The results
How well does it work?
These numbers come from 154 patients the model never saw while training. They are decent, not perfect, and it is worth being upfront about that.
- – Accuracy Share of all patients classified correctly
- – Recall Patients with diabetes that it caught
- – Precision Of those flagged, how many truly had it
- – ROC AUC How well it ranks risky patients above safe ones
ROC curve
The further the curve bows toward the top left, the better. The dashed line is pure guessing.
Where it was right and wrong
Every test patient falls into one of four boxes.
What it cannot do
Read these before trusting any single prediction.
- It makes mistakes. In testing it missed about 3 in 10 patients who really had diabetes, and raised a false alarm for about 1 in 4 who did not.
- The data is small and narrow. It covers 768 women aged 21 and over from one population, so results may not carry over to men or to other groups.
- It sees only eight numbers. Diet, activity, medication and a full medical history are all invisible to it.
- It is not a diagnosis. Only a blood test interpreted by a doctor can confirm diabetes.
Try it on a patient yourself
Enter measurements and see the estimate and what drove it.