Generative AI trends in healthcare - Moving from Hype cycle to Implementation cycle

A group of Google scientists released a groundbreaking report that showed that a Generative AI based medical solution gave better diagnostic accuracy and superior performance on multiple axes of clinical and consultation quality, which was rated from the perspective of specialist physicians as well as patient actors. AMIE (Articulate Medical Intelligence Explorer) is a research AI system developed by Google and is based on a LLM and optimised for diagnostic reasoning and conversations. In addition to strong standalone performance, clinicians assisted by AMIE system led to significant improvements in diagnostic accuracy of the clinicians in solving complex clinical cases.

Source – Google AMIE research report (January 2024)

Last year has seen many such studies highlighting the capabilities of Generative AI in healthcare. As 2024 unfolds, the domain of Generative AI (GenAI) in healthcare will see a transition from a phase of experimental hype to strategic implementation. Here are some trends we believe will change the way Generative AI applications are built for healthcare: 

1. New AI regulations that also stress the importance of explainability in AI

AI & ML algorithms have always had elements of a “black box” - technically complex, opaque, and uncertain on how they reach to a specific decision or prediction. The use of Generative AI in healthcare compounds this opacity, often leaving users (patients or doctors) wondering how it got its output and whether that output is reliable and accurate. Hence, a lot of research has already started on Explainable AI (XAI). XAI are a set of techniques and processes that give clarity on how AI models produce outputs, making them more understandable to humans.
Explainability aids in identifying flaws and biases in AI models, making it easier for improving the accuracy and fairness of the models and increasing the trust and adoption of Generative AI for industries like healthcare, where model outputs can have significant impact on individual’s lives.

With the European Union setting a precedent with its AI Act and the ICMR releasing ethical guidelines for AI last year; 2024 will see many countries coming up with stronger AI policies and regulations. Some themes that emerged from the EU AI Act include stricter regulations for high-risk AI, fostering transparency and trust in AI by setting out clear rules and regulatory sandboxes for innovating in high-risk AI applications, most of them pointing to the need of Explainable AI (XAI). And ICMR’s guidelines are more around “Do no harm for the patient”. 

Healthcare companies using Generative AI will have to focus on conducting comprehensive testing, documentation of data and set up accountability frameworks and LLM guard railings including human in the loop for high-risk use-cases.

2. Improving economics from reduced prices of Large Models and shift towards smaller, specialized models

We will very soon see the economics of LLMs getting better, making them more accessible and cost-effective, promising wider adoption in healthcare. There is aggressive innovation in software (efficient model architectures, sparse models, training optimization) as well as hardware (distinct chips for training and inferencing LLMs). All this innovation is bound to decrease the cost of training LLMs (infrastructure) as well as the cost of inferencing LLMs (prediction) significantly. In January 2024, OpenAI announced a major drop in their API pricing, reducing prices by 25-50%. We are likely to see similar announcements from other big tech players throughout the year.  

We have also seen emergence of Small Language Models (SLMs) which are significantly more computationally efficient than LLMs because of the number of parameters and domain-specificity. Hence, SLMs require less storage, are faster to infer and cheaper to deploy and run. Phi2 by Microsoft and Gemini Nano by Google are SLMs that have been launched recently. We expect that specialised disease specific SLMs focused on specific healthcare use cases will be launched in 2024.

We believe that it’s a matter of time that language models will be commoditized. And healthcare startups using Generative AI will have to focus on creating unique moats such as additional use-case specific data moat (proprietary data), technology moat (middleware on top of the foundational models) or commercialization/go-to-market moat (revenue generating or cost saving workflows).

3. Gold rush for AI Doctors and AI co-pilots

Fuelled by advancements in Generative AI, we will experience a resurgence of healthcare chatbots, which now can avoid the pitfalls of their predecessors by having a vast knowledge base from the start instead of building it from scratch, thanks to access to Large Language Models. Utilising the advanced capabilities of LLMs, AI co-pilots can address healthcare’s linguistic challenges while offering expertise in medical diagnosis with empathy. Generative AI based systems have shown clinical coverage and medical accuracy almost reaching 90%. Large language models have already shown excellent clinical capabilities on various benchmarks like MedQA (USMLE) and MedMCQA (Indian medical exams) and MMLU (Massive multitask language understanding). 

This progression towards AI-driven healthcare holds the promise to improve clinical performance and experience, as well as pave the path to enhance patient adherence and establish self-service habits. 

In conclusion, Generative AI is not only an inflection point in Artificial Intelligence but can prove to be one in healthcare as well. With Big Tech focusing on healthcare, and the speed of innovation in the industry, we expect to see meaningful use of AI in healthcare over the next few years. However, regulations, explainability and human oversight will be crucial factors for successful Generative AI implementation in healthcare.

By Pankaj Jethwani, Nikhil Hegde, Ashish Sharma – 2070 Health