Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification

Authors: Yihe Wang, Nan Huang, Taida Li, Yujun Yan, Xiang Zhang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on five public datasets under both subject-dependent and challenging subject-independent setups. Results demonstrate Medformer s superiority over 10 baselines, achieving top averaged ranking across five datasets on all six evaluation metrics.
Researcher Affiliation Academia Yihe Wang , Nan Huang , Taida Li University of North Carolina Charlotte {ywang145,nhuang1,tli14}@charlotte.edu Yujun Yan Dartmouth College yujun.yan@dartmouth.edu Xiang Zhang University of North Carolina Charlotte xiang.zhang@charlotte.edu
Pseudocode No The paper describes its method using natural language and figures, but does not include formal pseudocode or algorithm blocks.
Open Source Code Yes We release the source code at https://github.com/DL4m Health/Medformer.
Open Datasets Yes We conduct extensive experiments on five public datasets... Datasets. (1) APAVA [67] is an EEG dataset... (2) TDBRAIN [68] is an EEG dataset... (3) ADFTD [69, 19] is an EEG dataset... (4) PTB [70] is an ECG dataset... (5) PTB-XL [71] is an ECG dataset... URLs for these datasets are provided in Appendix B: https://osf.io/jbysn/ for APAVA, https://brainclinics.com/resources/ for TDBrain, https://openneuro.org/datasets/ds004504/versions/1.0.6 for ADFTD, https://physionet.org/content/ptbdb/1.0.0/ for PTB, https://physionet.org/content/ptb-xl/1.0.3/ for PTB-XL.
Dataset Splits Yes In this setup [Subject-Dependent], the division into training, validation, and test sets is based on time series samples. In this setup [Subject-Independent], the division into training, validation, and test sets is based on subjects. For the training, validation, and test set splits, we employ the subject-independent setup. Samples with subject IDs {15,16,19,20} and {1,2,17,18} are assigned to the validation and test sets, respectively. The remaining samples are allocated to the training set.
Hardware Specification Yes All experiments are run on an NVIDIA RTX 4090 GPU and a server with 4 RTX A5000 GPUs.
Software Dependencies No The paper mentions software like 'Time-Series-Library project', 'Py Torch', and 'Adam optimizer', but it does not specify version numbers for these software dependencies, which is required for reproducibility.
Experiment Setup Yes For all methods, we employ 6 layers for the encoder, with the self-attention dimension D set to 128 and the hidden dimension of the feed-forward networks set to 256. The optimizer used is Adam, with a learning rate of 1e-4. The batch size is set to {32,32,128,128,128} for the datasets APAVA, TDBrain, ADFD, PTB, and PTB-XL, respectively. Training is conducted for 100 epochs, with early stopping triggered after 10 epochs without improvement in the F1 score on the validation set.