OpenAI and Penda Health AI Clinical Copilot Study
OpenAI and Penda Health have developed and studied AI Consult, an LLM-powered clinical copilot that significantly reduces medical errors in primary care. In a study of 39,849 patient visits across 15 clinics in Nairobi, Kenya, clinicians using AI Consult saw a 16% relative reduction in diagnostic errors and a 13% reduction in treatment errors compared to those without the tool.
Closing the Model-Implementation Gap
While frontier models like GPT-4o and o3 show high performance on benchmarks such as HealthBench and in tasks like diagnostic reasoning, a "model-implementation gap" exists between model capability and real-world clinical use. To address this, OpenAI partnered with Penda Health, a primary care provider, to move beyond simple AI queries and integrate LLMs directly into the clinical workflow.
AI Consult: Technical Implementation and Workflow
AI Consult is integrated into Penda's electronic health records (EHR) and operates as a real-time safety net. It analyzes documentation (stripped of patient identifiers) and sends it to the OpenAI API at key points during a visit. The system provides feedback through a tiered alert system:
- Green: A checkmark indicating no concerns.
- Yellow: A ringing bell indicating moderate concerns, which clinicians can choose to view.
- Red: A mandatory pop-up for safety-critical issues that must be viewed before the clinician can proceed.
To ensure safety and local relevance, the system is designed so that clinicians remain in full control of all final decisions. Prompts are tailored to the Kenyan epidemiological context, local clinical guidelines, and Penda's standard operating procedures.
Impact on Quality of Care
An analysis of 5,666 randomly selected visits by 108 independent physicians revealed that AI Consult significantly lowered error rates across four key dimensions:
| Dimension | Relative Error Reduction |
|---|---|
| History-taking | 32% |
| Investigations | 10% |
| Diagnosis | 16% |
| Treatment | 13% |
For visits where AI Consult would have triggered a "red alert," the impact was even more pronounced, with diagnostic errors reduced by 31% and treatment errors by 18%. In absolute terms, the system is estimated to avert 22,000 diagnostic errors and 29,000 treatment errors annually at Penda alone.
The Role of Active Deployment
The study highlighted that model capability alone is insufficient; active deployment is critical for clinician uptake. During an initial "induction period," clinicians often ignored red alerts, and the "left in red rate" (the percentage of unremedied red alerts) was 35-40%, similar to the non-AI group.
Penda implemented an active deployment strategy to lower this rate, including:
- Peer Champions: Branch managers and peers provided one-on-one coaching and explained the tool's limitations.
- Measurement: Personalized coaching based on tracked interaction rates.
- Incentives: Recognition for clinicians and clinics that utilized the tool effectively.
Following these interventions, the "left in red rate" for the AI group dropped to 20%, while the non-AI group remained near 40%.
Clinician Feedback and Patient Outcomes
Clinicians reported that AI Consult acted as a "consultant in the room" and a "learning tool." Data showed that clinicians using the AI triggered fewer red alerts over time (dropping from 45% to 35% of visits), suggesting the tool helped them improve their own clinical skills.
Regarding patient outcomes, the study found:
- Patient Recovery: 3.8% of patients in the AI group reported not feeling better after eight days, compared to 4.3% in the non-AI group (not statistically significant).
- Safety: There were 7 patient safety reports in the AI group and 5 in the non-AI group. No AI Consult recommendations led to harm, though some could have prevented harm if heeded.
Future Directions
OpenAI and Penda Health view AI Consult as an early archetype. Future iterations may include voice-first interfaces to reduce documentation burdens or agents capable of taking actions within the health record upon clinician confirmation. Penda is currently conducting a randomized controlled trial with PATH to further measure the long-term effects on patient outcomes.