Waveform Perception (Signal Models)
Category: AI & ML
Node id: waveform-perception
Maturity: beta
Risk tier: B
Standards: HIPAA, GDPR, IEC 62304
Runs ECG waveforms through signal foundation models (ECGFounder, ECG-JEPA) via tietai-signal-service: embeddings and 150-label screening scores. Scores are screening output, never diagnoses.
What it does
- ECGFounder 12-lead / 1-lead: 1024-d embeddings + full 150-label screening head
- ECG-JEPA (multiblock / random): 768-d pooled embeddings
- Lead selection by name — send a 12-lead strip to an 8-lead model unchanged
- Fails closed when a classify response lacks the screening-framing marker
Use cases
- Screen MCT/Holter strips for 150 rhythm and morphology labels
- Produce cardiac embeddings for similarity search and probe training
- Feed signal_findings to reasoner nodes under the authoritative-payload contract
Runtime
- Performance tier:
variable
Input ports
| Name | Type | Required | Description |
|---|---|---|---|
strips | json | No | Array of strips from ecg-strips-source (each with channels, leads, sample_rate_hz, gain_uv_per_lsb). Preferred input. |
waveform | json | No | Single waveform as a 2-D array [leads][samples] in mV (or raw counts with gain_uv_per_lsb). Used when no strips input is wired. |
sample_rate_hz | number | No | Sample rate of the waveform input (required with waveform). |
leads | json | No | Lead names for each waveform row, standard 12-lead order assumed when omitted. |
gain_uv_per_lsb | number | No | µV per LSB when waveform carries raw ADC counts instead of mV. |
Output ports
| Name | Type | Required | Description |
|---|---|---|---|
findings | json | No | One finding per waveform: model provenance (id + weights sha256), embedding or ranked screening scores with framing, quality block, timing. |
success | boolean | No | True when every waveform produced a finding. |
Settings
Basic
| Setting | Name | Type | Required | Default | Allowed values | Description |
|---|---|---|---|---|---|---|
| Model ID | model | string | Yes | Signal model id from the service's /v1/models (e.g. ecgfounder-12lead, ecgfounder-1lead, ecg-jepa-multiblock). | ||
| Operation | operation | enum | Yes | classify | classify, embed | classify: ranked screening scores (ECGFounder models only). embed: fixed-dimension embedding. |
| Top K | top_k | number | No | 0 | Keep only the K highest scores (classify). 0 returns all 150 labels. | |
| Score Threshold | threshold | number | No | 0 | Drop scores below this value (classify). 0 keeps everything. |
Advanced
| Setting | Name | Type | Required | Default | Allowed values | Description |
|---|---|---|---|---|---|---|
| Timeout (s) | timeout_seconds | number | No | 120 | Per-request timeout. Must cover the service's documented cold start when min-instances is 0. | |
| Service URL | service_url | string | No | Override the signal service base URL (default from SIGNAL_SERVICE_URL). |
Errors
| Code | Description |
|---|---|
SIGNAL_CONFIG | Missing model/operation config or no waveform input wired |
SIGNAL_REQUEST | The signal service was unreachable or returned an error |
SIGNAL_FRAMING | A classify response lacked the screening-framing marker and was refused |
Runtime notes
- Requires SIGNAL_SERVICE_URL (or the service_url config). Auth: SIGNAL_SERVICE_TOKEN for dev, GCE metadata OIDC identity token in production (Cloud Run IAM perimeter).
- Scores are screening output with documented operating points, not diagnoses; keep the verifier chain downstream (RiskTier C applies when findings reach patient-facing content).
- Wire strips from ecg-strips-source for the PHI-clean path; inline waveform input is intended for research graphs.
See Build any workflow for how to put nodes together and Ports, types and data flow for how port types connect.
Something unclear? Send documentation feedback by email.