AI-Supported Feedback Raised Sepsis Bundle Compliance From 70 to 83 Percent

Sepsis care has a timing problem, and hospital learning has one too.

Clinicians must recognize severe infection and complete a web of time-sensitive steps. Hospitals then audit whether those steps happened. But the review can arrive months after the patient left, by which time the lesson is distant from the decisions it was meant to improve.

A randomized trial at two academic emergency departments tested a different rhythm. A small large language model read each eligible sepsis record at discharge, judged compliance with the federal SEP-1 quality bundle, and helped generate individualized feedback for the attending physician. The feedback arrived shortly after care instead of through the usual sampled review three to four months later.

Across 301 patient encounters, physicians assigned to the AI-supported feedback system achieved 82.9 percent SEP-1 compliance. Physicians receiving standard feedback achieved 70.1 percent. The model-adjusted absolute difference was 13.0 percentage points, with a 95 percent confidence interval from 2.5 to 23.4 points.

The advance is not an autonomous sepsis doctor. It is a faster learning loop: AI turns complicated records into a prompt for human review, education, and better documentation while the experience is still recent.

The Trial Changed Feedback Rather Than Bedside Authority

Researchers randomized 66 attending emergency physicians, 33 to each group, at UC San Diego’s Hillcrest and Jacobs medical centers. The trial ran from December 13, 2024, through July 8, 2025.

The unit of randomization was the physician. The measured outcomes came from patients under those physicians’ care. The final patient population included 301 encounters meeting Centers for Medicare and Medicaid Services criteria for severe sepsis or septic shock: 180 in the intervention group and 121 in the control group.

The study did not give an AI system control over treatment. It used Llama 3.1 8B Instruct to review structured and unstructured records against SEP-1, a detailed process measure covering time-sensitive tests, antibiotics, fluids, and reassessment. The system ran each case three times and used majority voting. Three experienced team members reviewed every in-measure case and made the final compliance determination.

When an intervention-group case met every requirement, the physician received a congratulatory email and a reminder of the health system’s sepsis policy. When it did not, the physician received an email identifying the missed elements and suggesting how to improve care or documentation. Some physicians also received a phone call from the sepsis program’s medical director.

Control physicians could receive similar feedback only when a case entered the hospital’s normal reporting sample—typically three to four months after discharge.

That design matters. The trial tested a package: AI record abstraction, human adjudication, rapid personalized feedback, policy reminders, and occasional physician outreach. The 13-point difference cannot be attributed to the language model alone.

Compliance Rose but Patient Outcomes Did Not

The primary analysis counted 82 of 117 assessable control encounters as compliant, or 70.1 percent, compared with 145 of 175 intervention encounters, or 82.9 percent. The mixed-effects model accounted for patient encounters clustered within physicians and produced an odds ratio of 2.10, with a 95 percent confidence interval from 1.15 to 3.81 and a P value of 0.02.

The largest difference appeared in the crystalloid-fluid component. The study recorded noncompletion in 16 of 121 control encounters, or 13.2 percent, and 3 of 180 intervention encounters, or 1.7 percent.

That finding needs careful interpretation. SEP-1 permits less than the default 30 milliliters per kilogram when a clinician documents a reason such as heart failure, kidney failure, or concern about fluid overload. The proportion actually receiving at least 30 milliliters per kilogram was 66.7 percent in the control group and 74.0 percent in the intervention group, a difference that was not statistically significant. Some of the measured improvement may therefore reflect better documentation of appropriate exceptions rather than more fluid being given.

The trial found no significant difference in intensive care admission: 20.7 percent in control and 20.0 percent in the intervention group. Thirty-day mortality was 10.7 percent in control and 13.9 percent in the intervention group; the odds ratio was 1.34, with a wide 95 percent confidence interval from 0.65 to 2.77 and a P value of 0.43.

Those results do not establish benefit or harm for survival. The study was powered to detect a difference in process compliance, not mortality, and its authors noted that the relationship between SEP-1 compliance and survival remains contested.

Research references: [1]

The Important Shift Is From Reporting to Learning

Most clinical AI stories focus on prediction: finding disease, forecasting deterioration, or recommending a treatment. This trial explored a quieter use—helping a health system learn from what already happened.

Traditional abstraction is expensive and slow because human reviewers must reconstruct a sequence of care from scattered notes, orders, laboratory results, and timestamps. The trial’s AI system made that reconstruction fast enough to support feedback near discharge.

The reported final AI assessment agreed with expert reviewers 92 percent of the time. That figure came from a workflow that included iterative system improvements and adjudication, so it should not be read as a locked-model external validation. It suggests useful automation inside this study’s setting, but not a basis for removing expert oversight. The intervention is better understood as AI-enabled quality assurance than as an AI replacement for reviewers.

The human possibility is a hospital that learns from every eligible case instead of a small monthly sample—and does so while clinicians can still connect the feedback to the choices they made. If independently replicated, the same pattern could support other complex quality measures where delayed auditing weakens improvement.

What the Evidence Does and Does Not Establish

The randomized design supports a causal claim about the combined feedback intervention and SEP-1 compliance within these two emergency departments. It does not establish that AI improved survival, shortened hospitalization, prevented intensive care admission, or delivered the same result in another health system.

The study also has design boundaries. It ran for seven months in one academic health system. The intervention and control groups accumulated different numbers of eligible patient encounters because physician shift assignments differed and some physicians left emergency-department duties. Thirteen randomized physicians were excluded from the final analysis because they changed duties or had no eligible encounters. Resident physicians provided much of the direct care but did not receive the feedback, although attending physicians remained responsible for supervision.

The trial was registered after completion, at the journal editors’ request, because the investigators initially treated it as quality improvement. The protocol and statistical analysis plan were published with the article, but retrospective registration weakens the protection that prospective registration provides against outcome switching.

Commercial relationships also deserve a clear record. Clairyon, a UC San Diego technology partner, supported work connected to several investigators. Authors disclosed cofounder, equity, chief medical officer, grant, patent, and other relationships involving Clairyon, along with relationships involving other health and technology companies. The study also received support from National Institutes of Health grants, a UC San Diego Department of Medicine pilot grant, and the Joan and Irwin Jacobs Center for Health Innovation. The paper states that funders had no role in the study’s design, conduct, analysis, interpretation, or publication decision.

Research references: [3]

A Warranted Conclusion

The study demonstrates a practical capability: AI-assisted record abstraction can help compress a hospital feedback cycle from months to near discharge, and in this trial the combined system raised compliance with a complex sepsis process measure by 13 percentage points.

That is meaningful progress in how hospitals learn. It is not yet evidence that patients lived longer or recovered faster. The next test is whether this faster learning loop can be reproduced across different hospitals, sustained over time, and connected to outcomes that patients feel—not only metrics that institutions report.