Today's deep dive: AI Clinical Summarization tools 🕵️♀️📚 One of the areas gen AI has immense potential in healthcare is its ability to surface and synthesize information from vast sets of unstructured documents and data. As Oscar Health cofounder Mario Schlosser puts it: “[LLMs] are uniquely capable at going from unstructured data into structured data, and the other direction.” So it’s no surprise that distilling information from a wide range of clinical documents and data is one of the most common applications of AI that we’re seeing come to market. There are MANY use cases for summarization across nearly every dimension of healthcare. To name a few: ↳ Pre-charting ↳ Referral summaries ↳ Discharge summaries ↳ Diagnosis and care gap identification ↳ Quality measurement and improvement ↳ Real world evidence curation ↳ Clinical trial matching ↳ Clinical registry submissions ↳ Clinical documentation improvement (CDI) ↳ Prior authorizations ↳ Legal and insurance use cases – e.g., workers comp, life insurance Despite their broad applicability, uptake of these solutions has been relatively slow, at least in comparison to ambient scribing solutions. A big reason is their perceived risk. With summarization, the AI is making a subjective call about what to include and what to leave out. There's also the potential it hallucinates something not present in the underlying information set. In a clinical context, both are major concerns. Still, between advancements in gen AI capabilities and improvements in risk mitigation techniques, we're seeing more and more organizations begin to evaluate and adopt summarization products, starting with lower risk use cases. Will be interesting to watch this space in the coming months to see how quickly the pace picks up, particularly for clinical use cases. Curious to hear what are others seeing. --- P.S. Here's our working list of AI Clinical Summarization solutions • Abstractive Health • Carta Healthcare • Credo • DigitalOwl • Layer Health • emtelligent • Fourier Health • Google Care Studio • Hona (YC W24) • Inference Analytics - Workforce Concierge • meMR Health • Mendel.ai • Navina • Oler Health • Pieces • Quench • Regard • Solstice • Wisedocs • Synthpop - Healthcare AI #healthcareai #genai #healthai #digitalhealth #healthtech
Really good call outs here. Healthcare is a different animimal. Everything we do is about safety and efficacy as we navigate the HHS, FDA, HIPAA, GDPR and all of the other governing bodies, guidelines and regulations. It’s slow to adopt because no one wants to be first and no one wants to be last. Being risk adverse is typical. Proven case studies and standards need to emerge. It is certainly an exciting time to see how AI and tech will reshape healthcare.
Thanks for including us (emtelligent) as well, Bobby Guelich! You are absolutely correct that there is concern with the accuracy and tendency for hallucinations. We believe emtelliPro+ has solved for this by leveraging our highly accurate data extraction product, emtelliPro. What I'm concerned with is solution fatigue. There is so much energy in trying out the different AI tools currently, but will there eventually be too much exhaustion if the accuracy problem isn't solved with the first few tools in the evaluation process?
A great way to mitigate this risk is by keeping a human in the loop! Even when a “human-in-the-loop” is still necessary, a drastic reduction in handle time can quickly make impact on workloads. Less time spent on notes —> [fill in the blank] - More time on patient care - More time spent with family - Ability to scale more efficiently
Having a strong clinical team for quality control and improvement initiatives is a great way to reduce to impact hallucinations have and will ease the minds of potential stakeholders as well as end users which have clinical licenses to think about as well!
Adjunct Professor of Medicine, UCSF | Electrophysiologist | Inventor | Entrepreneur | Generative AI for Healthcare
1yThis is great, Bobby. Thanks for the shout-out! You're performing an important role for the digital health ecosystem, which can be intimidating for folks who need actual solutions to real problems. At Quench, we serve physicians, nurses, attorneys, and insurance companies who need to analyze and reason over medical records. These records often arrive as 1000's of pages of poorly organized PDFs, including handwriting, tables, duplicates, blank pages, etc. Instead of relying on a single LLM, we approach these challenges with what Zaharia et al. recently described as "compound AI systems." By using multiple models (large and small), NLP, RAG, agents, etc. we can rapidly and accurately organize, summarize, and reason over a vast volume of content. Furthermore, by strictly limiting the context to the provided records, we essentially eliminate hallucinations. Solutions that simply wrap a front end on top of a proprietary LLM like GPT4 make for quick demos but rarely breakthrough to commercial success at scale because of inaccuracy, slow and expensive inference, and hallucinations.