Null on patient outcome (RCT)

A GPT-4o assistant built into the medical record improved Kenyan clinicians' notes and plans but did not reduce treatment failure for 9,691 patients.

Arandomized isolated pre-windowunreviewed

A pragmatic cluster-randomized trial in 16 Penda Health clinics in Nairobi (Nature Medicine, June 2026) gave 52 of 103 clinical officers the AI Consult tool; treatment failure within 14 days was 2.2% with it vs 2.0% without (adjusted OR 0.77, p=0.13), while blinded reviewers rated diagnoses, notes, and treatment plans better in the AI arm.

Sources

primaryNature Medicine: Generative AI-enabled clinical decision support system in primary care — a pragmatic, cluster-randomized trial (open access, PMC)pmc.ncbi.nlm.nih.gov
primaryLSHTM Research Online record (abstract and CC BY PDF)researchonline.lshtm.ac.uk
secondaryUniversity of Birmingham news releasebirmingham.ac.uk

Editor's notes

Pairs with the earlier Penda/OpenAI quality-improvement study of an almost identical tool (Korom et al., preprint), which reported fewer clinician errors without patient-outcome randomization. Clinicians fully followed only 19.5% of 1,000 reviewed red alerts. Protocol violations: 45 control-arm encounters had tool access through an EMR configuration error.

History

Oct 22, 2026The editor has not yet read the Nature Medicine paper; published unreviewed.
Oct 22, 2026Added by weekly research run; verified against the open-access full text.

Cite this entry

Permalink https://becauseofai.org/entry/kenya-ai-consult-null

“A GPT-4o assistant built into the medical record improved Kenyan clinicians' notes and plans but did not reduce treatment failure for 9,691 patients.” Because of AI, updated Oct 22, 2026. https://becauseofai.org/entry/kenya-ai-consult-null

Caveat: The event rate was lower than planned, so the trial could miss a modest benefit; one private urban network with a high baseline standard of care.
Achieved: Apr 22 – Jul 16, 2025
Published: Jun 26, 2026 · Published 11 mo after the benefit ended.
Compared with: Clinical officers with EMR plus LLM assistance vs EMR alone, cluster-randomized; primary endpoint expert-adjudicated treatment failure within 14 days
Effect: Adjusted OR 0.77 (95% CI 0.55–1.08), not significant; appropriate diagnosis in documentation aOR 1.74 (102/4,693 (2.2%) vs 94/4,654 (2.0%) treatment failures; death or hospitalization 14 vs 17)
Reach: 9,691 patients, 103 clinical officers, 16 primary care facilities, Nairobi
Conflicts: Funded by the Gates Foundation; run in Penda Health clinics, with two Penda authors holding stock options; OpenAI gave Penda in-kind compute credits and API guidance but had no stated role in the trial.
What would move the grade: A trial powered for patient outcomes (the authors suggest very large samples) showing fewer treatment failures.
Updated: Oct 22, 2026 unreviewed