Honest limits, stated plainly
This tool exists because early diabetes is quiet: millions of people carry it without dramatic symptoms, and the ones they do notice frequently go unconnected to each other. A cheap, instant pattern check that can nudge someone toward a simple blood test fills a real gap. But it has limits, and you deserve them spelled out.
Not a diagnosis. A score is a prompt to get tested. Only labs and a clinician can determine your status.
Small dataset. 520 rows that collapse to 251 unique profiles (exact copies removed before training), 61.5% positive on rows. Good for learning symptom patterns; not enough for certainty in edge cases.
Self-reported symptoms. Training data recorded what people said they felt — subjective and occasionally inconsistent with clinical reality.
Zero-symptom blind spot. 53 rows have no marked symptoms: 6 positive, all women — and five of those six are the same record (a 35-year-old woman, repeated). All 44 men with no symptoms are negative. A woman who marks no symptoms can score high and a man will score low, but both scores lean on a thin, skewed slice. The result card says so on the page when it happens. Treat any no-symptom score as unreliable, either direction.
No blood values. The model sees symptoms only — no glucose, HbA1c, or BMI entered in the training data. A real clinical risk model would combine both.
Not emergency guidance. If you have severe vomiting, deep rapid breathing, fruity-smelling breath or confusion, seek urgent care now — those can signal diabetic ketoacidosis, which no screening tool should triage.
Gender asymmetry is in the data. 90.1% of female records are positive vs 44.8% of male. The model reflects that faithfully — sex moves every score, and the result card names it when it matters. This questionnaire cannot tell you your personal risk; it can only place you against the people who filled in this hospital form.
No follow-up data. The dataset doesn't track who later developed diabetes, so this is about current symptom patterns — not a prediction of your future.
Why this is still worth doing
Despite all the above, the combination — cross-validated on real records, calibrated scores, instant and private, and explicit about failure modes on the page itself — is what a careful screening prompt should look like. The result is designed to get the right people to the right test, not to reassure or alarm anyone.