AI‑Based Clinical Decision Support in Primary Care
A real‑world study across 39,849 patient visits in Nairobi
Discover how Penda Health and OpenAI cut diagnostic and treatment errors—while preserving clinician autonomy—in one of the world’s largest live evaluations of an LLM co‑pilot.
WHY THIS STUDY MATTERS
Evidence, not hype
Large language models can ace exam‑style vignettes, but do they help real patients? Our cluster‑assigned quality‑improvement study measured the effect of AI Consult during everyday care in 15 busy Kenyan clinics.
Front‑line‑first design
AI Consult surfaces only when needed, uses a simple “traffic‑light” UI, and lets clinicians override or accept advice with a single click—ensuring workflow fit and clinician trust.
Scalable blueprint
We openly share the implementation playbook—data pipelines, change‑management tactics and regulatory safeguards—so other systems can replicate (or improve upon) our results.
KEY FINDINGS AT A GLANCE
32%
Reduction
History-taking errors
10%
Reduction
Investigation errors
16%
Reduction
Diagnostic errors
13%
Reduction
Treatment errors
75%
Clinicians reporting "quality substantially improved"
“AI provides an extra layer of comfort and confidence. It’s like having a fellow consultant in the room as i attend to my patients”
Adah - Clinician Penda Health
INSIDE THE PAPER
A full methods section, statistical code snippets and survey instrument are included in the appendix for teams who want to run similar evaluations.
109 clinicians, 39,849 patient visits, 15 clinics
Scale & setting
Cluster‑assigned intervention with independent physician raters and supplemental LLM raters to test concordance.
Design
AI Consult only interrupts when probability of harm is high, minimizing alert fatigue.
Safety net, not autopilot
Active coaching and feedback loops made the effect size grow over time as clinicians learned from the tool.
Uptake strategies
START YOUR HIGH‑VALUE AI JOURNEY
Ready to translate evidence into action?
Download the study, watch the overview, and contact our team to explore pilots, joint research or knowledge‑sharing sessions.