AI could transform health care
Recently, my colleagues and I at Stanford University launched a study of a promising AI tool to screen for colon cancer. Earlier trials had suggested the tool could help doctors detect potentially cancerous growths called polyps during colonoscopies. Our study was one of the first to roll it out in a realworld clinical environment.
What we found shocked us: The tool did not improve polyp detection rates or other outcomes. But then this is not the first artificial intelligence health care product that has failed to deliver.
During the COVID-19 pandemic, several AI applications promised to help with the detection and management of the disease but fell flat in clinical settings where human behavior and the real-world environment can interact with the tool and alter its efficacy.
AI has real potential to improve medicine, but it is hard to discern which tools hold the greatest promise. With billions of dollars flowing into health care AI, the technology must be rigorously tested. To do so, our existing approach to medical research and clinical trials needs to adapt.
AI is a game changer for medicine and research, especially when combined with smartphones and wearable devices that collect data from patients as they live their everyday lives. Researchers, physicians and patients are using AI to track and predict sleep, blood pressure, heart abnormalities, fertility, fitness and a host of other health markers.
The technology opens up a world of new health outcomes to measure. For example, diabetes care and research usually focus on hemoglobin A1c, a one-time metric taken in a lab that estimates someone’s blood sugar levels during the previous few months. But in a study I am co-leading, we give patients who are newly diagnosed with Type I diabetes continuous glucose monitors that track their blood sugar throughout the day, not just one single time point. We can use AI to adapt their care based on their current glucose levels and to offer insight into how diet, exercise, medication and other factors affect each individual’s management of the disease.
How humans — patients, the medical team and researchers — engage with AI-based tools also plays an important role in how much of the technology’s promise can be realized. To that end, we can tweak interventions in real time to evaluate and refine how the care team interacts with patients. Clinical research techniques like micro-randomization — where individuals may be randomized hundreds or thousands of times throughout the study — allow us to test different pathways through which a tool can improve health in a single study.
For example, we might have multiple times throughout the day when we can prompt someone to change their medication dosing and different ways to frame these messages. We can microrandomize participants to different combinations of engagement cadence or messaging language to find which approach works best and adapt it to the needs of different types of participants.
AI and digital health devices also enable greater interaction with patients. Technology can not only gather and analyze more data, but also continuously push information out to patients, making them active participants in the treatment process and medical research. But patients need help to make sense of all the health data that is increasingly available to them. An AIdriven tool that tracks your heart rate every second may sound cool but could be counterproductive if you don’t know what to do with that information.
More data opens up new possibilities, but it also brings more complexity and noise. My Stanford colleagues and I conducted a study using Apple Watches to detect irregular heart rhythms. We monitored roughly 420,000 people as they went about their day-to-day lives and referred more than 2,000 for follow-up, a scale that would be almost impossible to achieve using traditional clinical trial enrollment strategies.
However, we also had to contend with messy data. Was a patient actually wearing the watch at a given moment? Had they taken it off or lent it to a family member? If the watch battery was low, did that affect the accuracy of the timestamp on the data, scrambling the timing of important events?
The statistics and modeling we used for the Apple heart study were fairly simple. Managing the data was incredibly complicated, however. Only by leaning on a diverse research team that included data scientists and engineers, in addition to physicians, clinical experts and patients, were we able to find creative solutions for our data challenges.
Even more futuristic technologies are on the horizon (or already here), from immersive headsets like Apple Vision Pro to Band-Aid-like wearable sensors that wirelessly beam data to our phones. With all this data, it’s essential that data scientists and the engineers who build the apps and devices be integrally involved in evaluating AI health care products.
There are so many exciting ways AI is transforming medicine, but there is also plenty of unfounded hype. It is only through rigorous testing in the real world with strong consideration of the human-AI synergy that we will be able to discern between the two.
No matter how exciting they sound, AI applications have value only if they really move the needle on outcomes that matter to patients and if they help physicians and patients in meaningful, verifiable ways.