Comparative evaluation of ambient digital scribe systems in clinical documentation

To demonstrate a replicable in-house framework for evaluating ambient digital scribe (ADS) systems and, applying it, to assess 3 commercial systems and examine alignment between clinician Likert ratings and a manual transcript-level error audit.

Two standardized clinical encounters-a brief problem-focused visit and a comprehensive annual visit-were simulated and audio-recorded. Three ADS vendors generated notes from identical recordings. A multidisciplinary panel of 56 physicians rated the notes on 7-point Likert scales across 4 domains: relevance, accuracy, quality, and provider confidence. In parallel, physician reviewers cross-referenced each note against the source recording, enumerating hallucinations, omissions, and inaccuracies using a pre-specified rubric.

In this exploratory sample, Vendor A received the highest mean Likert ratings across all 4 domains (Friedman χ²(2) = 15.529, P < .001 short visit; χ²(2) = 10.753, P = .005 long visit). After Bonferroni correction, only Vendor A vs Vendor B on the short visit remained significant (P = .0098). The audit identified fewest errors in Vendor B's notes (n = 3), followed by Vendor A (n = 8) and Vendor C (n = 9). Physicians flagged hallucinations more often than the audit confirmed, a pattern we tentatively label "pseudo-hallucinations."

In this sample, clinician preference did not track audit-identified error counts, raising the hypothesis that note-style features influence ratings alongside documentation correctness; larger samples are required.

Pairing clinician Likert assessments with transcript-level error audits may be a useful adjunct in evaluating ADS and other generative AI clinical tools; confirmation in larger studies is needed before this framework drives procurement decisions.