Diagnostic Reasoning
Enough to be sure? A diagnosis, as the evidence arrives over days
A child has a fever, a rash, and joint pain. The clues do not all arrive at once: history on day one, imaging the next, the decisive lab days later. The real question is not just which cause, but when there is finally enough signal to commit. Ask an AI on day one and it will still answer, confidently, from partial evidence. This shows what a careful clinician, or a careful AI, actually needs: the point at which the case is truly solved.
Concept and interaction design by Hunter Whitney, with Veena R. Kumar, MD, pediatric emergency physician. Case published in Contemporary Pediatrics.
Fast, and now that the evidence is in, correct. The catch: asked on day one, it would have answered just as confidently, from far less. Knowing when there is enough is the harder, and more valuable, judgment.
The culture is the signal that tips it. The same answer, reached in the open, with a clear record of the moment it became safe to commit.
The audit trail. However the answer was reached, by an AI or by hand, this is the record a physician can inspect: not only which evidence promoted rat-bite fever and which findings ruled the others out, but when there was finally enough to commit, and what was still missing before that. Traceable, teachable, and a guard against deciding too early. The clinician stays in command of the reasoning.
The same mechanism, three places
Each node sums what the evidence pushes for and against it: supporting signals in green, opposing in red. When support crosses threshold, the node fires toward the next, the way a neuron does. Crucially, the inputs arrive over time, so a node can sit below threshold for days before a single decisive signal tips it. That is temporal summation: a neuron integrating inputs across time until it fires.
The same integrate-and-fire-over-time pattern is how a clinician reasons toward a diagnosis, how a learning system settles on an answer, and how I think about designing for AI: keeping the reasoning, and the moment it becomes sufficient, legible, rather than handing over a confident answer built on partial data. This concept was reaching for that idea years before the language caught up to it.
The clue that tips it: on day four, the joint-fluid culture returns positive for Streptobacillus moniliformis. Combined with the day-one history (the child handles and sleeps with a pet rat that bit both dogs), it crosses threshold and promotes a single infectious cause: rat-bite fever. Before that, the case was suggestive but not decisive, which is exactly what a careful reasoner, human or machine, should be able to say out loud.