Healthcare and AI hopes and expectations for 2024
Optimistic predictions from the ML for MDs community
2023 has been a whirlwind of new information and excitement, so what do we expect in 2024? The frontier of AI has been called jagged many times, and the concept of a uneven advances feels even more true in healthcare. I have patients who know the exact gene that caused their cancer, for example, but can’t get a ride to chemotherapy. Or who have surgery performed using a robot but have a drug reaction because we can’t predict those yet. I’m hopeful that AI can help even out some of the jagged edges of healthcare, even as it struggles through its own jagged frontier. Below is an optimist’s list of how we can build on AI advances to improve healthcare.
Note: The physicians at Machine Learning for MDs are some of the most knowledgeable experts at the intersection of healthcare and AI, and can speak from actual clinical and research experience about the technology. They were generous enough to add their thoughts below:
Actually getting this technology implemented
Healthcare adoption of new ideas has often lagged behind other industries. There’s an old study showing that 17 years pass between the development of a new idea and its adoption into common clinical practice. However, other studies note that the lag time is incredibly variable depending on the problem and how it’s measured. If the idea requires drug development and FDA approval (as was the case with the study citing 17 years) then it will take longer than products that are Software as a Service. The “last mile” problem in health tech is also frequently described.
In general, adoption of new ideas and technology take longer in healthcare due to the moral imperative to do no harm. Some authors cite slow technology adoption as being due to lack of financial incentives in healthcare and a slow managerial and bureaucratic process. This matches with what many healthcare founders from other industries tell me, which is that the sales cycle for hospitals is incredibly long, often 12-18 months. If we take that as a starting point, many generative AI companies weren’t even founded a year ago, so we can assume that many of them are about midway through the process. This fits with what I’m hearing from people: there’s interest but it has to make or save the hospital money (the financial incentive piece) and that founders are getting meetings but haven't closed deals.
The administrative tools will come much sooner than the clinical tools; there’s more focus on clinician burnout and retention, and I expect that offering ambient scribes and AI inbox management will be recruitment tools for clinicians in the near future. Physicians and nurses also do a lot of business tasks these days like Powerpoint presentations; using AI to make those parts of the job more efficient would be a huge win. The clinical tools will need more validation and monitoring, but at some point we’ll need to start asking ourselves if we’re doing more harm by not giving clinicians access to some of these tools.
Computer plus human studies
Matt Sakumoto, MD: I have been fascinated by the human vs computer and human + computer studies coming out of Adam Rodman’s group. These are mostly preprints but the question of Diagnostic Reasoning constructs that GPT4 mimics is very cool to look at.
Dr. Sakumoto is absolutely right: we need more studies to show us the best use cases of AI (and humans), and how the two interact. The studies he mentions are related to diagnostic reasoning, which can be a great use of AI if implemented correctly. Other studies looking at how often physicians override their own judgment or ignore correct AI suggestions are also fascinating. As one study of radiologists using AI diagnostic assistance noted, “average accuracy of human decisions drops from 78% when AI offers correct suggestions to 28% when AI offers incorrect suggestions (p-value: 2.2 e−16).” These studies help us understand not just the medicine and AI but also the human brain.
Increasing availability of clinical data repositories
Niraj Patel, DO: I’m hoping to see more clinical data repositories for researchers to train and test AI algorithms, such as the Medical Informatics Operating Room Vitals and Events Repository (MOVER): a public-access operating room database (no affiliation).
As Dr. Patel alludes to, getting access to real clinical data is hard due to institutional rules, HIPAA regulations, and very real concerns about re-identification of data with AI. In the future we may have better access to both synthetic data and real data for AI algorithms. Many of the current databases are desperately short on data from people of color which is especially important in fields like dermatology. Stanford recently released several large databases publicly, and hopefully other institutions and groups will follow.
Advances in LLMs interacting with each other
Hilary Lin, MD: I’m excited about advances in agents - with AutoGen and similar frameworks we can take LLM applications to a whole new level - dare I say, actually disrupt medical research and care entirely?
