If you are like most people, at some point in your life you will experience a medical misdiagnosis of some kind. As someone who has experienced many, some much higher-stakes than others, I can attest that even the lower-stakes errors can result in a lot of frustration, wasted time, and prolonged suffering.
It seems obvious enough to me that human error is one of the core issues. Although the advent of AI doctors seems like an inevitability, diagnostic conclusions for now still usually rely on humans to make the final call. MRIs or CT scans, for example, are usually sent to a trained radiologist, who reads the result and writes up a report that is then interpreted and communicated by the patient’s provider. The diagnosis and therapeutic intervention is therefore the result of expert intuition from a select pair of trained specialists. Equivalently, they are the result of these experts’ combined knowledge and training, and also their lack of knowledge and lack of training.
Back in 2018, it was first shown that AI outperforms most expert dermatologists at detecting skin cancer, and reams of papers have come out since then improving on those results. If we go a bit further back to 2015, we find evidence that individual dermatologists misdiagnose at much higher rates than they would if they combined their votes on each diagnosis (though this ‘collective intelligence’ is still much worse than AI). The authors argue that this effect is due to each doctor having their own biases - their own way of looking at the same signs, symptoms, and test results, and drawing their own unique and often incorrect conclusions.
This shouldn’t really be too surprising. Each doctor’s understanding is innately limited by their own unique experience, training, and of course cognitive abilities. In other words, each doctor can only fit their own limited mental model of disease to a limited sample of the data, leading to different doctors having different diagnostic models that are all suboptimal in their own unique way.
In the parlance of machine learning and artificial intelligence, these kinds of models are termed ‘weak learners’. Concretely, a weak learner is an oversimplified statistical model that is often only allowed to ‘see’ a limited subset of data.
In the ML literature, weak learners actually serve a very important role in producing very high performing models, by ensembling many weak learners on the same task and then making final predictions by taking a vote among your learners. In fact, for many types of problems, particularly those where the data is tabular, these types of models are consistently the best performing (see here and here). The best predictions emerge like a forest zooming out from lots of individual trees, where each tree is one guess that helps to build an overall consensus. But typically you need to ensemble across a huge number of these weak learners (i.e. hundreds, thousands, or more depending on the size of the data). It also helps to train the learners sequentially so that each subsequent learner is an improvement of the previous.
One of the findings of the 2015 skin cancer study was that, indeed, if you ensemble the predictions of all doctors, you get predictions that are on average far better than those of any individual doctor’s. Even so, AI now vastly outperforms the best human experts at skin cancer detection as well as a host of other imaging based diagnostics.
But, two important questions remain: 1. Are doctors aware that they’re not very accurate? 2. Are they communicating this to their patients?
Logically, I would think “no.” First of all, most established diagnostic methods are taken for granted - their exact sensitivity and specificity are often not considered in a patient setting. Second, determining real-world diagnostic accuracy would require that doctors get unbiased feedback from an unbiased sample of their real patients, which isn’t happening when disgruntled patients decide to change doctors. Third, even if doctors were aware, it’s not clear that sufficient incentives are in place to guarantee learning.
Anecdotally, I think patients would also say “no” - i.e. that they often can’t tell if their doctor is really sure or just guessing. I could be wrong though.
Finally, there is some data. In one study, general internists were given 2 easy and 2 difficult case vignettes to diagnose. Doctors were right 55.3% of the time in the easy cases, but only 5.8% in the difficult cases. Regardless, their confidence was nearly the same - 7.2 out of 10 in easy cases versus 6.4 in difficult. That’s so close that the authors suspect the difference is “clinically insignificant.”
So we can establish that an individual doctor/provider can’t be all that accurate, because they’re basically a weak learner. But in many cases they’re also overconfident, and that’s a problem, because there’s a connection between error rates and doctors’ individual mental models. More generally, that suggests a connection between doctors’ overconfidence and misdiagnosis across the healthcare spectrum.
I’m not sure how providers should try to estimate their uncertainty, but it seems obvious that they would be better able to calibrate and recalibrate over time if the labels they were producing were at least tracked somewhere and subsequently evaluated. One possibility is that with some regular cadence, the EHR system prompts providers/clinics to reach back out to patients, especially the ones that they **aren’t* seeing anymore, until the SOAP assessment is confirmed or disconfirmed by the patients themselves (for example based on their reported outcomes following the recommended treatment plan or via further evaluations at other clinics). My hypothesis is that simply knowing when you’re wrong should be enough to effect at least some amount of relevant change in most providers. This assumes that most providers take seriously their code of ethics, and I think it’s OK to assume that.
Even so, this wouldn’t be a trivial system to implement. Standardizing a protocol for follow-up analysis coud be a major challenge technically, legally, and behaviorally. However, using some sort of EHR is practically mandated by most clinics and providers nowadays already, despite legally it being an opt-in system. Bolting on additional features for follow-up and reporting seems technically solvable. Legally, there’s already multiple bills that mandate providers to undertake specific measures for continual quality improvement. Relative to those bills e.g. ACA, CoPs, HCQIA, this would be highly aligned and could be relatively limited in scope. And now I’m venturing into territory I really don’t know so I’ll stop here. Hope you enjoyed reading.