Diagnostic models perform well on conditions clinicians already recognise easily and poorly on the ones that are missed. The inversion is a direct consequence of how the models are trained.
Why frequency drives performance
A supervised model learns patterns from labelled examples, and the number of examples for a condition tracks how often it occurs.
Common conditions supply thousands of cases across a wide range of presentations, so the model learns their variation thoroughly.
A condition appearing a handful of times in the dataset produces a weak and unstable pattern, if the model learns it at all.
How the training objective makes it worse
Training minimises overall error, and predicting the common condition is the strategy that minimises error across a realistic case mix.
A model can achieve excellent aggregate accuracy while never once identifying a rare condition correctly, because those cases contribute almost nothing to the total.
Headline accuracy figures therefore say very little about performance on the cases where diagnostic support would change an outcome.
Why the clinical need runs the other way
Common conditions are diagnosed reliably by clinicians who see them constantly, so a model agreeing with them adds limited value.
Rare conditions produce long diagnostic delays, repeated referrals and years of uncertainty, and that is where assistance would matter most.
The tool is consequently strongest where help is least needed, which is the central difficulty in this area rather than a temporary limitation.
What is done about it
Oversampling rare cases during training and weighting their errors more heavily both improve detection and increase false positives.
In a screening context, false positives carry real costs in unnecessary investigation, patient anxiety and clinician time, so the trade cannot simply be pushed in one direction.
Pooling data across institutions builds larger rare-disease datasets and runs into privacy constraints and inconsistent labelling between sites.
Where the framing is more productive
Systems designed to flag that a presentation is unusual, rather than to name the condition, avoid needing examples of each specific disease.
Detecting that a case does not fit any common pattern is a different task with far more available training signal, and it produces a useful prompt to investigate further.
That reframing matches how experienced clinicians actually work, since the diagnostic breakthrough on a rare case usually begins with someone noticing that the picture does not fit, rather than with immediately identifying what it is.