UnivaBio 2026 · AI for Human Health
Screen faster than the disease spreads.
Indonesia carries 10% of the world's TB burden, and about 200,000 people each year transmit TB before ever being detected. SIGAP compresses the care loop with one calibrated risk state — and says “uncertain” when it should.
Priority referral
p(TB) estimate
99%
Send the patient for GeneXpert testing at the nearest facility today. Record them as a presumptive TB case.
Illustrative result — real model output on a public dataset case.
0.887
AUROC on unseen hospitals
19
training images removed after our leakage audit
0.71
sensitivity at 90% specificity (external)
Evidence, reported honestly
Measured & open
- External validation on two unseen hospitals (800 images), with bootstrap intervals, calibration (ECE) and age/sex subgroups.
- Data audit before training: 338 internal duplicates and 19 training images that leaked into the external test set were removed. The frozen split manifest is published.
- Representation study across three model families; the winner (RAD-DINO self-supervised probe) is open in the repo and as downloadable artifacts.
Not done yet — honestly
- No prospective validation in Indonesia yet; this is a research prototype, not a medical device.
- Triage thresholds come from public data, not from puskesmas populations; site calibration is planned.
- Cough audio analysis and SITB integration are not implemented.
Under the hood
RAD-DINO (self-supervised chest X-ray) + linear probe, temperature scaling, cross-family disagreement deferral. Served with FastAPI and a bilingual Next.js UI; trained on Kaggle GPUs via API.
Try it now
Both links run the real ensemble: upload a chest X-ray and get a triage band, an attention map and a referral letter.