TAAD-ECG Model Card
Model Description
TAAD-ECG is a research model for identifying an ECG warning signal associated with acute type A aortic dissection among adults evaluated for acute chest pain. The model combines a 12-lead ECG backbone, age and sex, and a text-query network. Its weights accompany the manuscript "An AI-enabled electrocardiographic warning signal for acute type A aortic dissection at chest-pain triage."
The public bundle contains the ECG backbone, text-query network, and a fixed query representation. It does not contain source clinical data or the runtime text encoder.
Intended Use
The model is released for reproducibility, method development, and external research validation using appropriately governed, de-identified data.
It is not a medical device. It must not be used to diagnose or exclude aortic dissection, to determine imaging or treatment, or for autonomous clinical decision-making. Performance can change with ECG systems, PDF layouts, populations, disease prevalence, and care settings.
Input
- Model tensor: 12 leads x 1000 samples at 100 Hz.
- Short leads occupy samples 0 to 249; the lead-II rhythm strip occupies samples 0 to 999.
- Age in years.
- Sex encoded as female 0 and male 1, consistent with model development.
The companion PDF parser supports only the vector 4 x 3 plus rhythm-strip layout validated in the study.
Output
The scalar output lies between 0 and 1 but is not a calibrated disease probability. The manuscript's fixed validation-derived tiers are:
- Low model-score tier: score below 0.0462.
- Intermediate model-score tier: score from 0.0462 to below 0.4275.
- High model-score tier: score at or above 0.4275.
Limitations
- A low score cannot exclude disease for an individual patient.
- The model was evaluated in participating hospitals in China; independent validation in other populations and health systems is required.
- PDF extraction depends on vector structure and vendor layout.
- The attention output is descriptive and should not be interpreted as a causal physiological explanation.
- The public demo has not been evaluated as a clinical intervention and does not establish effects on imaging time, treatment, surgery, or outcomes.
Privacy
No patient-level data are distributed. Do not upload identifiable clinical records to public services.
License
The model weights are released under CC BY-NC 4.0 for non-commercial research use. The source code is Apache-2.0 licensed.
Citation
The accompanying manuscript is under submission. Citation metadata will be updated after publication.
