Part 1: The 3P's of AI in Healthcare

Building in the Productivity Space

Image generated using Open AI's DALL·E 3.

Despite a growing set of challenges, the primary goal of healthcare systems around the world remains unaltered - to provide superior patient care. In recent years, we’ve seen the opportunity to apply Artificial Intelligence (AI) expand exponentially, not only to optimise cost and alleviate workforce pressures, but also to provide superior methods of patient care. 

This article represents the first in a series of 3 articles on AI in healthcare. We believe that AI has the potential to revolutionise healthcare, with a myriad of solutions ranging from automation to preventive and personalised medicine. It’s no secret that building successful companies in healthcare is laden with challenges from regulation to interoperability. Join us as we breakdown and analyse the landscape through the lens of 3P's: productivity, preventative medicine and personalised medicine. These represent what we see as the greatest and most expensive challenges facing healthcare today that AI can tackle. 

Breaking down the productivity problem

Not a day goes by that we don’t read about the massive inefficiencies and rising costs in healthcare. Spending is growing at a rate greater than 5% in both the US and the UK. In 2019, the US spent 3.8 trillion or 18.4% of GDP on healthcare alone, and is predicted to spend as much as 12 trillion by 2040. Much of this, however, goes to waste, with a whopping 25% written off due to administrative complexities, duplicative services, unnecessary treatments, high drug prices, and hospital readmissions. With pressure to reduce costs, a greater burden is put on clinical staff resulting in lower retention rates and worse working conditions. Just this year, the UK has seen the largest NHS strike action on record and a shortfall of 43,000+ nurses. 

How might we make healthcare workers and processes more productive using AI? And where are the main opportunities? 

Administration and financials. Administrative spending, which relates to the non-clinical costs of running a medical system, was estimated to contribute 9.8% of healthcare spending in the USA and 1.9% in the UK in 2022. The companies identified in this space relate to insurance-related expenses (Co:helm, Slingshot, Latent, Phare Health, Alaffia Health, Regard, Cair Health), automated patient billing (Pledge Health) and patient registrations (Healthtech-1). Many of these companies leverage Large Language Models (LLMs) out of the box as they are particularly well suited for generating and filling expense and registration forms.

Clinical note taking and planning. In 2016, a study estimated that doctors spend 67% of their time doing paperwork. Automation of paperwork can enable more productive clinical staff and reduce costs. Advances in speech signal processing and LLMs are enabling automated note taking for clinicians, whereby clinical notes providing a synopsis of a patient’s admission or care can be transcribed and coded into electronic health record fields appropriately. The companies identified in this space relate to digitalisation of conversations (Knowtex, Eleos, Ambience, DeepScribe, Nabla) and providing an all-in-one platform for note taking and insights (Glass, Abridge).

Medical understanding and search. Current advances in LLMs are aiding the interpretation of large amounts of unstructured data in medicine. The companies identified here are offering solutions for intelligent summarisation of notes and relevant patient medical histories (Science.io), for reviewing emerging evidence for advanced medical procedures (OpenEvidence, i.e. XYLA), reviewing medical regulation for compliance (Medwise.ai), and consumer medical search tools (MediSearch). The challenge using LLMs out of the box for this application is that they could provide misleading non-factual information, i.e. they hallucinate. These companies bypass this limitation by using LLMs to interpret queries and then retrieving factual information from an existing database (e.g. PubMed) to supplement responses.

Digital care and telemedicine. Solutions in digital care and telemedicine offer digital (rather than in-person) access to medical resources for the public, thereby alleviating resource and productivity pressure on clinical staff. Start-ups identified in this space are offering solutions to digitalise patient-doctor interactions (Videra), manage personal health (Syndi), and improve the mental health of clinical staff who burn-out due to high workloads and staff shortages (Vitalizecare, Maida). It's interesting to consider the future of the digital care and telemedicine space as LLMs that excel at medical understanding become multimodal, and the current interface with the technology evolves beyond text prompts.

Early stage start-ups operating in the healthcare productivity space. Our list is not exhaustive.

Finally, some companies, like Anima Health, are building across multiple healthcare productivity challenges. For example, Anima’s platform combines automated clinical documentation, streamlined admin and financials with their LLM, Annie, which enables quick access to patient data and care suggestions.

