Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation
Authors: Yanlong Chen, Mattia Orlandi, Pierangelo Rapa, Simone Benatti, Luca Benini, Yawei Li
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations across various tasks and datasets within each physiological modality demonstrate that Physio Wave consistently achieves state-of-the-art performance, underscoring its broad applicability and robustness (See Section 2.4). |
| Researcher Affiliation | Academia | 1IIS, ETH Zurich 2DEI, University of Bologna 3DIEF, University of Modena and Reggio Emilia |
| Pseudocode | Yes | Algorithm 1 Frequency-guided masking |
| Open Source Code | Yes | Code and data are available at: github.com/Forever Blue816/Physio Wave |
| Open Datasets | Yes | Leveraging the Physio Wave architecture, two large-scale, modality-specific foundation models, Physio Wave-ecg and Physio Wave-emg, are pretrained on the most extensive open-access corpora currently available for their respective signal types (see Tables 15 and 16). Physio Wave-ecg is trained on approximately 182 GB of twelve-lead ECG recordings, while Physio Wave-emg utilizes about 823 GB of EMG data. For each modality, we provide three parameter configurations: Small (5M), Base (15M), and Large (37M). Both Physio Wave models share the same Transformer backbone; however, their learnable wavelet front-ends are modality-aware (see Tables 12 and 13 in Appendix C for full architectural and training details). The pretrained encoders are evaluated on the datasets listed in Table 17. All downstream experiments follow the singleand multi-modal procedures detailed in Section 2.4. Each benchmark is split by subject into 6:2:2 train/validation/test partitions to prevent subject leakage. |
| Dataset Splits | Yes | Each benchmark is split by subject into 6:2:2 train/validation/test partitions to prevent subject leakage. |
| Hardware Specification | Yes | Pretraining lasts for 50 epochs with a global batch size of 64 on 16 NVIDIA A100 GPUs, whereas all downstream experiments are carried out on 4 A100 GPUs. |
| Software Dependencies | No | Low pass and high pass filter taps hlo, hhi RK0, where K0 is the original wavelet filter length are extracted from a chosen discrete wavelet using Py Wavelets [24]. |
| Experiment Setup | Yes | C HYPERPARAMETER SETTINGS We employ Adam W optimizer with a weight decay of 0.01 and moment coefficients β1 = 0.9 and β2 = 0.98. The learning rate is linearly warmed up from 5 10 7 to 5 10 5 over the first ten epochs and then follows a cosine decay to a floor of 1 10 6. Pretraining lasts for 50 epochs with a global batch size of 64 on 16 NVIDIA A100 GPUs, whereas all downstream experiments are carried out on 4 A100 GPUs. During downstream training, the same Adam W optimizer and cosine scheduler are retained, but the learning rate is reduced to 1/10 of its pretraining value. Table 12: Hyperparameters for masked ECG pre-training with Physio Wave-ecg. Table 13: Hyperparameters for masked EMG pre-training with Physio Wave-emg. Table 14: Hyperparameters for downstream fine-tuning with Physio Wave. |