A new arXiv paper argues that clinical AI models often carry more detail than any actual medical decision requires.
The researchers propose what they call decision-sufficient abstraction: a model should only be as detailed as the decision it's meant to support, not as detailed as the data allows. They argue clinical reasoning is usually ordinal-first, built around crossing thresholds rather than precise numbers. A doctor might first ask whether mortality risk is below 10 percent, and only if that doesn't settle the choice, ask whether it's below 5 percent. The paper formalizes this as ordered threshold refinement, where a model needs just enough resolution to preserve the outcomes relevant to whichever threshold has been reached so far.
That's a real change in emphasis for a field that mostly chases higher accuracy and finer-grained predictions by default. If a coarse threshold already resolves the clinical choice, refining the model further doesn't help anyone - it just adds complexity that has to be validated, explained, and trusted for no practical gain. The framework also implies a smarter way to spend scarce clinical evidence: aim it at whichever threshold is still undecided, instead of trying to improve the whole model uniformly.
It's worth noting this is a theoretical framework, not a deployed tool - the real test is whether hospitals can identify the right thresholds for a given decision, and whether model builders are willing to build less instead of more.