On the way to solve the hardest problem with LLMs - Medical AI
Medical AI provides mind-blowing opportunities. In the US, 4.9T USD is spent annually on healthcare. Creating new products in this domain will make healthcare more accessible to millions if not billions. But how can LLMs help with that?
Google created Med-PaLM 2, which focuses on documentation and question-answering tasks for medical professionals. Gradient has chosen almost the same approach with the release of its HIPAA-compliant AI Platform for Healthcare. Woebot shared its experiments with LLM but highlighted that WoeBot doesn't use LLM to generate responses for the end user.
These examples show that companies play safe, and regulatory approval is the main reason. LLM chatbot will require regulatory approval as a medical device. Near an infinite number of inputs and outputs, low transparency of foundational models, proven clinical effectiveness, hallucinations, and prevention of harmful responses are just a few problems to solve. Bringing Medical AI will cost the same amount of money as bringing a new frug to the market.
Causal machine learning can answer many of these problems. In health, Causal ML helps estimate treatment effects and answer 'what if' questions. What happens if a patient quits smoking? The goal here is to infer causal relationships between variables rather than merely identify correlations.
Two things make LLMs ideal for Casual ML. The first is instant access to domain knowledge, and the second is extracting key primitives such as necessity, sufficiency, normality, etc. The CARE-CA framework uses these abilities to enhance causal reasoning.
Here is an example - "My body cast a shadow over the grass.". The first hypothesis is "The sun was rising, " and the second is "The grass was cut.". Which hypothesis is correct? To answer this question, CARE-CA enriches the premise with ConceptNet(an open, multilingual knowledge graph). Then CARE-CA generates 'what-if' scenarios. Having a deep understanding of concepts(first step) and 'what-if' scenarios, CARE-CA is ready to choose the correct hypothesis.
In the example above, we can use any available LLM on the market. However, there is an alternative - Teaching Transformers Causal Reasoning through Axiomatic Training. The paper shows how to train a new model specifically for Causal Reasoning. The results are impressive—a 67M parameter model works at the same level as GPT-4. The paper's key finding is that the model trained on linear causal chains can generalize to new scenarios.
Resources:
https://arxiv.org/abs/2407.07612 - Teaching Transformers Causal Reasoning through Axiomatic Training
https://www.sequoiacap.com/article/generative-ai-for-healthcare-perspective/ - Bringing Generative AI to Healthcare
https://www.nature.com/articles/s41591-024-02902-1.epdf - Causal machine learning for predicting treatment outcomes
https://woebothealth.com/why-science-in-the-loop-is-the-next-breakthrough-for-mental-health/ - Why Science in the Loop is the Next Breakthrough for Mental Health
https://www.nature.com/articles/s41591-023-02412-6.epdf - Large language model AI chatbots require approval as medical devices
https://www.cnbc.com/2023/10/09/google-announces-new-generative-ai-search-capabilities-for-doctors-.html - Google announces new generative AI search capabilities for doctor
https://gradient.ai/blog/gradient-releases-hipaa-compliant-ai-platform-for-healthcare - Gradient Releases HIPAA-Compliant AI Platform for Healthcare
https://www.nature.com/articles/d41586-023-03803-y - Generative AI could revolutionize health care
https://a16z.com/commercializing-ai-in-healthcare-the-jobs-to-be-done/ - Commercializing AI in Healthcare: The Jobs to be Done
https://www.wsj.com/tech/ai/medical-ai-tools-can-make-dangerous-mistakes-can-the-government-help-prevent-them-b7cd8b35 - Medical AI Tools Can Make Dangerous Mistakes. Can the Government Help Prevent Them?
https://a16z.com/the-next-webmd-an-llm-as-your-front-door-to-healthcare/ - The Next WebMD: An LLM as Your Front Door to Healthcare
https://arxiv.org/abs/2305.00050 - Causal Reasoning and Large Language Models: Opening a New Frontier for Causality
https://arxiv.org/abs/2402.18139 - Cause and Effect: Can Large Language Models Truly Understand Causality?
https://arxiv.org/abs/2403.09606 - Large Language Models and Causal Inference in Collaboration: A Comprehensive Survey
https://quarter-on-causality.github.io/tools/ - Introduction to causal inference, Elise Dumas
