Trustcat/queenbee-ecg-transformer

🐝 QueenBee ECG-Transformer

Foundation model for 12-lead ECG diagnostic classification

Part of the TrustCat sovereign medical AI stack - no cloud, no compromise.

Model Description

A transformer-based architecture for multi-label ECG classification trained on PTB-XL, the largest freely accessible clinical ECG dataset (21,799 records).

Architecture Highlights

• Patch Embedding: 100ms temporal patches (50 samples @ 500Hz)
• Lead Embeddings: Learnable embeddings for each of 12 leads
• Rotary Position Embeddings: Better temporal awareness than sinusoidal
• Multi-head Self-Attention: 8 heads × 6 layers
• Multi-task Heads: 5 superclasses + 44 SCP diagnostic codes

Input: [B, 12, 5000] (12-lead ECG, 10 seconds @ 500Hz)
  ↓
Patch Embedding (50 samples per patch)
  ↓
Lead Embeddings + Position Embeddings
  ↓
Transformer Encoder (6 layers, 8 heads, dim=256)
  ↓
[CLS] Token Pooling
  ↓
├── Superclass Head → [B, 5]  (NORM, MI, STTC, CD, HYP)
└── SCP Code Head   → [B, 44] (Detailed diagnostic codes)

Performance

Superclass Classification (5 classes)

Class	Description	AUC
NORM	Normal ECG	92.8%
MI	Myocardial Infarction	90.1%
STTC	ST/T Change	91.6%
CD	Conduction Disturbance	90.2%
HYP	Hypertrophy	81.0%
Mean	All Classes	89.1%

Overall Metrics

Metric	Value
Superclass Mean AUC	89.1%
Superclass F1	64.6%
SCP Code Mean AUC	84.9%

Training Details

Parameter	Value
Dataset	PTB-XL v1.0.3
Train/Val/Test Split	17,084 / 2,146 / 2,158
Epochs	50 (best @ epoch 10)
Optimizer	AdamW
Learning Rate	1e-4 (cosine decay)
Batch Size	64
Hardware	2× NVIDIA RTX 5090
Training Time	~18 minutes

Usage

import torch
import wfdb
from scipy import signal as scipy_signal
from model import ECGTransformer

# Load model
model = ECGTransformer(
    num_leads=12,
    signal_length=5000,
    patch_size=50,
    embed_dim=256,
    depth=6,
    num_heads=8,
    num_superclasses=5,
    num_scp_codes=44
)

checkpoint = torch.load("ecg_transformer_best.pt", map_location="cpu")
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Load and preprocess ECG (example with WFDB)
signal, meta = wfdb.rdsamp("path/to/ecg_record")

# Resample to 500Hz if needed
if meta['fs'] != 500:
    signal = scipy_signal.resample(signal, int(len(signal) * 500 / meta['fs']), axis=0)

# Normalize per-lead
signal = (signal - signal.mean(axis=0)) / (signal.std(axis=0) + 1e-8)

# Pad/truncate to 5000 samples (10 seconds)
if len(signal) < 5000:
    signal = np.pad(signal, ((0, 5000 - len(signal)), (0, 0)))
else:
    signal = signal[:5000]

# Inference
x = torch.tensor(signal.T, dtype=torch.float32).unsqueeze(0)
with torch.no_grad():
    superclass_logits, scp_logits = model(x)
    probs = torch.sigmoid(superclass_logits)[0]

classes = ["NORM", "MI", "STTC", "CD", "HYP"]
for cls, prob in zip(classes, probs):
    print(f"{cls}: {prob:.1%}")

Superclass Definitions

Code	Full Name	Clinical Significance
NORM	Normal ECG	No significant abnormalities
MI	Myocardial Infarction	Heart attack (current or prior)
STTC	ST/T Change	Ischemia, electrolyte abnormalities
CD	Conduction Disturbance	Bundle branch blocks, AV blocks
HYP	Hypertrophy	Ventricular/atrial enlargement

Intended Use

• Primary: Clinical decision support for ECG interpretation
• Secondary: ECG screening and triage
• Research: Cardiology AI research and benchmarking

Limitations

⚠️ Not FDA Cleared - For research use only

• Trained on PTB-XL dataset (German population, 1989-1996)
• HYP class has lower performance (smaller training set)
• Requires 12-lead ECG input (not suitable for single-lead devices)
• Should be validated on target population before clinical use

Citation

If you use this model, please cite:

@misc{trustcat2026queenbee,
  title={QueenBee ECG-Transformer: Foundation Model for 12-Lead ECG Analysis},
  author={TrustCat},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Trustcat/queenbee-ecg-transformer}
}

License

Apache 2.0

---

Built with diamond hands by TrustCat 🐝

Sovereign Medical AI - No Cloud, No Compromise