Healthcare Startups & Gen AI: the Stakes are High

Thanks Steven Krein for the intro and Unity Stoakes for moderating this very interesting panel on conversational AI accelerating health moonshots, featured in the StartUp Health newsletter this week! How Conversational AI and Chatbots Are Accelerating Health Moonshots (campaign-archive.com)

A couple of thoughts that I took away from the discussion, encouraged by both common & dissimilar opinions across the panelists:

(1a) Everybody's been applying & training AI in for their own specific use cases, creating a variety of useful point solutions, similar to how medical imaging diagnostics with traditional AI models have been very specialized to e.g. one type of #cancer per model. Old-school neural networks have been prone to problems with sparse data (https://en.wikipedia.org/wiki/Sparse_matrix) and catastrophic forgetting (https://en.wikipedia.org/wiki/Catastrophic_interference), making these highly specialized models necessary for good performance. However, the characteristics of the transformer architecture (https://arxiv.org/abs/1706.03762) behind large languages models like #ChatGPT should allow much more holistic healthcare AI going forward, with a single monolithic model able to learn from e.g. a variety of cancer image & diagnoses data. Very promising for pooling use cases and data together to take on new approaches such as Marina Simian , who I met yesterday at #SFTechWeek, building Oncoliq to detect all forms of cancer from a single blood draw. Or, an AI care plan recommender that works both for older adults and for children with developmental disabilities? When done right, a transformer model that generates effective care plans for geriatric patients will tend to do it better, not worse, when also trained to produce effective care plans for pediatric patients. It's a bright future for generalizable solutions, mitigating silos, and delivering on the promise of whole-person, whole-lifespan care.

(1b) This didn't come up in the panel, but the corollary to (1a) is that there is likely to be a lot of merger & acquisition activity coming up, as larger players with "Layer 2" models (see https://www.nfx.com/post/generative-ai-tech-5-layers -- thanks NFX ) in particular healthcare domains are able to not only view acquisitions as accretive from a typical M&A perspective in increasing sales growth/value for the combined company, but literally accretive to their flagship Layer 2 models. A traditional software company acquisition would result in lines of code from two product platforms remaining substantially separate lines of code, with possibly some changes and integration code added, and some cross-selling of products. A Layer 2 AI startup that acquires another related Layer 2 AI startup could pool their data and model training resources such as reinforcement learning with human feedback cycles into a single Layer 2 model that is from a strictly technical standpoint better than the sum of its parts. Due to this technical accretive potential of transformer-based startups, it's a particularly exciting time to be in this space, whether as an investor or operator.

(2) There's a tension between innovation in tech and safety in healthcare. Those of us with the means to apply technology, esp. rapidly evolving #generativeAI approaches, in a safe and effective way in healthcare (e.g. here at care.coach we have healthcare conversational safety/situation classifiers with AUROC >0.9 and the ability to escalate to our 24x7 team of human health advocates) have a responsibility to do so and demonstrate at-scale, irrefutable clinical and financial returns quickly, before one of the many newcomers apply this tech to healthcare screws something up and causes chill, not unlike how a certain fingerstick blood testing company did for that whole technology category. These are genuinely life-changing technologies that need to be developed and applied to improve the health & wellbeing of millions or billions of people, but all it takes is a single screw up to put our healthcare industry years further behind in tech adoption. The difference with other types of medical technology is that generative AI is evolving so quickly, a few years of chilling effect will translate to an almost unimaginable degree of forgone benefits to society. We need to get it right, at scale, before someone gets it wrong.

In case you're wondering what exactly care.coach has been doing in this field over the past few years, here's a summary:

Feel free to get in touch if you're interested in improving the health & wellbeing of millions of people together, or just learning more :)

Special thanks to StartUp Health for sharing our journey and National Institute on Aging (NIA) for supporting our latest technology development, funded in large part under awards SB1AG082634 & R44AG062014.

cc Renalis Health Mediktor IndyGeneUS Bio AgeTech Collaborative? from AARP Rick Robinson Amelia Hay Sasha Spellman Sheila D. Collins, MBA CalPACE National Institute on Aging (NIA) National Cancer Institute (NCI) Nex Cubed Blueprint Health LLC Logan Plaster Holly Zoeller Nicole Kinsey Microsoft for Startups Sally Ann Frank Scott Peifer Paul Sonnier Nate Dalton Jennie Chin Hansen Ana Gupte

#genAI #LLM #digitalhealth #seniorcare #medicare #medicaid #PACE