Marzyeh Ghassemi
Faculty Member, Vector Institute and Assistant Professor, Depts. of Computer Science and Medicine, University of Toronto
in Machine Learning for Health [at I WAS MOTIVATED TO DO MY PHD
MIT] because I believe we can improve healthcare with machine learning and AI. During my program, I would shadow physicians at morning rounds in the ICU, and I would often notice that two patients who looked very similar on paper would get very different treatments. At other times, the same patient would get very different treatment recommendations from two different doctors. When I asked my colleagues, ‘What is going on here?’, I was shocked to be told that doctors often have to make choices in an absence of evidence.
Traditionally, what we have used as ‘evidence’ in healthcare has been the results of randomized controlled trials (RCTS), which entail recruiting a study population. The question I’m concerned with — and that we should all be concerned with — is, do these studies generalize? And if so, for whom do they generalize? Even with the best intentions during recruitment, certain kinds of people see RCT advertisements. As a result, in practice, much of what we have learned about medicine has related to particular sub-groups of the population. And that translates into a lack of relevant evidence at care time. To make matters worse, healthcare professionals disagree — not just with each other, but with themselves. For radiologists, inter-rater reliability on the same image is around 67 per cent. If we don’t know for certain how to label our data, how can we use it to make predictions?
Looking ahead two years, I think the biggest impact of AI will be the automation of processes that are currently cumbersome and inefficient for people; things like order sets, or checking whether medications conflict with one another, or scheduling. Automating these things is going to make a big difference.
In 10 years, I believe AI can provide doctors with new ‘super powers’. For example, one thing that is often missed by the healthcare system is domestic violence. This is a really hard thing to recognize if you’re not looking for it. A patient shows up to the clinic with a small fracture here, a bruise there. In order to catch it early, the physician would need to sit down and review the patient’s entire record, have time to think about it, and have some insight into the person’s home situation. Given the extraordinary constraints on clinical staff ’s time and attention, that is not a likely scenario. But these kinds of patterns could be detected with algorithms, and used to direct attention as needed.
A clinician once told me that he felt the top 10 per cent of doctors operate in lock-step: If you give them a similar patient, they will probably have almost exactly the same recommendation, because they’ve all read exactly the same articles and seen the same volume of patients with the same range of severities. If variation in care comes from inexperience or lack of access, that would be another huge win for machine learning. We should focus on all the tasks that are not valuable for a human to do — all the things that we could train a machine to be very good at. That way, people will be freed up to make good judgment calls about what should happen next — and to talk to patients in a compassionate way about it.