JMIR Publications

ABSTRACT

Background:

Pharmacogenomics uses patient genetic information to tailor medication selection and dosing and to reduce adverse drug reactions. Despite its potential, clinical adoption remains limited. Existing predictive models address what happens after a pharmacogenomic service has been offered, using electronic health records to identify patients who complete testing once it is already on the table. The earlier question of which patients would accept the service in the first place has not been examined as a prediction problem.

Objective:

This study aimed to develop and evaluate an interpretable, calibrated machine learning model that predicts patient willingness to accept pharmacist-provided pharmacogenomic services from patient-reported survey data, and to identify which patient-level factors are most associated with that willingness.

Methods:

The 2021 National Consumer Survey on Medication Experiences and Pharmacists' Roles, a national quota-sampled survey of 1521 adult US residents, was analyzed. The target was willingness to have a pharmacist review genetic test results and advise the physician. Six classifiers were compared under stratified 5-fold cross-validation with class weighting, with logistic regression as the designated baseline. The selected model was recalibrated with sigmoid scaling fitted within cross-validation, its threshold set by the Youden J statistic on calibrated cross-validation predictions, and it was evaluated once on a held-out test set (n=457). The central finding was tested by domain ablation with bootstrap confidence intervals, a Parsimonious Relationship Model and a general-positivity adjustment for shared-method variance.

Results:

Discrimination did not differ significantly across models (DeLong P=.058 for the 2 strongest), and logistic regression, selected for interpretability, was also the best-performing model. On the held-out test set it achieved an area under the curve (AUC) of 0.652 (95% CI 0.600-0.702) with well-centered calibration (slope 1.005, 95% CI 0.26-1.36; intercept 0.001, 95% CI −0.26 to 0.65; Brier score 0.215) and net benefit over default strategies in decision curve analysis across threshold probabilities of approximately 0.33 to 0.90. The central finding was that the patient-pharmacist relationship was the only domain with a significant unique contribution: domain ablation reduced AUC by 0.049 (95% CI 0.025-0.073). A relationship-only 3-feature model recovered 97% of the full model's discrimination (AUC 0.638). After adjusting for general positivity, the relationship domain still added signal (AUC gain 0.061, 95% CI 0.027-0.096), and feeling listened to (odds ratio [OR] 1.41, P=.001) and an established relationship (OR 1.29, P<.001) remained significant. The finding was stable across alternative outcome cut points.

Conclusions:

Willingness to accept pharmacist-provided pharmacogenomic services is most strongly associated with the quality of the patient-pharmacist relationship, a potentially modifiable factor, rather than with demographics, clinical burden, or attitudes alone. The calibrated model offers an interpretable prototype decision aid intended to support, not replace, clinical judgment. Because the data are cross-sectional and self-reported, the association cannot be interpreted causally.