The Work After the Draft

The randomized trial showed what happened to a narrow time measure; the UCI analysis asks what clinicians actually changed after an ambient system produced a note. Its subject is Abridge-generated documentation, paired with the clinician-finalized version of the same note section. The useful question is whether post-editing is mainly a light vocabulary cleanup or a deeper act of compression, organization, and clinical framing. That distinction reveals the work a time-in-note measure may not see.

The answer is descriptive rather than causal. This was a retrospective analysis of text preserved before and after clinician editing, not a randomized comparison of people with and without Abridge. It cannot tell us how many minutes an edit took or whether the draft was faster than composing from scratch. It can show, at large scale, the kinds of transformations that occur between machine-generated language and a final clinical record.

The unit of analysis was a note section, not an encounter

The study examined 71,173 paired AI-draft and finalized note sections from 34,726 encounters, authored by 241 clinicians. The data came from a UCI Health pilot that ran from December 2023 through April 2025. Clinicians accessed Abridge through Epic Haiku and a web interface; encounter audio was transcribed after patients were notified, and the draft and final text were preserved through EHR audit trails.

A note section is a narrower unit than a visit. The researchers grouped text with the same note identifier and section label into a draft-final pair. A single encounter could generate more than one clinical note, and a note could contain several sections. The analysis therefore did not treat 71,173 sections as 71,173 independent patient encounters. It treated each section-level pairing as an opportunity to compare what the system proposed with what the clinician kept.

The four sections had different jobs: History of Present Illness, or HPI; Assessment and Plan, or A&P; Physical Exam; and Results. That choice matters because a section’s communicative purpose changes what counts as a useful edit. HPI often carries the patient’s account, chronology, and context. A&P expresses the clinician’s interpretation and management decisions. Physical Exam and Results are more constrained by established clinical formats. A change in one section cannot automatically be interpreted as the same kind of work in another.

The pairing also gives the analysis a useful before-and-after structure. It does not compare one clinician with another or rely on memory about what changed. It places the AI draft and the final text from the same note-section side by side. That makes it possible to ask whether a phrase disappeared, whether a different phrase appeared, and whether the two changes occurred together. But the structure remains observational: the text shows an outcome of editing, not the clinician’s complete reasoning while editing.

The transformation test deliberately counted only a narrow kind of word change

The researchers built a consumer-to-clinical terminology dictionary from the Consumer Health Vocabulary and other plain-language health resources. A consumer expression is a lay or conversational phrase such as “high blood pressure,” while its mapped clinical equivalent might be “hypertension.” The study’s algorithm did not count every replacement as evidence of professionalization. It required two conditions within the same note-section.

First, the consumer term had to appear in the AI draft and be absent from the finalized text. Second, the clinical term mapped to it had to be absent from the draft and newly present in the final text.