UNIVABIO 2026 / AI FOR HUMAN HEALTH
Understand the denominator.
An interactive field guide to screening uncertainty. Explore fictional populations, compare experiments, inspect the assumptions behind repeat testing, predict before revealing the counts, and explain the result to an on-device AI coach.
Working prototype demonstration
109.4 seconds of the final application, with the entrant’s edited supplied narration. Actual browser interactions are shown as edited screen-state captures, with readable holds and cuts rather than a continuous recording. Staged estimates are demonstration inputs, not learner-study data. The model results and privacy scope are qualified on screen. No synthesized voice or reconstructed words are used.
Open or download the finished MP4 ↗ · Demo methods and claim qualifications ↗
What you can investigate
- Keep detection and clearance fixed while changing prevalence. Pin experiment A, change B, and compare their positive groups and a continuous prevalence curve.
- Switch from population squares to a natural-frequency tree. The table and tree retain raw expected counts; only squares are rounded.
- Explore a second test among first-positive people. Choose independent tests, perfectly repeated results, or explicit conditional rates. A repeated result can add no information.
- Complete three predict/reveal/denominator cases and export a numeric learning notebook. Correct a wrong denominator before proceeding. The notebook retains first/final choices and first numeric guesses for this tab only; learner explanations are never exported.
- Inspect influential model features, label alternatives and uncalibrated votes. Override the suggested topic, answer a factual check, and revise your explanation. Votes are model diagnostics, not health probabilities or proof of reasoning quality.
Verification and limitations
- 336 arithmetic and population-rounding combinations; 2,525 prevalence-curve points; 405 conditional repeat-test combinations; five test suites, practice/export checks and a deterministic model audit.
- 100 synthetic English training explanations. On a separate 30-example synthetic holdout: 29 correct raw classifications; uncertainty checks accepted 28 labels, of which 27 were correct.
- Known error: one correct distinction between sensitivity and predictive value was misread. A 24-case development stress audit found further semantic failures. Feedback is now provisional; classifications and original accuracy are unchanged. Negation, mixed ideas, unfamiliar wording and other languages can fail. Votes are uncalibrated.
- No independent learner study, clinical validation or measured health benefit. All scenarios are fictional; no personal diagnosis, risk estimate or treatment advice.
Read the model card and audit ↗ · Competition and originality assessment ↗
Privacy and media
No patient data, analytics, remote inference, external scripts or fonts. Coach text stays in tab memory, without persistent app storage or server transmission. Exports save numeric answers and selected choices to a file and exclude explanations. Loading app/model files and this page’s video metadata/media still contacts GitHub Pages; hosting providers may retain ordinary request logs. Optional connected browser assistants have separate data handling. The entrant supplied the narration recording and must understand and explain the work to judges.
How the app and classifier work ↗ · Data handling ↗ · Attribution ↗
Reproduce the work
The repository includes the complete static app, synthetic data, standard-library Python trainer, evaluation details and Node verification checks. The source ZIP runs locally without app dependencies. The one-page description states what was built, what was verified and what remains unvalidated.