Google DeepMind is launching a research initiative focused on developing an AI co-clinician, aiming to enhance healthcare delivery by augmenting doctors' capabilities. The project, detailed on Deepmind, addresses a global shortage of clinical experts.
The AI co-clinician is envisioned as a collaborative team member, interacting with patients under direct clinical oversight. This approach seeks to extend clinicians' reach while maintaining their ultimate judgment and control.
Augmenting Clinician Expertise
The research prioritizes AI systems that are trustworthy and factually grounded, crucial for clinician adoption. In blind evaluations comparing AI responses to leading evidence synthesis tools, physicians consistently preferred the AI co-clinician's output.
Using the "NOHARM" framework, the system was tested for errors of commission and omission. In 98 primary care queries, it recorded zero critical errors in 97 cases, outperforming two widely used AI systems.
The AI also showed significant progress in complex medication knowledge and reasoning, particularly when answering open-ended questions. It surpassed other frontier AI systems on the OpenFDA RxQA benchmark, a challenging test of medical knowledge.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Exploring Multimodal Patient Interaction
Beyond clinician support, DeepMind is investigating the AI's role in patient-facing scenarios. Recognizing that medicine involves more than text, the team is exploring real-time multimodal capabilities using audio and video.
Leveraging advancements from Gemini and Project Astra, the AI co-clinician is being tested in simulated telemedical calls. The goal is to support diagnosis and management, with expert supervision.
In simulations involving synthetic clinical scenarios, the AI demonstrated the ability to guide patients through physical examinations. It successfully corrected inhaler techniques and assisted in identifying injuries.
While expert physicians outperformed the AI overall, particularly in identifying critical "red flags," the AI co-clinician performed comparably to or better than primary care physicians in 68 out of 140 assessed consultation skills.
Engineering Trust and Safety
Clinical-grade AI requires robust safeguards. The AI co-clinician employs a dual-agent architecture with a "Planner" to monitor and ensure the "Talker" agent remains within safe clinical boundaries.
The system also prioritizes clinical-grade evidence, performing verification and citation checking for retrieved information.
