Files
Abstract
Importance: Time spent interacting with electronic health records (EHRs) is strongly associated with clinician burnout. Artificial intelligence (AI) scribes may offer a promising solution to EHR-related burnout. However, previous studies on their effectiveness are limited by selection bias.
Objective: To evaluate the association of an AI scribe with EHR efficiency using a pre-post analysis among AI scribe users and a comparison of AI scribe users with a covariate-balanced control group of nonusers.
Design, Setting, and Participants: This retrospective cohort study included ambulatory clinicians at an academic health system during a 3-month pilot period (July 1 to September 30, 2024).
Exposure: Use of an AI scribe.
Main Outcomes and Measures: Primary outcomes were time spent in the EHR, time spent in notes, and after-hours time spent documenting (“pajama time”) (all per appointment). Secondary outcomes were time to close encounters, appointment length, and monthly appointment volume. Two analyses were conducted: a within-individual pre-post comparison of AI scribe users (n = 125) and nonusers (n = 478), and a between-group comparison of AI scribe users and nonusers using propensity score overlap weighting to balance covariates.
Results: A total of 125 AI scribe users (83 women [66.4%]; 69 [55.2%] with >10 years in practice; 46 [36.8%] in a medical subspecialty, 45 [36.0%] in surgery, and 34 [27.2%] in primary care) and 478 covariate-balanced AI scribe nonusers (267 women [55.9%]; 248 [51.9%] with >10 years in practice; 233 [48.7%] in a medical subspecialty, 155 [32.4%] in surgery, and 90 [18.8%] in primary care) were included. In the pre-post analysis, AI scribe users experienced significant reductions in median time in the EHR per appointment (baseline: median, 22.2 minutes [IQR, 12.1-37.0 minutes]; intervention period: median, 20.2 minutes [IQR, 11.5-31.4 minutes]; difference, −2.0 minutes; P < .001), time in notes per appointment (baseline: median, 7.5 minutes [IQR, 4.3-13.4 minutes]; intervention period: median, 7.0 minutes [IQR, 3.6-10.8 minutes]; difference, −0.5 minutes; P < .001), and time to close encounters (baseline: median, 24.4 hours [IQR, 7.7-94.0 hours]; intervention period: median, 17.3 hours [IQR, 5.4-57.0 hours]; difference, −7.1 hours; P < .001), with no significant differences in after-hours time spent documenting, appointment length, or appointment volume. In the weighted generalized linear regression, AI scribe use was associated with an 8.5% (95% CI, −12.8% to −3.9%; P < .001) lower mean EHR time (ie, 2.4 minutes) and a 15.9% (95% CI −21.2% to −10.4%; P < .001) lower mean time in notes (ie, 1.8 minutes) with no significant differences in other outcomes.
Conclusions and Relevance: In this retrospective cohort study, clinicians using an AI scribe spent significantly less time in the EHR and in notes in both pre-post and propensity score analyses. These findings suggest that AI scribes may improve documentation efficiency and reduce clinician workload.