To evaluate the safety and efficacy of an AI workflow for drafting hospital course summaries in discharge documentation, focusing on potential harm and efficiency.
Key Findings:
88% of reviewed summaries were rated as having no potential for harm, indicating a positive safety profile.
One summary was rated as likely to cause moderate harm but was later deemed safe after independent review.
Omissions were reported in 25% of summaries, inaccuracies in 20%, and hallucinations in 2%, underscoring areas for improvement.
Physician burnout scores decreased after implementation, suggesting potential benefits for clinician well-being.
67% of physicians reported perceived time savings, with 32% estimating over 15 minutes saved per summary, indicating efficiency gains.
Interpretation:
The AI system effectively reduced the risk of harm in discharge summaries, though addressing significant omissions remains a critical concern.
Limitations:
The single-unit academic setting limits generalizability of findings.
The small number of participating physicians and voluntary feedback may introduce bias.
Lack of subgroup analyses and a contemporaneous control group limits the robustness of conclusions.
Absence of systematic assessment of error rates in physician-authored summaries complicates the evaluation of the AI system's impact.
Conclusion:
While the AI workflow mitigated hallucinations, addressing omissions will require better integration with structured EHR data and alignment with physician judgment to enhance safety and efficacy.
Biomarker-guided patient selection may help identify patients with sepsis who could benefit from endotoxin-targeted therapy, although confirmatory evidence is still needed.
