Training AI for Health Care: Munjal Shah’s Mission to Do No Harm

Munjal Shah, founder of startup Hippocratic AI, wants to leverage large language models (LLMs) to provide crucial but nondiagnostic healthcare services without the risk of AI overreach. His vision is to create an LLM specially trained in medical content to communicate health information accurately and empathetically at scale.

Just one year after OpenAI unveiled the revolutionary chatbot ChatGPT, the rapid advancements of LLMs have entrepreneurs like Shah considering new applications. With healthcare worker shortages widening the gap between patient demand and provider capacity globally, Shah sees an opportunity for AI to fill the breach responsibly.

Hippocratic AI’s name reflects Shah’s commitment to ‘first, not harm’ in applying AI. Rather than making diagnoses or treatment decisions, the LLM will focus on chronic care coordination, patient navigation, and nutrition—reducing provider burnout to concentrate on higher-risk judgment calls. Shah believes LLMs are uniquely suited to absorb research and convey it conversationally, with potentially more empathy and patience than overburdened human clinicians.

A recent JAMA study supports this idea, finding ChatGPT’s patient counseling responses were preferred to doctors 79% of the time, with higher marks for quality and empathy. While cautious on conclusions, researchers and innovators like Shah see communication-centric use cases where AI could augment overwhelmed healthcare teams.

For any clinical application, accuracy is paramount. Shah stresses Hippocratic AI’s LLMs are explicitly trained on evidence-based medical content rather than the broader internet sources of mass-market chatbots. This specialization better equips them to understand and advise patients on care basics accurately. The training also includes reinforcement learning with clinician feedback to improve responses continually.

So far, Hippocratic AI has outscored GPT-4 on 114 medical exams and bedside manner benchmarks. Shah sees responsible AI integration as inevitable and aims to lead on training tools focused entirely on patient education and support. By lightening human health workers’ communication and coordination burdens, the goal is to free their time for judgment calls only highly trained professionals can make.

Comments are closed.