Dataset Preprocessing README
🔥 June 2, 2025 Update for TUAB and TUEV
The previous preprocessing code for the TUAB and TUEV datasets was inherited from the BIOT and LaBraM repositories. These original implementations included random elements in the data splitting process. Even with fixed random seeds, different hardware environments could lead to inconsistent Train/Val/Test splits. This issue has been carried forward into CBraMod.
In the performance comparison presented in the CBraMod paper, I directly cited the results reported in the BIOT and LaBraM papers without having access to their exact dataset splits. Therefore, I cannot guarantee that the comparisons were made using the same dataset partitions. As a result, the evaluation may not be entirely fair.
Moreover, others may also be unable to reproduce a fair comparison with CBraMod on TUAB and TUEV under the same conditions.
To fully address this issue, I have updated the preprocessing code for TUAB and TUEV to ensure fixed, deterministic dataset splits. If you are conducting experiments on these two datasets, please use the latest version of the preprocessing code to generate the splits.
For accurate and fair comparisons, it is strongly recommended to re-implement existing methods such as BIOT, LaBraM, and CBraMod on the same fixed splits.
⚠️ Please note: The TUAB version used in our experiments is 3.0.1, and the TUEV version is 2.0.0. Updates to the datasets may result in changes to the total number of samples.
📌 If you are using different versions of these datasets, do not refer to our sample counts. Instead, reproduce the results directly on your own data splits.
✅ We also provide dataset splits for TUEV v2.0.1. Please refer to the sample counts below for details.
📊 Current Sample Counts (Updated Preprocessing)
TUAB (v3.0.1):
- Train: 297,103
- Validation: 75,407
- Test: 36,945
- Total: 409,455
TUEV (v2.0.0):
- Train: 67,436
- Validation: 15,634
- Test: 29,421
- Total: 112,491
TUEV (v2.0.1):
- Train: 68,445
- Validation: 15,487
- Test: 29,421
- Total: 113,353