By Venkat Srinivasan (Managing Director), Mark Legare (Principal), and Yuri Ahuja (TIR)
Introduction
The pandemic laid bare the inefficiencies in health care delivery and inequities in access to quality health care. Inefficiencies reflect antiquated systems and lack of integrated health data compounded by privacy concerns, while inequities speak to inconsistent access to qualified medical professionals due to cost or availability.
On the supply side, there is a shortage of qualified healthcare workers globally. Hearing from experts in continuing medical education, there are also significant gaps in knowledge even among qualified U.S. clinicians in different geographies, e.g., Boston vs rural New Mexico. The knowledge gap across healthcare workers is exacerbated by the constant flow of new research findings further amplified by the widespread usage of pre-print archives, e.g., arxiv.org.
There is also increasing evidence of significant burnout among healthcare workers nationwide. Among U.S. healthcare professionals, over one-half of physicians and one-third of nurses are reported to be experiencing burnout symptoms(1). Beyond poorer clinical outcomes, burnout also has significant economic costs. On a national scale, it is estimated that approximately $4.6 billion in costs related to physician turnover and reduced clinical hours is attributable to burnout each year in the United States(2).
Shockingly, the 3rd leading cause of death in the U.S. is medical errors. A Johns Hopkins study found that across all care settings including hospital and clinic-based care, an estimated 795,000 people die or are permanently disabled from diagnostic errors(3).
In the U.S., there are also exogeneous factors which greatly affect the success of any efforts at improving affordable health care. The sector has stakesholders with potentially conflicting incentives – pharma companies are interested in maximizing sales of their drugs, doctors are influenced by reimbursement and malpractice insurance, and insurers are interested in minimizing their expense ratios.
The Immutable Care Triangle
The immutable care triangle reflects the notion that access, cost, and quality are co-dependent and in the current environment, no one side can be improved without affecting the others. That’s why it is immutable. As long as the care triangle remains immutable, it is going to be difficult, perhaps impossible, to achieve the goal of affordable high quality care for all.
Inadequate health insurance coverage is one of the largest barriers to health care access(3) and the unequal distribution of coverage contributes to disparities in health. At the current rate of annual increases, health insurance premiums in the U.S. are simply not sustainable. Out-of-pocket medical care costs may lead individuals to delay or forgo needed care (such as doctor visits, dental care, and medications), and medical debt is common among both insured and uninsured individuals. People with lower incomes are often uninsured, and minority groups account for over half of the uninsured population.
The Immutable Care Triangle
Can AI help solve these challenges? On paper, AI has the power to change this, enabling the provision of higher quality care at lower cost. In fact, only AI can help us change the immutable characteristic of the access-quality-cost co-dependency.
We believe healthcare represents a very large opportunity area for AI and expect a great deal of innovation and startup activity to that end. We agree with Microsoft, Apple and Google, who have all made public statements how healthcare may well be the most pressing and critical area of application for AI. How much of this we will realize in the short vs long run is a function of how trustworthy and reliable these AI applications are and how the regulatory system handles such innovations. In practice, these are still very early days.
For mission critical use cases, AI will only provide significant value if we can solve some well known challenges – explainability, generalizability, and tractability. For other use cases, where the cost of errors is less significant, black box AI is likely to gain faster adoption.
Clinical decision support could finally come of age with AI
AI has the potential to completely upend current clinical delivery models. Clinical decision support systems [CDSS] have been around at least since 1984. They were symbolic rules-based systems and didn’t have the benefit of empirical knowledge across a large number of experts/data. Will they be adopted now? As long as the CDSS remain black boxes, we are skeptical they will get adopted. CDSS can provide transparency and explainability, we believe will unlock adoption pathways.
In the next generation of CDSS, Foundation Models [FMs] will play a key role. In an earlier article titled, ‘The AI Renaissance’, we had forecast the emergence of a new industry centered around ‘computational bundles of knowledge(4). Foundation models are exactly that. At the heart of ‘Generative AI’ are Foundation Models [FMs] trained on widely available data reflecting a broad pre-curated knowledge base. On top of these FMs, startups that add/create proprietary data or knowledge can develop valuable proprietary models with domain-specific expertise.
We can think of two kinds of FMs – unimodal and multimodal. Multi-modal FMs will integrate health data across multiple modailities – EMRs, images, research, and others. We believe it is likely we will see unimodal FMs become successful first and evolve into successful multimodal FMs over time.
While FMs hold a lot of promise to reduce the knowledge gap and to prevent misdiagnosis, they will only be adopted if they solve the challenges identified earlier, particularly explainability. While there is a rush to adopt the current crop of black box FMs, in mission critical use cases, the degree of adoption and value will highly depend on how well the challenges listed earlier have been addressed.
Precision Medicine and Real Time Diagnostics will gain traction
Despite the age old wisdom, ‘Prevention is better than cure’, historically diagnostics startups have not been the most sought after venture investments. We believe this will change with precision AI models. CMS has recently introduced reimbursement codes for chronic condition management focused squarely on prevention and disease progression. Finally, we are seeing a shift towards prevention.
The objective of precision medicine is to use individual biology rather than population biology at all stages of a patient’s medical journey(6). Precision AI models will be able to provide clinicians with personalized recommendations, e.g. which of the available drugs might be most effective on a specific patient or predict the occurrence of a certain clinical presentation. We believe real time diagnostics and feedback are game changers for managing disease progression, especially for chronic conditions like diabetes. We have several portfolio companies that have innovated real time diagnostic models and devices.
