Related Experiment Videos
Bridging the trust-adoption gap for AI scribes in rural communities: A machine learning approach using the 2024
Zhaoqiang Zhou1, John Geracitano1, Sandy Hatoum2
1Carolina Health Informatics Program, University of North Carolina at Chapel Hill, NC, USA; School of Data and Information Sciences, University of North Carolina at Chapel Hill, NC, USA.
Background:
Ambient AI scribe tools that capture clinician-patient conversations and generate draft notes are increasingly deployed to reduce documentation burden, but patient-facing acceptance may influence implementation success, especially in rural areas.
Objectives:
To characterize rural respondents' attitudes toward AI scribes across (1) trust in documentation accuracy, (2) perceived impact on patient-provider interaction, and (3) preference for future use, and to examine how attitudes vary by demographic, socioeconomic, health, and digital-access characteristics.
Methods:
We conducted a cross-sectional analysis of 1,050 rural respondents in the 2024 Canadian Digital Health Survey. Each outcome was dichotomized. XGBoost classifiers were trained for each outcome using prespecified predictors (sex, age group, race/ethnicity, education, employment status, household income, chronic disease, self-reported health, and high-speed internet access). Models demonstrated strong overall performance on a held-out test set. Subgroup differences were summarized using marginally standardized predicted probabilities with bootstrap 95 % confidence intervals.
Results:
Predicted endorsement decreased across three attitude domains, from trust in documentation accuracy to interaction benefit and future-use preference. Predicted endorsement was higher among males than females across outcomes (e.g., future-use preference: 0.388 vs 0.313). Higher education and chronic disease were consistently associated with more favorable responses (e.g., future-use preference: graduate degree 0.466 vs less than high school 0.308; chronic disease 0.408 vs no condition 0.298). Compared with White respondents, visible minority, non-Indigenous respondents had lower future-use preference (0.293 vs 0.349), while Indigenous respondents showed higher predicted future-use preference (0.423 vs 0.349). Predicted probabilities were similar by internet access status across all three outcomes.
Conclusions:
Among rural respondents, trust in AI scribe accuracy does not fully translate into perceived interaction benefit or willingness to use AI scribes in future encounters, supporting rollout strategies that prioritize clear communication, privacy transparency, and meaningful choice.
