BoneBot

🏆 Winner, Hack-Nation's 6th Global AI Hackathon, London hub.

BoneBot is a hormone-aware bone-health screening assistant for postmenopausal women. It turns a woman's health profile (menopause history, risk factors, and simple labs) into an estimated T-score with an uncertainty range, the factors driving it, and a clear prompt to confirm with a DXA scan when warranted. It is a screening flag, never a diagnosis: an NHANES-trained model makes every prediction, and the LLM only explains it in plain language.

🔴 Live demo: https://bonebot.vercel.app/ No login needed.

Built in 24 hours for the Hack-Nation 6th Global AI Hackathon: Challenge 05, Women's Hormonal Health.

Why

1 in 2 postmenopausal women will suffer an osteoporotic fracture, and most are never screened until they break a bone. Estrogen loss after menopause accelerates bone loss, but existing risk tools are static questionnaires. BoneBot flags who needs a scan earlier, from data women already have.

How it works

1. Triage gate: a lightweight model (age, BMI, postmenopausal status) decides whether the fuller questionnaire is worth the user's time. Its threshold is set at a 1% predicted-risk cutoff, chosen for safety (validation sensitivity ≥ 99%); see docs/TRIAGE_THRESHOLD_AUDIT.md for the held-out audit.
2. T-score estimate: a fuller model (age, BMI, years since menopause, activity, prior fragility fracture, glucocorticoids, smoking, alcohol, vitamin D, calcium, rheumatoid arthritis, and hormone therapy) produces an estimated T-score with an uncertainty range and per-factor contributions in T-score units. Activity can be entered as seven-day average steps and active minutes or confirmed from an activity-app screenshot; these are disclosed proxies for the wrist-movement measure used during training.
3. Explanation: the LLM (via the Vercel AI SDK) explains the model's output. It never sets or adjusts the risk itself. Ambiguous answers are clarified once, then treated as unknown, never as a positive risk factor, and flagged on the results page.

Model training and the threshold audit live in model/; canonical input definitions are in docs/INPUT_SPEC.md; the clinical evidence constraining the explanations is in docs/EVIDENCE.md.

Model performance

Trained and evaluated on 1,119 postmenopausal women in NHANES 2013–2014 with femur DXA and wrist-accelerometry (activity averaged over valid wear days only). Split is a held-out test set the model never saw during fitting; the DXA T-score is the label only and never a feature (anti-leakage rule).

T-score model (Ridge regression, 13 features → estimated femoral-neck T-score):

Metric	Value
Mean absolute error (held-out)	0.727 T-score units
R² (held-out)	0.276
Train / held-out test	788 / 280
95% interval empirical coverage under ~50% missing inputs	0.946 (target 0.95)

The uncertainty range shown to the user is a genuine 95% prediction interval (complete-data half-width ±1.85 T-score units) that widens automatically when inputs are imputed, so a sparser profile visibly shows less confidence.

Does the extra signal help? Against an age + BMI baseline that stands in for today's simple tools, on the same held-out split (osteoporosis classification):

Model	AUC	PR-AUC
Baseline (age + BMI) ≈ current tools	0.772	0.280
OST reference tool (age + weight)	0.779	n/a
Full model (+ menopause, wearable activity, risk factors, labs)	0.789	0.350

Triage gate (age + BMI + postmenopausal status): AUC 0.867. The referral threshold is set at a 1% predicted-risk cutoff, chosen so validation sensitivity stays ≥ 99%, the safety priority for a screen (missing a case is worse than an extra questionnaire). Full held-out audit in docs/TRIAGE_THRESHOLD_AUDIT.md.

The benchmark task, split, and reproduction steps are documented in model/README.md.

Limitations & responsible use

BoneBot outputs an estimate, not a diagnosis. Only a DXA scan measures bone density. The model is trained on NHANES data and its applicability outside that population is untested. Uncertain inputs are surfaced, not hidden, and users are directed to discuss results with a clinician.

Development

npm run dev      # local dev server
npm run build    # must pass before pushing; main auto-deploys
npm run lint

Stack: Next.js 16 (App Router, TypeScript), Tailwind, Vercel AI SDK, Vercel hosting.

Team

Built in 24 hours by a team of four at the London hub:

• Emre Yavuz (GitHub)
• Francesco Coccia (GitHub)
• Paula Juan Royo (GitHub)
• Macropodux (GitHub)

License

Code is released under the MIT License. The benchmark, evidence documents, and documentation (docs/) are released under CC BY 4.0. See LICENSE for details.