Model Catalogue

Four frozen models share one input and output layer. Bundle verification runs before prediction so a corrupted or substituted artifact is rejected before deserialization.

Identifier	Leads	Model family	Classes	Artifact
12lead-conformer	12	Lead-wise shared Conformer	acceptable / unacceptable	best_model.pt
12lead-rbfsvm	12	RBF-SVM over profile-defined SQIs	acceptable / unacceptable	model.joblib
singlelead-conformer	1	GM-mechanism Conformer	good / medium / bad	ckpt_best.pt
singlelead-rbfsvm	1	RBF-SVM over profile-defined SQIs	good / medium / bad	model.joblib

Artifact integrity

Model	SHA-256
12-lead Conformer	f08e97226ed0bed5cadcced470708572b32dcc0894e5d842a59866da307a3654
Single-lead Conformer	17d7dc331de40862943d3f04b372719cdc53194ee0d10d41fc3e239239dc1c7a
12-lead RBF-SVM	3d739eac5d378d08c0a47e913979a5f127618fe6a0a868a3095d2cb6a69031c2
Single-lead RBF-SVM	906ed289c203cf6fab9f85ec4eee0f99f8d07d5f65b8f05e3aa9fc38b1f6e7ee

Profiles are hashed separately in pretrained/inference/manifest.json because they define feature order, thresholds, class order, and normalisation context.

Common input contract

All models receive complete 10-second windows at 125 Hz: 1,250 samples by the required number of leads. The public loader accepts .npy, .npz, numeric .csv, and WFDB records and converts them to samples-by-leads before segmentation.

Output schemas

Binary models return:

Column	Meaning
raw_class	acceptable or unacceptable
display_class	usable or unusable
prob_unacceptable	Poor-quality probability
prob_acceptable	Acceptable-quality probability

Three-class models return raw_class, display_class, prob_good, prob_medium, and prob_bad.

The record runner adds record identifiers, segment indices, start/end time, model name, and input path. See Inference and Docker.

Intended use

These are research artifacts for reproducible ECG quality experiments. They are not medical devices and must not be treated as independent evidence for clinical diagnosis or deployment readiness.