If you're trying to keep up with AI without wasting time on the wrong layer, this is the detail that matters: in rare childhood disease, the first AI job is not reading genes. It is turning messy charts into HPO, a shared symptom language, before the gene model even starts [C001].
That sounds like the boring part. It is also the part that decides whether the rest works. If one hospital writes a symptom one way and another writes it differently, the downstream gene step starts half-blind. Better models do not clean the input for you.
One study trained on 2,671 patients from a rare-disease network and tested on 16,357 real clinical records outside training. The full notes-to-HPO pipeline scored 0.70 vs 0.58 for an older baseline. That is enough to matter at the shortlist stage, not just on paper.
A separate paper shows why this bottleneck is real: only 2.2% of billing codes mapped directly to HPO, and in real health records the mapping stayed under 50%. If the symptom layer is this patchy, better gene ranking is still starting from bad input.
"Using AI to help physicians diagnose rare genetic diseases affecting children" [C001] is the headline. The decision-changing version is simpler: an AI update matters when it fixes the bottleneck, not when it adds more model magic. Share this with anyone still obsessing over the gene model.