The Questions Accuracy Cannot Answer
On medicine, uncertainty and artificial intelligence
I have been thinking about ECGs more than I expected to over the last few months.
Not because I have suddenly become interested in cardiology. My work still revolves largely around diabetes and metabolic medicine. But somewhere along the way I realized that an ECG is a surprisingly good place to think about a much larger question: what exactly do we mean when we say that an artificial intelligence system “understands” medicine?
Whenever a new AI model for ECG interpretation is reported, the first number that appears is usually its accuracy. Ninety-seven percent. Ninety-eight percent. Occasionally even higher. Those numbers are impressive, and they should be. Modern machine learning has achieved things that would have sounded implausible only a decade ago.
Yet I find myself lingering on a different question.
What happens when the ECG refuses to be read?
Every physician has encountered such tracings. They are rarely dramatic. In fact, they are often ordinary at first glance. The rhythm seems obvious until one lead raises a small doubt. A baseline wanders just enough to make atrial activity uncertain. A premature beat appears straightforward until its timing suggests another possibility. Nothing is obviously wrong, but neither does everything fit comfortably together.
These are not necessarily rare ECGs. They are simply the ones that resist certainty.
Get sanyalsoham’s stories in your inbox
Join Medium for free to get updates from this writer.
Medical education teaches us diagnostic criteria. Clinical practice teaches us something more subtle. It teaches us that before interpreting a tracing, we often end up interpreting the tracing itself. Is the recording reliable? Is there too much noise? Could a technical issue be creating an illusion? Am I trusting one lead more than another? We perform these small acts of judgement almost unconsciously, yet they shape many of our decisions.
The more I think about artificial intelligence in medicine, the more I wonder whether this quiet process of judging the quality of evidence deserves greater attention.
Much of the discussion around clinical AI revolves around whether an algorithm reaches the correct diagnosis. That is, of course, an important question. But it assumes that every diagnostic problem begins with dependable information. Real clinical practice is less accommodating. Data are often incomplete, noisy, contradictory or simply untidy.
Perhaps this is why experienced clinicians become less impressed by certainty as they grow older. Experience rarely eliminates doubt. If anything, it teaches us to recognize situations where confidence should be tempered by caution. Some of the safest physicians I have worked with were not the ones who sounded the most certain. They were the ones who knew when certainty itself deserved scrutiny.
I sometimes wonder whether this is where the conversation around artificial intelligence should evolve. Rather than asking only how accurately an algorithm classifies disease, perhaps we should also ask how it behaves when confronted with imperfect evidence. Does it recognize ambiguity? Can it communicate uncertainty? Does it know when another ECG, another lead, or simply another opinion might be appropriate?
I don’t know the answers to those questions, and I suspect neither medicine nor computer science has fully answered them yet.
What I do know is that these questions seem more interesting to me than another percentage point of accuracy.
Medicine has never been practiced across datasets. It has always been practiced one patient at a time. Every so often, a patient presents us with an ECG that quietly refuses to be read. Those moments can be frustrating, but they also remind us why medicine is more than pattern recognition.
Perhaps the future relationship between clinicians and artificial intelligence will not be defined by who recognizes patterns more efficiently. Perhaps it will be defined by who recognizes uncertainty more honestly.
That, at least, is a question I think is worth sitting with.
