JMIR Publications

ABSTRACT

Background:

Generative artificial intelligence (AI) is increasingly embedded directly within hospital electronic health record (EHR) systems to assist clinical documentation. Most evidence to date concerns general-purpose chatbots or ambient voice scribes used by attending physicians; how EHR-data-integrated generation tools are taken up across the training continuum, and whether they are associated with reduced self-reported documentation time, remain poorly characterised, particularly outside North American settings.

Objective:

We examined survey-reported use, self-reported documentation-time effects and perceived module-level maturity of an EHR-integrated generative-AI documentation tool (OneRecord, within a HIS 3.0 system) across professional groups, and identified the trainee and tool factors associated with survey-reported use and with self-reported time improvement.

Methods:

We conducted a single-centre, cross-sectional electronic survey of clinical trainees and staff, reported per CHERRIES and STROBE. Four professional groups were analysed: undergraduate clinical learners (UGY), residents (R), postgraduate-year trainees (PGY) and nurse practitioners (NP). Respondents reported documentation time for admission and progress notes with and without AI assistance using ordinal time bands and rated content-module maturity. Paired ordinal changes were assessed with Wilcoxon signed-rank tests, transition matrices and marginal-homogeneity tests (Bowker, Stuart-Maxwell); estimated minute savings (band midpoints, bootstrap 95% CIs) were reported as a secondary metric with sensitivity analyses. Multivariable logistic regression modelled survey-reported use and self-reported admission-time improvement; module maturity (structured-data vs synthesis modules) was compared using a respondent-clustered (GEE) logistic model. All time outcomes were self-reported in ordinal bands and should be interpreted as perceived documentation-time changes rather than objective EHR-measured efficiency.

Results:

Of 282 respondents (UGY n=111, R n=88, NP n=49, PGY n=34), 165 (58.5%) had used the EHR-integrated tool. In multivariable analysis, PGY trainees were substantially more likely to have used the tool than UGY (adjusted odds ratio [aOR] 6.37, 95% CI 2.27-17.84). Among 165 tool users, admission-note time distributions shifted toward shorter categories (Stuart-Maxwell P<.001; 53% improved, 95% CI 46%-61%), with an estimated mean saving of 6.0 minutes (95% CI 4.8-7.3) that was robust across sensitivity assumptions. Progress-note changes were smaller (Stuart-Maxwell P=.002). After adjusting for baseline documentation time, professional group was not independently associated with improvement (baseline category aOR 1.95, 95% CI 1.35-2.83), indicating that larger apparent gains among early learners reflected longer baseline times. Structured-data modules were far more likely than synthesis modules to be rated usable without further optimisation (aOR for synthesis vs structured 0.29, 95% CI 0.21-0.41; 55.5% vs 24.9%). Because no invited-user denominator was available, all estimates of use refer to survey-reported use among respondents.

Conclusions:

Over half of the analysed respondents (165/282, 58.5%) reported using the EHR-integrated generative-AI documentation tool. Tool users reported reductions in documentation-time categories, especially for admission notes. However, early clinical learners remained the longest-duration documenters, and improvement was explained primarily by baseline documentation time rather than by training stage. Module-level ratings showed higher perceived maturity for structured-data summarisation than for clinical synthesis. These findings support reframing AI-assisted documentation as a supervised practice of verification, editing and accountability. We propose Digital Charting Entrustment as a candidate EPA-aligned framework to guide future assessment research.