Nurse-Led Ambient AI Scribe for Patient Safety Incident Investigation Reports (Project NARRATE): Retrospective...

Patient safety investigation reports support organizational learning only when they are complete, usable, and sufficiently detailed. Conventional free-text reports are often inconsistent and may omit information needed for review and learning. Project NARRATE (Nursing AI-Refined for Accurate Transcription of Events) is a nursing-led ambient artificial intelligence workflow that uses prompts aligned with the World Health Organization Minimal Information Model for Patient Safety Incident Reporting and Learning Systems, Situation-Background-Assessment-Recommendation output, and visible missing-information cues to support structured supervisor reporting.

This study aimed to compare the completeness and narrative quality of conventional and NARRATE-period supervisor investigation reports for falls and medication administration-related incidents.

We conducted a retrospective pre-post document-quality study at a tertiary academic medical center in Singapore. We reviewed 150 deidentified completed supervisor investigation reports: 75 conventional reports from June to August 2025 and 75 confirmed NARRATE reports from January to March 2026. NARRATE use was voluntary, and recorded use represented approximately 40% of eligible postimplementation reports. Two blinded reviewers rated reports using a World Health Organization (WHO)-aligned completeness checklist and an adapted 8-domain Physician Documentation Quality Instrument (PDQI). Report-level comparisons were adjusted for repeated reports by the same supervisor using random-intercept linear mixed-effects models. A stratified 60-report plain-paragraph rerating examined whether visible structure influenced ratings.

All 150 reports were analyzed. Unadjusted mean WHO total completeness was 11.81 (SD 3.39) for conventional reports and 13.61 (SD 2.54) for NARRATE reports; the unadjusted difference was 1.80 points, and the cluster-adjusted mean difference was 1.95 (95% CI 0.91-3.00; P<.001). The adapted PDQI mean was 3.61 (SD 0.52) and 4.13 (SD 0.34), respectively; the unadjusted difference was 0.52 points, and the cluster-adjusted mean difference was 0.53 (95% CI 0.37-0.69; P<.001). In the plain-paragraph sensitivity analysis, the completeness advantage remained (adjusted mean difference 1.70, 95% CI 0.27-3.14; P=.02), as did the adapted PDQI mean advantage (adjusted mean difference 0.25, 95% CI 0.06-0.43; P=.009). Explanation, organization, and comprehensibility remained significantly higher after deformatting; actions were borderline (P=.05), and synthesis, internal consistency, and fairness/balance were not statistically significant.

Among voluntary early adopters, NARRATE use was associated with more complete reports and higher adapted PDQI mean scores after accounting for supervisor clustering. Because recorded use represented approximately 40% of eligible postimplementation reports and users self-selected, findings may reflect adopter and supervisor characteristics. Results support the structured workflow as a whole, not any single AI component, and do not demonstrate downstream patient-safety effects. Confirmatory evaluation under broader adoption with a concurrent, reliably classified comparison group is needed.