JMIR Publications

ABSTRACT

Background:

Documentation burden contributes to physician burnout, with trainees spending up to 50% of clinical time on electronic health record (EHR) tasks. Ambient artificial intelligence (AI) scribe technology uses speech recognition and large language models to generate clinical notes from patient-clinician conversations, yet evidence of its impact on trainee workload in dermatology remains limited.

Objective:

This study evaluated the impact of an ambient AI scribe (Abridge) on subjective workload and documentation burden among dermatology residents at an academic medical center.

Methods:

We conducted a single-arm, pre-post study at Yale School of Medicine. Pre- and post-implementation surveys were administered via Qualtrics and matched by participant identity, yielding 13 paired respondents (from 18 pre- and 14 post-survey completions; 1 non-user excluded). The primary outcome was the NASA Task Load Index (Raw TLX, 6-item composite, 0–100 scale; lower = less workload). Secondary outcomes included the AMIA TrendBurden instrument (5 items, 0–100 visual analog scale [VAS]), self-reported documentation time at 3 clinical timepoints, and 9 post-implementation experience items. Paired analyses used Shapiro-Wilk normality testing to select between paired t-tests and Wilcoxon signed-rank tests. Bonferroni correction was applied to TrendBurden comparisons (α = .010). Effect sizes were calculated as Cohen's d.

Results:

Participants were predominantly female (62%), aged 25–34 years (85%), and distributed across PGY-2 (31%), PGY-3 (23%), and PGY-4 (46%) training levels. The NASA-TLX composite score decreased significantly from 67.3 (SD 13.2) at baseline to 44.0 (SD 16.6) post-implementation (P = .003; Cohen's d = 1.01), representing a 34.7% reduction. Five of 6 individual workload dimensions improved: Mental Demand (P = .006), Physical Demand (P = .004), Temporal Demand (P < .001), Effort (P = .011), and Frustration (P = .010). Performance self-assessment was unchanged (P = .636). Two of 5 TrendBurden items were significant with Bonferroni correction: "Documentation impedes quality care delivery" decreased from 78.6 to 36.6 (P = .002), and "Ease of EHR documentation" improved from 34.7 to 61.5 (P = .006). All 5 positive post-implementation experience items exceeded the scale midpoint (means 59.6–68.2 on a 0–100 VAS), and trainees endorsed continued use of the AI scribe (mean 75.9). Concerns about AI hallucinations (mean 41.2) and missed information (mean 33.2) remained below the scale midpoint.

Conclusions:

Implementation of an ambient AI scribe significantly reduced subjective workload and documentation burden among dermatology residents, with a large effect size replicating prior multi-site findings. Two documentation burden items survived Bonferroni correction, and trainee concerns about AI accuracy remained below the scale midpoint. These findings support integration of ambient AI scribes into graduate medical education in dermatology. Clinical Trial: Not Applicable.