I think the correct way to think about AI in general is that it is a tool. Not a friend, not a savior, not a new entity. It is merely a tool, like a hammer or a saw. Through this lens, I would like to consider near-term uses that will have the most utility for physicians.
Electronic health records (EHRs) are ubiquitous in medicine. In my experience to this point, they have been developed in such a way that they offer little more than improved legibility over paper charts. Otherwise, they behave no better than physical filing cabinets with a rudimentary search function, forcing physicians to simply dig through everything. Even worse than filing cabinets, computers allow us to produce so many more notes and so much more text and other data that useful information is rapidly buried. On top of that, even if there are useful features (like a list of chronic problems attached to a patient’s chart), few physicians have the time to appropriately organize or update them, so even these useful features become untrustworthy quickly.
So to start, and this work has in fact started to a degree, AI should summarize the chart. Currently, there are tools that briefly summarize an admission as well as current events (the last 24 hours) for the hospitalization. These are useful, but it’s time to go a step further. There should be an LLM layer on top of the EHR. It should be trivial to ask the LLM about the patient’s chart, chatting with it to get up to date on the patient’s history and their current and past workup. When I say chat, I do not mean for it to act as a consultant, or that it is directly coming up with differential diagnoses or plans. Instead, I mean it to act like a much-improved search engine, allowing the physician to ask directed questions about the information located in the patient’s chart. I believe this can be done safely (meaning no hallucinations) by making the only reference data for the EHR LLM the data that is within the patient’s chart. To further decrease hallucinations, at least in our current state, the bar for the LLM to say that it does not know should be relatively low. More gracefully, it can link a physician to a part of the chart it determines may have the answer for the physician to review themselves when its confidence is moderately low (this is in a addition to the LLM always giving links to references when it does pull information). In essence, the LLM should make it trivial for physicians to find information in the chart along with the ability to summarize, trend, and link to the patient’s history and data as well as the current clinical course. The goal should be near-zero time hunting through the chart and all the time spent actually reviewing the desired information.
As mentioned previously, the EHR contains tools that can be helpful (like chronic history, past admission diagnoses) but are often disorganized or out of date. AI should be able to assist with cleaning these pieces up. This would include what I have already mentioned, as well as the patient’s current problem list (or active diagnoses), medical history, family history, surgical history, social history, and vaccination history, among others. This should be built in a way that the physician can confirm these recommended changes and additions. There should be nothing that is occurring completely out of view of the physician. The AI is not the main architect of the chart, merely an assistive tool.
Just as you wouldn’t simply start a buzz saw and let it run roughshod over a pile of wood to try to build a house, AI should not be allowed any more freedom. Seeing AI as a tool and not some advanced entity will stop any decisions from being made that allow AI to run unregulated and unmonitored. Our goal is to develop a tool to make the tedious work more efficient and allow physicians to focus on what they are there for. The goal of AI is no different from that of any other tool.
Searching through the chart, manually typing the history of present illness, updating the patient’s various histories, these efforts have become so bloated that they have crowded out the actual care of the patient: the human-to-human interaction and the deep consideration of the assessment and plan. The aforementioned pieces need to be considered, but the physician should spend little time manually entering them. I believe we are finally in a place where we can begin to reverse what we have wrought with the EHR, using AI to get the tedium out of the way.
Finally, we come to transparency, which we have touched on a bit. It should be very clear when AI is updating something, entering text, where it is pulling information from, etc. The physician should have an easily attainable bird’s-eye view of exactly what the AI is doing and suggesting. This keeps the physician engaged as they still hold the ultimate responsibility, while also offering a safe on-ramp to progressively increasing AI involvement in boots-on-the-ground medical care. Although all of this mainly focuses on the functionality of the AI, here we must consider the human element: the design of how the tool functions. We must consider how to easily reveal the AI’s changes and suggestions, the bird’s-eye view I mention above. We don’t want to remove tedious work only to replace it with other tedious work. It should be easy to see and approve what the AI is doing, like a dashboard of the AI’s work.
Considering and designing AI as outlined will decrease the tedious work of the physician, make EHRs much more useful and efficient than they have ever been, and keep physicians directly in the loop of the most important work, ensuring they produce what only they can appropriately produce right now, and giving them time to focus on just that.
In service of transparency: AI wrote no part of this piece. It did do a grammar, spelling, and punctuation check which I then double-checked, hand-typing the changes.