We believe integrated closed-loop drug delivery systems with on-demand therapeutic capabilities will gain traction. Tissue-interfacing sensors and robotic therapeutics already enable the real-time monitoring and treatment of diabetes, wound healing, and other critical conditions. These systems will provide timely and personalized treatments that correspond to chemical, electrical, and physical signals of a target morbidity(7).
We have seen a lot of activity in remote patient monitoring and intelligent patient engagement through conversational AI agents. We expect that such agents will be anchored on models of disease progression. While these are early days, we expect the focus to be on the demonstrated clinical utility of such models – e.g., a reduction in the number of ER visits, a noticeable arrest of disease progression, among others.
We also believe AI will play a greater role in making robotic surgery and clinical procedures routine. This should reduce the current error rates in manual surgical procedures.
Automating Administrative processes can be the low hanging fruit
Administrative processes in healthcare represent a relatively less risky set of opportunities for AI. They are less mission critical than clinical care. There are numerous processes that are inefficient and can be automated with intelligent technology. Many physicians complain they spend an inordinate amount of time on non-clinical activities. One study estimates that on average a nurse in the U.S. spends 25% of work time on administrative activities including complying with regulatory requirements.
An area where we are seeing a lot of startup activity is conversational interaction with patients. The extent to which these NLP based conversational agents can be helpful in clinical versus non-clinical interactions will depend whether the applications can solve the AI challenges listed earlier.
Pre-authorization, claims processing and billing verification are areas which remain inefficient and can benefit from AI. There have been many attempts at automating claims processing in the past. There are two aspects to these automating processes – efficiency and fraud prevention. Fraud remains a significant burden on the healthcare system.
The Human in the Loop
While we see enormous possibilities for AI in healthcare, we do not foresee wholesale direct replacement of clinicians or healthcare workers as a result of AI. We do, however, envision significant reduction in human effort in all areas of healthcare and opportunities for improved clinical care. Reduction in human effort will certainly have the effect of reducing the shortage of healthcare workers. We believe the use of AI will give rise to a differently skilled healthcare worker - one who is skilled in the use of such AI models. And, above all of else, we are optimistic that one of the greatest benefits of AI-driven care will accrue to the physician-patient relationship, where automation of certain tasks can enable more time spent on the human elements of care delivery.
A Higher Level View
We view the healthcare AI startup universe along two dimensions – those that improve operational efficiency and those that improve clinical outcomes. An example of the first is an AI model to automatically identify and apply the appropriate billing codes for a procedure. Applications of AI to extract and normalize data from unstructured documents represent a common example of an application to improve operational efficiency. An example of the second dimension is an AI model that interprets a patient’s medical history and relevant new data to provide a summary, diagnosis, and putative plan to the physician.
AI Startups in Healthcare
The 2x2 above is an attempt to characterize startups across the two dimenions from an investor lens. Startups with low levels of expected impact on clinical outcome and/or operational efficiency are obviously not likely to be venture scale opportunities. Startups that are expected to have a high level of clinical impact will generally carry a higher level of R&D risk and need longer adoption cycles. We also believe ‘explainability’ will be crtical for these use cases. On the other hand, startups focusing on operational efficiency will carry much lower R&D risk and face easier adopton challenges but offer relatively more modest returns. The rare startup that can impact both operational efficiencies and improve clinical outcomes will of course have the combined characteristics of the two dimensions.
We believe an appropriate portfolio strategy might require a careful allocation of capital across the three quadrants. Further optimization will be required based on specific investment mandates.
Summary
At Innospark, we believe AI opportunities in healthcare are enormous but require careful navigation. It is easy to get carried away with the hype around LLMs. We find it is critically important to understand the context around which AI is proposed to be used in order to develop conviction around adoption and adequacy. At this time, we expect CDSSs will be the hardest to scale in terms of adoption and adequacy. Until the fundamental challenges with black box LLMs get resolved, their adoption in mission critical applications is likely to be limited. However, many entrepreneurs fully understand these challenges and are working around them. In the aggregate, we are bullish that the next decade and more will see the transformative impact of AI in healthcare. The care triangle will become mutable and we might well see significant progress towards quality healthcare for all.
References
(1)Reith T P (December 04, 2018) Burnout in United States Healthcare Professionals: A Narrative Review. Cureus 10(12): e3681. DOI 10.7759/cureus.3681
(2)Shasha, H, Shanafelt, T.D. et al (2019). Estimating the Attributable Cost of Physician Burnout in the United States, Annals of Internal Medicine (2019), https://doi.org/10.7326/M18-1422, June 4, 2019.
(3)Newman-Toker DE, Nassery N, Schaffer AC, et al. Burden of serious harms from diagnostic error in the USA. BMJ Quality & Safety Published Online First: 17 July 2023. doi: 10.1136/bmjqs-2021-014130
(5)Bohr A, Memarzadeh K. The rise of artificial intelligence in healthcare applications. Artificial Intelligence in Healthcare. 2020:25–60. doi: 10.1016/B978-0-12-818438-7.00002-2. Epub 2020 Jun 26. PMCID: PMC7325854.
(6)The AI Renaissance, Innospark, https://medium.com/innospark-ventures/the-ai-renaissance-8bf48823c782, Nov. 2022.
(7)Yuri Ahuja, Generative Large Language Models in Healthcare, Innospark, Internal Research, 2023.
(8)Jeff Knox, Healthcare Patient Data, Innospark, Internal Research, 2022.