UnivaBio 2026 · AI for Human Health
About SIGAP
Indonesia has the world's second-highest tuberculosis burden: about 10% of global cases and an estimated 200,000 people each year who transmit TB without ever being detected. The root cause is latency — the care loop moves slower than the disease.
How it works
- 01
Chest X-rays are scored by a self-supervised chest X-ray representation (RAD-DINO) with a linear probe, trained on audited public data and evaluated on unseen hospitals.
- 02
Scores are calibrated (temperature scaling) and mapped to three bands: priority referral, uncertain (human re-read) and negative — with thresholds from a sensitivity-target operating point.
- 03
The screening result becomes the starting risk state on a patient record; daily check-ins update it and escalate when the patient worsens.
Rigor you can check
Leakage audit: perceptual hashing removed 338 internal duplicates and 19 training images that leaked into the external test set before any training ran.
External evaluation across two unseen hospitals with bootstrap AUROC intervals, sensitivity at fixed specificity, calibration (ECE) and age/sex subgroups.
Limitations are reported, not hidden: public radiographic labels, aggressive re-processing may evade de-duplication, and no Indonesian data yet.
Code, data and live demo
Frozen split manifest, evaluation scripts and model artifacts are public. The live demo runs the real ensemble.
Honest statement: this is a research prototype built for a competition — not a medical device. Do not use it for clinical decisions without validation and oversight by health professionals.