Medical Condition Classification Platform

A machine-learning platform that suggests likely ICD-10-related condition categories from institutional healthcare records. The model was trained on real institutional data with substantial class imbalance and deployed on AWS behind a load balancer with HTTPS. It is presented here as a decision-support ranking tool, which is what its measured accuracy actually supports.

Institutional healthcare records carry a large number of condition categories, and their distribution is severely uneven: a handful of categories dominate while a long tail appears rarely. A classifier trained on that data will look competent on aggregate metrics while being close to useless on the categories that most need help.

The framing question was therefore not 'can a model predict the condition' but 'is there a form of prediction that is actually usable by someone doing this work'.

Top-1 accuracy is 29.26%. On its own that is a weak classifier, and presenting it as anything else would be dishonest.

The useful figures are the ranked ones: the correct category is in the model's top three 53.79% of the time and in its top ten 88.76% of the time. The gap between 29% and 89% is the whole finding. It says the model has learned a great deal about which categories are plausible for a given record, and comparatively little about how to choose between the few that remain.

Against a large, heavily imbalanced label space, that is the expected shape. A single prediction forces the model to spend its confidence on a decision the data does not support; a short ranked list matches what it actually knows.

A tool that proposes one category and is wrong seven times out of ten will be ignored after a week. A tool that narrows hundreds of categories to ten, and is right about nine times in ten that the answer is somewhere in that list, removes most of the search while leaving the judgement where it belongs.

That reframing changed what the system had to be. It is not an automated classifier; it is a shortlist generator whose job is to be reliably inclusive rather than occasionally exactly right.

This is an academic project. It has not been clinically validated, it is not a medical device, and it must not be used to make or support a diagnosis.

Two limitations matter most. Class imbalance means the rare categories — often the ones where assistance is most valuable — are the ones the model has seen least and predicts worst; aggregate accuracy hides this entirely. And a ranked shortlist creates its own risk: a plausible-looking list can anchor a reader towards its contents and away from a correct answer that is not on it.