AI in healthcare: How engineers and clinicians at U-M are unlocking its potential together

When we think about what artificial intelligence is, it’s the ability to reason—to make predictions and decisions. But the workhorse behind the advancements we’re seeing today is machine learning. It can be integrated into clinical workflows in a lot of different ways. With large language models, you get tools like AI scribes that can help clinicians with notes.

Machine learning was also behind the revolution in computer vision starting around 2012 that led to better models for recognizing conditions in medical images.

Machine learning needs data. In today’s healthcare system, we’re collecting an immense amount of data, whether it’s clinical, medical imaging or vital sign measurements entered into the electronic health records, or data from wearables like a continuous glucose monitor. It’s too much for an individual or even a team to review. So AI in healthcare starts with approaches that can organize and make sense of all these data and turn them into actionable insights.

Practicing medicine is hard. And it feels like it’s getting harder. We’re awash in all this data, as Jenna said. Also the pace of medical discovery is shockingly fast now. It’s hard for folks to keep up. So I think of AI as systems that can help clinicians deliver really high quality care amid these challenges. That may mean summarizing large quantities of text to identify insights, helping clinicians recognize things they might not have appreciated themselves, making recommendations for what to consider doing next or unburdening them from a lot of the other aspects of patient care like documentation.

It’s hard to predict what the future will look like, honestly. We all can feel that things are really starting to accelerate and these technologies are finally being brought to the bedside. But it’s still early days.

We all dream of a future where health care is better and costs less, right? Several key pieces are needed before we reach that point. We’ve only recently figured out how to collect data inside the hospital. A lot of types of care, especially preventative care, can happen outside the hospital. So we also need the ability to collect data outside of the hospital, and then the infrastructure to connect the data across these different sources so that they can be used as input to machine learning models.

Once we’re able to scale the delivery of high-quality care, it becomes about personalization and optimization—how do I find the right treatment for you?

There’s no way we could do what we do without clinical collaborators. It would be more than just suboptimal, it could be dangerous. As an example, we’re working on developing a medical imaging model to improve diagnosis when patients come in with shortness of breath. Early on in the project, we trained a model to identify the cause of acute respiratory failure based on chest X-rays. It seemed to be doing really well. Then we went to see what it had learned, so we looked at what part of the images the model was using to make its predictions. In one patient, it had pointed to heart failure, but the clinicians could tell the part of the image it focused on wasn’t the heart. It was a pacemaker. This is a type of spurious correlation or ‘shortcut’ in this diagnostic task. Relying on this shortcut, instead of relevant radiological findings, could result in poor performance of the model if the relationship between heart failure and pacemaker changes.

We’re working on this project because respiratory conditions can be life-threatening. And it’s not always immediately clear what’s going on. Sometimes, in up to 30% of cases, clinicians get the initial diagnosis wrong. The earlier we get the right diagnosis and we can provide the right treatment, the better the patient’s outcome.

We have a preliminary model in which we’ve addressed the shortcut issue that we think works pretty well, and we’re now working on studies to better understand how it affects clinicians’ decision making.

Our work goes beyond developing the model. We need to determine how to ensure that clinicians use it in an optimal way. It’s a human-AI collaboration. But it’s ultimately the clinician that’s making the diagnostic decision.

AI is being integrated into clinical workflows in many different ways. One example is risk prediction or early warning systems models. We built a model that predicts a patient’s risk of acquiring a C.diff bacterial gastrointestinal infection during their hospitalization. It’s now integrated in the health record and the pharmacist’s workflow, and it’s identifying patients at higher risk due to being treated with antimicrobials. We’re actually moving the needle on antimicrobial stewardship.

There are also other early warning systems in use that haven’t been built in-house—things like the Epic sepsis model and deterioration index. Pathology also relies on a lot of machine learning and automated tools to process histopathology data.

In radiology new AI tools point out particular areas that might warrant further review, like on a chest X-ray. Another type of tool that is having a big impact are AI scribes. When you visit the doctor, he or she may ask to turn on a recording so AI can help generate notes. Clinicians are saying it’s changed their life in terms of not having to spend hours at home every night finishing clinical documentation. The AI scribe produces a first draft of the clinic visit notes, which the doctor would then review and edit as needed.

The AI scribe is an active area of research for us. A complete draft might not be the optimal solution. We want these tools to enable collaboration, right? Part of a collaboration is doctors using their own knowledge and experience and we want to leave space for that.

The sense I get from my colleagues is that patients mostly appreciate it. When the AI scribe is in the room, the doctor can focus on talking to the patient as opposed to also typing on a computer. The hope is that it can help doctors be better at things they want to do, like building a relationship and listening to the patient.