Accuracy & Its Limits
A screening model runs return "healthy" on all 1,000 patients, misses the single sick one, and still prints 99.9% on its report card. On paper it is the best model in the building, in the clinic it is worthless.
- ▸Flag no one on a rare disease, score 99.9%, and catch not one sick patient
- ▸Two screens can share one score yet clear very different numbers of sick patients
- ▸No cut-off catches every case without crying wolf, so pick by which mistake costs more
A hospital wants a screen for a rare disease that shows up in about 1 of every 1,000 patients. You build the laziest possible model: it clears every patient, sick or not, without looking. Run it on 1,000 patients and it is right 999 times, because 999 of them really are healthy. Its score reads 99.9%. By that number it is nearly flawless, and it caught not a single sick person.
The number is not lying about the arithmetic. It is lying about what matters. A score that just counts how often the model was right will always flatter whichever answer is common, and here the common answer is healthy. The rarer the disease gets, the higher the lazy model climbs, because being wrong on the one case anyone cares about barely moves an average built from 1,000 patients. One headline number cannot tell you the model is useless, it is too busy being technically correct.