AI Copilot Reduces Clinical Errors in 40,000-Patient Study at Penda Health

The research, conducted across 15 clinics and nearly 40,000 patient visits, offers early evidence that large language models (LLMs) can improve patient safety in real-world healthcare settings.

A new study by OpenAI and Penda Health has found that an AI-powered clinical copilot helped reduce diagnostic and treatment errors among frontline clinicians in Kenya.

The research, conducted across 15 clinics and nearly 40,000 patient visits, offers early evidence that large language models (LLMs) can improve patient safety in real-world healthcare settings.

The system, called AI Consult, was integrated into Penda Health’s electronic medical records platform. It provides real-time feedback to clinicians during patient consultations, flagging potential safety issues based on the clinical documentation entered during the visit.

Clinicians using the copilot recorded a 16% relative reduction in diagnostic errors and a 13% drop in treatment errors, compared to those working without it. The study was approved by Kenya’s national and regional health regulators, including AMREF Health Africa Ethical and Scientific Review Committee, the Ministry of Health’s Digital Health Agency, and the Nairobi County Department of Health.

The AI Consult tool is powered by OpenAI’s GPT-4o model and was designed in close collaboration with Penda’s clinical teams. Unlike previous iterations that required doctors to manually request input, the latest version runs in the background and flags issues automatically at key points in the visit.

The system returns one of three signals to the clinician:

• Green: No concerns detected
• Yellow: Moderate concern, optional to view
• Red: Safety-critical alert, must be reviewed

According to Penda’s Chief Medical Officer, Dr Robert Korom, this version of AI Consult is the result of multiple rounds of clinical feedback. “We focused on embedding the tool within clinician workflows so that it supports rather than disrupts care,” he said.

Penda Health operates 16 primary care clinics across Nairobi and sees over 500,000 patient visits annually. As a social enterprise, it has maintained a strong focus on quality of care, with structured clinician training programs and past experimentation with early AI copilots.

The copilot’s feedback includes local context such as Kenyan disease prevalence, epidemiology, and national clinical guidelines. Penda said these adjustments were essential to ensure relevance and trust in the AI’s suggestions.

OpenAI and Penda have now published the full findings, including implementation details, as a case study for others exploring how to safely use AI in frontline health systems.

Stay tuned for more such updates on Digital Health News