Thai-CGT · research preview

β-Thalassemia · audit

Guardrail integrity & decision trail

Governance

Synthetic research-preview demonstrator — IS_SYNTHETIC and NO_PHI hold. Not pre-registered; not externally validated. Every prediction carries researchPreview + uncertainty:"illustrative"; the no-overclaim guard (G08/G14) and the G-CONF confidential-payload gate run at build and render time.

Local infrastructure

arm64 · Mac Studio

App: TanStack Start (Nitro node-server) · Docker
Backend: Prisma 7 + PostgreSQL (read contract + Decision/PiSignoff/AuditEvent)
Model service: ml/bthal · FastAPI (XGBoost, arm64)
Reproduce: make -C ml/bthal all

Reporting compliance

TRIPOD+AI checklist · PROBAST self-audit

Research preview

TRIPOD+AI checklist

22/30 covered · 5 partial · 3 not yet

1a	Title — identifies study as developing/validating a prediction model	✓ covered
1b	Title — names target population and outcome	✓ covered
2	Abstract — structured summary of objectives, methods, results	✓ covered
3a	Background — clinical context and rationale	✓ covered
3b	Objectives — pre-specified	✓ covered
4a	Source of data — registries, trials, EHR linkage	✓ covered
4b	Data eligibility dates and follow-up window	✓ covered
5a	Participants — setting and recruitment	✓ covered
5b	Participants — eligibility criteria	✓ covered
5c	Participants — treatments received	✓ covered
6a	Outcome — definition and assessors	✓ covered
6b	Outcome — blinding of outcome assessment	⚠ partial
7a	Predictors — definitions and measurement	✓ covered
7b	Predictors — blinding for outcome at measurement	⚠ partial
8	Sample size — justification	⚠ partial
9	Missing data — handling strategy	⚠ partial
10a	Statistical analysis — predictor selection	✓ covered
10b	Statistical analysis — model type and assumptions	✓ covered
10c	Statistical analysis — calibration and discrimination	✓ covered
10d	AI/ML specifics — hyperparameter search and seeds	✓ covered
10e	AI/ML specifics — fairness/subgroup performance	⚠ partial
11	Risk groups — thresholds and rationale	✓ covered
12	Development vs validation — split rationale	✓ covered
13a	Participants — flow diagram	✓ covered
13b	Characteristics — baseline table	✓ covered
14	Model specification — equations / weights released	✓ covered
15	Model performance — discrimination, calibration, DCA	✓ covered
18	Limitations — bias, generalisability	○ not yet
20	Interpretation — clinical implications	○ not yet
22	Supplementary materials — code, data, model artifacts	○ not yet

Collins, Moons, Dhiman et al. BMJ 2024;385:q824. DOI 10.1136/bmj-2023-078378

PROBAST risk-of-bias self-audit

4 domains · 20 signaling questions

Participants

low ROB

2 signaling questions

Multi-source registries + trial cohort.

Predictors

low ROB

3 signaling questions

Pre-specified, blinded to outcome at measurement.

Outcome

unclear

6 signaling questions

EBMT 2023 refined VOD criteria new; adjudication still maturing.

Analysis

high ROB

9 signaling questions

Fragile: events-per-candidate-predictor margin thin (§E.2); missing-data handling under review.

§E.2 fragile spots: sample size vs candidate-predictor count · missing-data handling

Wolff et al. 2019. Open-access at probast.org

Self-audit refreshed 2026-05-28 by Anirach M.

Decision-capture trail · 0 entries

No decisions captured yet. Record MDT agreement/override on /bthal/decisions and it appears here (and in the Postgres audit log when running with a database).

Model-lock integrity

bthal · manifest sha256 —