This study used machine‑learning to predict deprescribing — defined as medication discontinuation or dose tapering — for high‑risk medications among older adults using structured electronic health record (EHR) data. The analysis included 2,979 adults aged 65 years or older and 158 candidate EHR features that captured patient attributes, primary care provider (PCP) characteristics and EHR‑use behaviors, and aspects of deprescribing‑tool design. Eight predictive models were compared, and SHAP (SHapley Additive exPlanations) was applied to evaluate feature importance. The model TabPFN achieved the highest positive predictive value (67.71%), AUROC (74.30%), and AUPRC (60.95%), though overall predictability was considered moderate. PCP EHR‑use measures dominated the top ranked features, accounting for 19 of the 25 highest TabPFN features. The authors note that routinely captured structured EHR fields only moderately predict deprescribing, implying that important determinants are not captured in these fields. They recommend incorporating richer clinical context and performing external validation before models are used to inform personalized deprescribing support.
The analytic cohort comprised 2,979 adults aged 65 years or older. The study drew on structured EHR elements to assemble 158 candidate features. Those features spanned three broad domains: patient characteristics, PCP characteristics and EHR‑use behaviors, and attributes of the deprescribing intervention or tool design. The source reports that all data produced in the study are available upon reasonable request to the authors.
Investigators considered 158 structured EHR features. These included routine patient‑level variables, measures reflecting PCP characteristics, and multiple PCP EHR‑use behavior metrics. The deprescribing‑tool design (intervention design) was also represented among candidate predictors. The report emphasizes that PCP EHR‑use measures comprised a substantial portion of candidate predictors and subsequently accounted for a large share of the top ranked features in model interpretation.
Eight machine‑learning models were trained and compared for their ability to predict deprescribing events (discontinuation or dose tapering). The authors used SHAP to examine the relative importance of features in the best‑performing model. Performance metrics reported for model comparison included positive predictive value (PPV), area under the receiver‑operator curve (AUROC), and area under the precision‑recall curve (AUPRC). The study represents, according to the authors, the first machine‑learning analysis of these trial data for predicting deprescribing.
Among the eight models evaluated, TabPFN achieved the highest reported performance: positive predictive value 67.71%, AUROC 74.30%, and AUPRC 60.95%. Despite these top scores, the authors describe overall predictability as moderate, indicating that model performance, while informative, was not highly predictive across the cohort using only the structured EHR inputs available.
SHAP‑based feature importance analysis identified multiple PCP EHR‑use measures among the highest ranked predictors. Specifically, 19 of the 25 highest‑ranked features in the TabPFN model were PCP EHR‑use measures. The prominence of these measures likely reflects both their prevalence among candidate predictors and their association with the outcome in the dataset. The authors interpret the moderate overall predictability and the dominance of EHR‑use variables to suggest that key clinical determinants of deprescribing decisions are not captured in routine structured fields.
The study provides an initial, informative reference for health systems interested in predicting deprescribing of high‑risk medications using routinely captured EHR data. Key conclusions reported by the investigators are:
The authors emphasize that their findings should inform how health systems target and scale deprescribing support, but they caution against immediate clinical implementation without further model refinement and validation.
The preprint notes that the work received ethical approval from the ethics committee/IRB of Mass General Brigham (2019P002167). The authors state that all relevant ethical guidelines were followed and that patient consent procedures were completed as required. The manuscript is a preprint and has not been peer reviewed, and the authors caution it should not be used to guide clinical practice. All data produced in the study are available upon reasonable request. Funders declared include the National Institute on Aging, PCORI, the National Library of Medicine, and the Swiss National Science Foundation. Competing interests were disclosed for one author who serves as a consultant and has equity in outside companies; all other authors reported no conflicts relevant to this work.