What are some of the key challenges?

Accessing data. As the old saying in AI goes: garbage in, garbage out. To ensure a high quality AI product, high quality and relevant data is needed in training. Whilst LLMs are capable of great performance out of the box, it may seem that additional training data isn't required. But LLMs are a double-edged sword, they are also an extremely accessible technology. Companies need to find a competitive edge for their application. Building a data moat could be one way to do this. For example, DeepScribe has an exclusive dataset containing over 2 million patient encounters which they have used to fine-tune their LLM note taking product - making it difficult to replicate their model and performance. Critical for developing a high quality data moat is to build long-term trust-based partnerships with healthcare providers.

Interoperability. Healthcare providers often have established systems in place, and these systems may vary between hospitals. Integrating new productivity tools with legacy systems can be challenging on its own, but building a product that can seamlessly generalise between different systems presents its own challenge. To combat this, some of the companies we've listed have developed API product versions, for example, Nabla that offer an API for their note taking co-pilot. Whilst other companies, such as Co:Helm, are tackling the problem head-on with their 'interoperability engine'. These strategies decrease the need to train (already very busy) staff on how to use your product, and instead focus on integrating with systems the staff are already trained to use, therefore aiding adoption and growth.

Regulation. In the productivity space it can be difficult to tell at first if you're building a product that could be classified as a medical device. In general, any technology that provides information regarding a medical condition will be classified as a medical device. This increased regulation will change your company strategy from the outset, as it could take millions in capital and many years to build the evidence for compliance. Check-out this post by the FDA on what software functions classify as a medical device, and this post on what it may take to get medical device approval for a LLM.

Making healthcare productive at London's Generative AI x Healthcare Hackathon

Recently we hosted the inaugural Generative AI x Healthcare Hackathon at Imperial College London, which attracted 70+ participants from industry, academia and the clinic. Many of the teams built productivity solutions aimed at improving patient outcomes and healthcare administration. To give some insight into innovative and emerging products in this space, we highlight the work of the winning teams.

• Paco built a patient companion app to support patients throughout their clinical journey. Patients are kept informed of what’s required of them and can clarify any remaining questions - all whilst patient engagement and understanding can be monitored from the clinical side and flagged for contact. Their solution aims to reduce negative preventable outcomes such as missed appointments and poor post-op outcomes. They won our first place prize: £2000 in cash, £2500 in OpenAI credits, and meetings with Hummingbird and Microsoft for Startups. Congratulations to the Paco team, Martin van der Heijden, Vinayak Athavale, Dr. Paul Jewell, David Ensor, Dr Gani Nuredini!

• LabGPT took home our second place prize by building a chat-based molecular biology assistant for experimental design, speeding up the design process from 10-20 minutes to seconds. Congrats to the LabGPT team, Timon Schneider, Maxime Theisen, Le He!

• Raytwelve took our third place prize, by presenting an API to tackle misinformation in healthcare by comparing input text prompts to an approved healthcare knowledge graph to obtain a score that quantifies the trustworthiness of a statement. Congrats to the Raytwelve team, Tina Gogna, Jonathan Rubin, Tiantian He, Kien Hang, Elle Yang, Seth Howes!

The team behind Paco that won our first place prize at the hackathon. From left to right, Dr Gani Nuredini, Dr. Paul Jewell, Martin van der Heijden, David Ensor and Vinayak Athavale.

Conclusion

As healthcare embraces the productivity problem, AI stands out as a significant tool to help; particularly LLMs. With its capacity to streamline tasks and enhance clinical practices, companies in this space are already having huge clinical impact. Accessibility of the underlying technology has opened up a huge number of opportunities, but we believe those that focus on building data moats, seamless product integration, and successfully navigating regulation, will emerge on top. Stay tuned for our next article, where we delve into AI for preventative medicine; the second of our 3P's.

At Hummingbird, we’ve had the privilege to back entrepreneurs across four continents challenging Eroom’s Law such as BillionToOne, Enveda Biosciences, Automata, Basecamp Research, Pristyn Care, Eden and Anima Health, among others. We’re incredibly excited about the next decade of change in healthcare and lifesciences, so if you’re building in this space please get in touch!