AI in Healthcare: Solving the Right Problems with the Right Tools

AI in Healthcare: Solving the Right Problems with the Right Tools

The most eye-opening experience of the ability to use data to improve care quality came early on in my career. I was tasked with pulling a list of patients who died after surgery and compiling some statistics on outcomes. What struck me most as I worked on the project was how many of the surgeries were like those that I had done. Very few of the cases involved patients who died immediately after surgery; rather, they were relatively ill patients who died weeks later. I realized how much information we as doctors were missing, and how better information could help us make better decisions. To me that is AI, not artificial intelligence but, augmented intelligence.

As both a physician and someone deeply involved in AI development, I’m frequently asked whether we are at the tipping point of when artificial intelligence is ready to transform healthcare. I understand the urgency of the question; I've felt it myself during late nights staring at fragmented records, wondering what I might be missing on a given patient. AI is not a panacea, and trying to retrofit it into every aspect of care is not only misguided but potentially dangerous. The key is to match the tool to the task.

The challenges we ought to tackle first

Identifying high-risk patients

Healthcare operates under tight resource constraints—clinicians don’t have the time, staff, or infrastructure to provide every possible test or procedure to every patient. Current society guidelines often fail to account for these constraints assuming everything is available to everyone, a particular mismatch in lower income countries. AI models can help prioritize patients, spotlighting those who would benefit most from early intervention, frequent monitoring, or targeted testing. For instance, not everyone needs a colonoscopy, mammogram, or PSA test at the same age or frequency. And yet, we’ve all seen patients who were over-tested, and others who fell through the cracks. AI can help stratify risk and create personalized screening models, reducing unnecessary procedures and catching disease earlier for those who need it most.‍

Resource planning

Most hospital staffing models are static—frozen in time. Yet demand varies by the hour, day, and season. Imagine if Starbucks scheduled the same number of baristas for 2 p.m. on a Wednesday as they do for the Monday morning rush. It would be chaos. Hospitals face the same mismatch. AI can help forecast patient flow, adjust staffing in real time, and even dynamically set appointment durations based on patient complexity. Sicker or more complex patients can be scheduled for longer visits, while routine follow-ups may require less time. This optimizes both efficiency and patient care while reducing stress for burned-out providers.

It's just as important to recognize what AI is not ready for and may never be.

Replacing Clinicians

Healthcare is not a game of probabilities alone. The “right” answer to a clinical question is not always what an algorithm determines statistically, but what a patient decides in conversation with their physician, based on values, preferences, and ethics. A model might suggest surgery is statistically optimal, but what if the patient values quality of life over longevity? That’s not a decision AI can—or should—make.

Correspondence or interpreting test results

Individual results that are “normal” or “abnormal” in isolation may take on a whole new meaning when you understand the entire patient history. Consider two “positive” biopsy results for cancer – one in a patient who has never been diagnosed and the other in a patient undergoing chemo. The same result may have very different meanings. But AI doesn’t really understand anything; it makes a lot of inferences, which means it may deliver the wrong message to a patient. ‍

Treatment selection

AI has shown flashes of potential here, particularly with techniques like reinforcement learning that can simulate decision paths. But so far, no model has come close to replicating the layered, nuanced decision-making clinicians do every day. Our choices aren’t just based on lab values and guidelines; they’re shaped by context, co-morbidities, patient history, psychosocial dynamics, and a healthy dose of human intuition. These are incredibly difficult to codify, let alone replicate in an algorithm.

So where does this leave us?

The most successful AI implementations won’t come from building models and searching for problems, but from starting with a clinical need and designing purpose-built solutions to address it.

In other words, let clinicians lead. Let us identify the inefficiencies, the bottlenecks, the pain points that slow us down or make care less safe. Then let technologists collaborate with us to solve them. Just last week, I had a patient who was scheduled for surgery that got cancelled because someone missed the fact that she was on Ozempic and the fasting guidelines were different. A tool to identify these patients at scale and get their personalized fasting guidelines isn’t fancy, or even complicated, but it does make a real difference in patient care. It’s not about taking the fanciest LLM to solve the problem; it’s about solving the biggest problems. Rearranging the schedule for that patient took over an hour out of my day, and the operating room was left empty meaning another patient had to wait longer for their surgery.

That’s how we’ll unlock AI’s full potential—not by dreaming of robot doctors, but by creating smarter, more responsive systems that support the ones we already have. The future of healthcare isn’t AI versus doctors. It’s AI with doctors, for better, more humane care.