, with internal validation AUC of 0.759 (95%CI: 0.753 –0.764) and external validation AUC of 0.842 (95%CI: 0.816 –0.867). SHAP summary plots showed that the number of medications and sleep disorders were the most influential variables in basic prescription information and diagnoses, respectively. A predicted probability cutoff point of 29% was determined to classify low- and high-risk PIP categories. The optimal model was deployed as a web application (https://stoppip.online/pipview) for clinical use, and a WeChat mini-program was developed to facilitate self-assessment of PIP risk during outpatient follow-ups or home medication use.

Conclusion

The model was not only developed to predict PIP but can also be used by medical staff and older patients for self-assessment.