The concept of agents in LLMs is really exciting. Usually there are multiple LLMs, each with a specific knowledge set, and they can interact with each other to figure out the best answer or solution to a problem. They can also be given roles within the ecosystem so they “act like” something or someone in real life. At least in theory, they should be able to do things like fact-check each other and decrease the risk of single-point failure. The research on these is developing now and I agree with Dr. Lin that we’ll see more of these in the near future.
Faster and better drug development
Drug development has always been a numbers game to some extent: companies would make financial bets on a variety of drug candidates, knowing that about 10% of candidates would actually succeed. AI has sped up that process with advances like AlphaFold, which decreased time to identify a protein structure from its amino acid sequence from months/years to hours/days. Now tools like RFDiffusion can create images of protein structure and inverse folding tools can suggest amino acid sequences based on a desired target structure. It almost feels like science fiction compared to how hard it was to do these tasks when I was doing bench research as an undergrad.
Research on using AI for clinical trials and digital twinning has also accelerated and provided hope that AI may safely shorten the (very long) timeline of clinical trials. These digital twinning efforts are happening both at the cellular and whole person levels, but have yet to be approved by the FDA. Still, pharmaceutical companies may be able to focus on more promising drug candidates more quickly as this technology improves.
Scribes rolled out to more clinical specialties and more professions
One of my continued confusions about AI scribes and other AI tools that reduce AI scribes is that they’re usually limited to outpatient primary care physicians. There are many other people and places in the healthcare system that would benefit from this technology. Yes, primary care physicians spend a lot of time charting, but so do all outpatient physicians including specialists. Also, many of the ambient scribe platforms are browser-based, so it seems easy for inpatient physicians to bring an iPad into a patient’s room and have the ambient scribe ‘work’ in the background. Inpatient doctors spend a huge amount of time on documentation: more than four hours a day according to one study. If these ambient scribes work as well as I’ve heard from colleagues, let’s get them rolled out to physicians across specialties and settings.
And let’s not stop at physicians. Nurses and other healthcare professionals bear a huge amount of the documentation burden. Many surgical NPs and PAs do the large bulk of documentation with physicians signing off on notes. Nurses have a truly ridiculous documentation burden, and it’s imperative that we keep our nurses engaged and happy as we deal with the ongoing nursing shortage. Many of the screening questions nurses ask could be done by a text-based survey, for example, or entered into the chart by an EHR-integrated ambient scribe. It’s great that some primary care physicians have access to these tools. I’m just surprised I haven’t seen or heard of more innovation to use AI to decrease documentation burden throughout the healthcare ecosystem.
More patient-facing AI applications
Matt Sakumoto: I haven’t seen or heard enough about patient-facing LLM advances. With OpenNotes and lab results seen by patients before clinicians, can patients put that info into a LLM-powered product? Will DrGPT end up being better or worse than Dr Google when it comes to the patient-clinician interaction? I’m hopeful this will lead to more informed and empowered patients, and that it will be a net-positive benefit.
Similar to the hope for more physician vs computer studies, I expect we’ll see an increase in patient-computer interaction studies related to AI. Dr. Sakumoto makes the great point that patients have access to many similar tools and we can expect them to use AI-powered assistants in their medical care whether we think it’s a good idea or not. And as he mentions, patients can now immediately see their notes, labs, and imaging data on their patient portals. I know a few companies are trying to use AI to help people make sense of that information, and I agree that giving patients access to reliable information about the data in their medical record could be great.
More reimbursed AI-enabled SaMD
There are over 500 FDA-approved AI/ML SaMD products but only 16 of them are reimbursed currently and only four of them have over 1,000 claims. Most of these have only been approved for about a year and are likely still undergoing proof of efficacy studies to help negotiate with insurance payors, but the field is growing, as seen in the graphic below. I’d expect to see more AI-enabled devices in the coming months/years.
What do you expect to see in healthcare AI in 2024? What did I miss?