jialucode/FSL_ECG_QA_Dataset

FSL ECG QA Dataset

FSL ECG QA Dataset is a benchmark dataset used in paper ["Electrocardiogram–Language Model for Few-Shot Question Answering with Meta Learning"] (https://arxiv.org/abs/2410.14464). It supports research in combining electrocardiogram (ECG) signals with natural language question answering (QA), particularly in few-shot and meta-learning scenarios.

Dataset Highlights

• 🧠 Task Diversification: Restructured ECG-QA tasks promote rapid few-shot adaptation.
• 🧬 Fusion Mapping: A lightweight multimodal mapper bridges ECG and language features.
• 🚀 Model Generalization: LLM-agnostic design ensures broad transferability and robustness.

1. Datasets

1.1 Load the classes

python load_class.py --base_path /your/actual/path/to/ecgqa/ptbxl/paraphrased --test_dataset ptb-xl
python load_class.py --paraphrased_path /your/actual/path/to/ecgqa/mimic-iv-ecg/paraphrased --test_dataset mimic
(all tested mimic-iv-ecg dataset is listed in "data/processed_test_30k.json")

1.2 Load ECG QA FSL dataset

python data_loader.py
--model_name /your/actual/path/to/model/download/from/hugging/face
--paraphrased_path /your/actual/path/to/ecgqa/mimic-iv-ecg/paraphrased
--test_dataset ptb-xl

1.3 Sample of ECG QA FSL dataset

episode = {
    # --- Support Set ---
    "support_x": [  # ECG feature tensors or preprocessed arrays
        ecg_sample_1,  # typically a NumPy array or tensor
        ecg_sample_2,
        # ...
    ],
    "support_y_q": [  # Question token sequences (padded)
        [12, 45, 78, 0, 0, 0],
        [23, 67, 89, 90, 0, 0],
        # ...
    ],
    "support_y_a": [  # Answer token sequences (padded)
        [1, 0, 0, 0, 0, 0],
        [1, 1, 0, 0, 0, 0],
        # ...
    ],
    "support_y_q_mask": [  # Mask for question tokens (1 = valid, 0 = pad)
        [1, 1, 1, 0, 0, 0],
        [1, 1, 1, 1, 0, 0],
        # ...
    ],
    "support_y_a_mask": [  # Mask for answer tokens
        [1, 0, 0, 0, 0, 0],
        [1, 1, 0, 0, 0, 0],
        # ...
    ],
    "flatten_support_x": [  # File paths to raw ECG signals
        "/path/to/support/ecg_1.npy",
        "/path/to/support/ecg_2.npy",
        # ...
    ],

    # --- Query Set ---
    "query_x": [
        ecg_query_1,
        ecg_query_2,
        # ...
    ],
    "query_y_q": [
        [34, 78, 56, 0, 0, 0],
        [90, 12, 45, 76, 0, 0],
        # ...
    ],
    "query_y_a": [
        [0, 1, 1, 0, 0, 0],
        [1, 0, 0, 0, 0, 0],
        # ...
    ],
    "query_y_q_mask": [
        [1, 1, 1, 0, 0, 0],
        [1, 1, 1, 1, 0, 0],
        # ...
    ],
    "query_y_a_mask": [
        [1, 1, 1, 0, 0, 0],
        [1, 0, 0, 0, 0, 0],
        # ...
    ],
    "flatten_query_x": [
        "/path/to/query/ecg_1.npy",
        "/path/to/query/ecg_2.npy",
        # ...
    ],
}

1.4 Class distribution

Question Type	Attributes	Answers	Classes (train:test)	Samples	Example
Single-Verify	94	yes/no	156 (124:32)	34,105	Q: Does this ECG show 1st degree av block? A: yes/no
Single-Choose	165	both/none/attr_1/attr_2	262 (209:53)	47,655	Q: Which noise does this ECG show, baseline drift or static noise? A: baseline drift /static noise
Single-Query	30	attr_1/attr_2/.../attr_n	260 (208:52)	63,125	Q: What direction is this ECG deviated to? A: Normal axis/...
All	206	yes/no/both/none/.../attr_n	678 (541:137)	144,885	...

2. Citation

@article{tang2024electrocardiogram,
  title={Electrocardiogram-Language Model for Few-Shot Question Answering with Meta Learning},
  author={Tang, Jialu and Xia, Tong and Lu, Yuan and Mascolo, Cecilia and Saeed, Aaqib},
  journal={arXiv preprint arXiv:2410.14464},
  year={2024}
}