EchoBridge: Long-Tail-Aware ECG--Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings

Abstract

Standardized echocardiography conclusions provide meaningful supervision for learning ECG representations of echocardiography-derived cardiac findings. Global ECG--text alignment may entangle modality-specific factors, while long-tailed finding distributions provide sparse positive supervision for low-prevalence conditions. We propose EchoBridge with Complementary Shared--Private Projection (CSPP) and Adaptive Prototype Boundary Calibration (APBC). CSPP maps each modality into shared and auxiliary private projections, reduces directional redundancy via within-modality orthogonality, and bidirectionally aligns normalized shared projections. APBC organizes the shared hypersphere with class-specific prototypes, training-frequency-adaptive angular margins, and spherical Riesz repulsion. We evaluate EchoBridge on EchoNext-Mini and independent PKUPH and SHTMU cohorts under four protocols: prompt-based inference without downstream classifier training, in-domain frozen linear probing, target-domain cross-center frozen linear probing, and source-only cross-center transfer, supplemented by finding-specific analyses. EchoBridge improves classifier-free AUROC, AUPRC, and F1 over the strongest baselines by 7.88, 5.61, and 4.54 points, respectively, and achieves the highest point estimates across all in-domain and target-domain probing budgets and both source-only transfer cohorts. Finding-specific analyses show gains for most conditions, including several low-prevalence valvular findings.

Method Highlights

• CSPP: maps ECG and text into shared and private projections, enforces within-modality orthogonality, and bidirectionally aligns normalized shared embeddings.
• APBC: calibrates class prototypes on the shared hypersphere with frequency-adaptive angular margins and spherical Riesz repulsion for long-tailed findings.

Evaluation Protocols

Protocol	Description
Prompt-based inference	Classifier-free matching with text prompts
In-domain frozen linear probing	Frozen ECG encoder + linear head on EchoNext
Target-domain cross-center probing	Frozen encoder, probe on external centers
Source-only cross-center transfer	Train on source, evaluate on PKUPH / SHTMU

Repository Structure

EchoBridge/
├── data/
│   ├── dataset.py              # ECG--text dataset loaders
│   ├── echonext_labels.json    # Prototype / label schema
│   └── README.md               # Data layout details
├── engine/
│   ├── encoders.py             # ECG backbone (ResNet18)
│   ├── model.py                # EchoBridge (CSPP + prototypes)
│   ├── losses.py               # Alignment / APBC losses
│   ├── train.py                # Pretraining entry
│   └── utils.py
├── linear_probing/
│   ├── train.py                # Frozen linear probing
│   └── test.py                 # Linear probing evaluation
├── zeroshot/
│   ├── prompts.json            # Finding prompts
│   └── zeroshot.py             # Prompt-based inference
└── outputs/                    # Checkpoints and eval reports

Requirements

• Python >= 3.9
• PyTorch (CUDA recommended)
• transformers
• numpy / tqdm / scikit-learn

Example install:

pip install torch torchvision transformers numpy tqdm scikit-learn

The text encoder defaults to ncbi/MedCPT-Article-Encoder.

Pretrained Weights

Pretrained checkpoints are available on Hugging Face: XiaochengFang/EchoBridge.

File	Description	Link
best_model.pth	EchoBridge alignment model (CSPP + APBC)	download
best_linear_head.pth	Frozen linear probing head (EchoNext, 100% budget)	download

Place the downloaded files as:

outputs/EchoNext/align/best_model.pth
outputs/EchoNext/linear_probing/best_linear_head.pth

Or download with Hugging Face CLI:

pip install -U huggingface_hub
huggingface-cli download XiaochengFang/EchoBridge best_model.pth \
  --local-dir outputs/EchoNext/align
huggingface-cli download XiaochengFang/EchoBridge best_linear_head.pth \
  --local-dir outputs/EchoNext/linear_probing

Data Preparation

EchoBridge expects paired ECG--echo-text samples plus a label CSV keyed by ECG_Id. Supported storage formats (NPZ / LMDB) and field definitions are documented in data/README.md.

Minimal checklist:

• train / val / test splits are ready
• Each sample has filename, ecg, echo
• Parsed ECG_Id matches the label CSV
• prototype_cols in data/echonext_labels.json match CSV column names

Pretraining

Train EchoBridge alignment (CSPP + APBC) on EchoNext:

cd EchoBridge

python engine/train.py \
  --data_path <ECHO_NEXT_DATA_ROOT> \
  --csv_path <ECHO_NEXT_LABEL_CSV> \
  --labels_json data/echonext_labels.json \
  --text_model ncbi/MedCPT-Article-Encoder \
  --proj_hidden 256 \
  --proj_out 256 \
  --batch_size 64 \
  --max_epochs 15 \
  --gpu 0 \
  --checkpoint_dir outputs/EchoNext/align

Best checkpoint is saved as outputs/EchoNext/align/best_model.pth.

Key loss weights:

Argument	Default	Role
--lambda_align	1.0	Shared-projection bidirectional alignment
--lambda_prototype	1.0	Prototype supervision (APBC)
--lambda_prototype_cons	0.5	ECG/text prototype consistency
--lambda2	0.1	Shared--private decoupling

Prompt-based Inference

python zeroshot/zeroshot.py \
  --npz_path <TEST_NPZ> \
  --labels_csv <ECHO_NEXT_LABEL_CSV> \
  --checkpoint outputs/EchoNext/align/best_model.pth \
  --prompt_json zeroshot/prompts.json \
  --labels_json data/echonext_labels.json \
  --gpu 0 \
  --out_report outputs/EchoNext/zeroshot/zeroshot_eval.txt

Set --bootstrap_iterations 0 to disable 95% CI estimation.

Frozen Linear Probing

Train a linear head on frozen ECG features:

python linear_probing/train.py \
  --data_path <ECHO_NEXT_DATA_ROOT> \
  --labels_csv <ECHO_NEXT_LABEL_CSV> \
  --align_checkpoint outputs/EchoNext/align/best_model.pth \
  --subset_ratio 1.0 \
  --max_epochs 30 \
  --gpu 0 \
  --output_dir outputs/EchoNext/linear_probing

Evaluate:

python linear_probing/test.py \
  --data_path <ECHO_NEXT_DATA_ROOT> \
  --labels_csv <ECHO_NEXT_LABEL_CSV> \
  --align_checkpoint outputs/EchoNext/align/best_model.pth \
  --head_checkpoint outputs/EchoNext/linear_probing/best_linear_head.pth \
  --gpu 0 \
  --output_file outputs/EchoNext/linear_probing/linear_probing_eval.txt

Use --subset_ratio 0.01 / 0.1 / 1.0 for low-shot probing budgets.

Citation

If you find this repository useful, please cite:

@article{echobridge2026,
  title={EchoBridge: Long-Tail-Aware ECG--Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings},
  author={},
  journal={},
  year={2026}
}

(Paper / camera-ready citation will be updated upon release.)

License

Code in this repository is provided for research purposes. Please follow the licenses of EchoNext / PKUPH / SHTMU data and of third-party models (e.g., MedCPT).

Acknowledgement

We thank the EchoNext community and related ECG--text alignment baselines that inspired this codebase.