Separating causation and correlation in AI systems is a challenge because most machine learning systems look for trends, but not 'counterfactual' information, which is more like the way we, humans, and doctors think.
People, like doctors, make decision based on what they know to be true and untrue, and build causal reasoning into a diagnosis. Most #machinelearning systems don't build causality: they are built on associations / correlations. We don't care the sky is probably blue when we get a cold, but we do care your T-cell count is low when you get a cold (causation vs. correlation). Counterfactual data? "Lets get a chest x-ray / ultrasound / CBC ..." i.e., some data that rules out other possibilities to see how symptom relates to disorder (directly or indirectly). But what rules can you build for machine learning? (Un)surprisingly this paper shows it can be simple, (because thats how *our* brains probably work): disease should be consistent with diagnosis, rule out stuff that isn't possible, and keep it simple: 1 Dx fitting M symptoms is better than N Dx fitting M symptoms. They go on to define things called "expected disablement" and "expected sufficiency". The former is obvious, but the latter is like "sufficient cause", and state theorems, one of which is that disablement and sufficiency are sufficient conditions for the rules above. But real data is noisy and murks the variables and so there needs to be a way to account for noise. (insert mathy stuff here). Thats all fine, but the litmus test is "How does this compare to actual clinical decisions?" In short, a physician achieves higher accuracy in diagnosing a disorder for simpler problems, and the algorithm outperforms for more complex problems. Thats good for rare disease classification. That makes sense as the story of #machinelearning and #AI in medical diagnoses suggests utility in a role as a 'decision support tool', but not a fully autonomous one. The difference here is that the model behaves more like a clinician would. For you Bayesians ... when you first learned Bayes' Theorem I bet you pondered "Why can't we do counterfactual inference in medical diagnosis? ...policy making? ... court decisions?". This article is a nice progression of how we can use AI based on causation - not just correlation. Don't believe me? Read for yourself.
No comments:
Post a Comment