Deconstructing the diagnostic reasoning of human versus artificial intelligence

The place of AI in medicine has a certain fascination, mostly of the sci-fi variety, but I thoroughly appreciated this article's description of the mechanics of how AI "learns" as well as "reaches a diagnosis". I would go a bit further than the authors though.

The quality of an AI diagnosis is strictly limited by the quality of it's model, and the "clean-ness" of the dataset it learns from, and then induces from. Our human understanding of disease causality is neither perfect, nor complete, making creation of good models challenging. Much more challenging is "clean data", meaning symptom X, or physical finding Y, or test result Z, are true descriptions of the patient’s condition. This can be due to a multitude of factors. Some (e.g. test results) might in future be made more accurate, and some (e.g. "I'm feeling dizzy”) will very likely never be accurate in the sense required by AI. This is because a human patient (their own history, and the multitude of external factors affecting them) is infinitely complex - figuratively, and perhaps literally. Reducing all of these potential causative factors to computable “data points” requires both simplification and evaluation, both of which tend to change the data itself. That’s why AI has really only proven effective at diagnostic tasks that involve relatively clean “data” - diagnostic imaging, photographs of skin lesions and retinas, and certain conditions with relatively well-understood and objective markers.

On the other hand, humans have had hundreds of thousands of years of practice at picking up on subtle cues emanating from other humans, and when this is combined with quality medical training and "practiced experience”, a physician's diagnostic ability is generally very good (and certainly very efficient, despite it’s potential flaws). However, we share something with AI - the requirement for quality feedback on whether we’re right or not. This is particularly challenging for generalists like family physicians, typically dealing with undifferentiated problems having a wide differential, 80% of which resolve on their own. Anesthetists, on the other hand, probably have a more dependable intuition, because they tend to get immediate feedback on their decisions. The point is that AI is likely to be more accurate than humans when the diagnostic model is relatively simple, the data involved is high-quality and “clean”, and especially where errors related to human reasoning are common.

The other point to make is that the practice of medicine isn’t simply about diagnosis, but rather helping the patient attain, or regain, wellness. And this is why AI will never replace physicians. The ability to imagine novel solutions to complex problems, to feel empathy, and to make decisions when it’s entirely unclear if there’s a “right answer”, are all uniquely human abilities which, by it’s nature, AI is incapable of. Physicians are not competing with AI; we have different, and sometimes complimentary, strengths and weaknesses. We should carefully assess which type of questions and answers AI can help us with, and just as carefully consider how our own powers can best be used, to the benefit of the patient.

Seeking to compare the reasoning of human and artificial intelligence in the context of medical diagnosis is an overly optimistic anthropomorphism. The term artificial intelligence (AI), as used to describe machine learning algorithms employed in this domain, is itself a misnomer. This is apparent when comparing modern machine learning algorithms based on artificial neural networks to non neural algorithms (e.g. logistic regression). Unfortunately, this comparison was not made by the authors.

Logistic regression, established in the 1800s, is the machine learning algorithm most commonly applied to structured medical data for diagnostic and prognostic purposes (e.g. Framingham Risk Score, Kocher Criteria, etc.). The same nomenclature of “learning” or “training” is equally well applied to this algorithm that we have been using for centuries. Simply put, machine learning algorithms are mathematical formulae with free parameters derived retrospectively from clinical data. These formulae are not intelligent according to even the most generous of definitions, and they have no capacity for reasoning.

It is notable that non-neural machine learning algorithms are still the most accurate for structured clinical data, and continue to dominate the field. Neural network algorithms bring not intelligence, but rather the capacity to model more complex unstructured data (specifically natural language, images, and time series) and incorporate this information into our predictive tools. This capability for modelling complex inputs is the most compelling advantage of neural network algorithms.

However, the downside of this complexity is that neural network models are typically uninterpretable, meaning humans can not understand or explain how the prediction is derived from the clinical data. The greatest risk in deploying neural network algorithms, or any machine learning algorithm with limited interpretability, is to assign too much trust to it and ignore the potential for unknown confounders or biases. Such confounders and biases could harm patients and are notoriously hard to identify and correct for. By conflating human and machine intelligence, we further this risk.

Fortunately or unfortunately, true artificial intelligence with the capacity for reasoning, remains for now in the realm of science fiction. We should not pretend otherwise, because while there are many benefits of the technology currently available to us, there is also real capacity for harm by using it inappropriately.