Intra-Inter Subject Self-Supervised Learning for Multivariate Cardiac Signals

Authors: Xiang Lan, Dianwen Ng, Shenda Hong, Mengling Feng4532-4540

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three real-world datasets were conducted. In a semi-supervised transfer learning scenario, our pre-trained ISL model leads about 10% improvement over supervised training when only 1% labeled data is available, suggesting strong generalizability and robustness of the model. Experimental results demonstrate ISL outperforms current state-of-the-art methods in the downstream task under various scenarios.
Researcher Affiliation Academia Xiang Lan1, Dianwen Ng3, Shenda Hong4, 5 , Mengling Feng1, 2 1Saw Swee Hock School of Public Health, National University of Singapore, Singapore 2Institute of Data Science, National University of Singapore, Singapore 3School of Computer Science and Engineering, Nanyang Technological University, Singapore 4National Institute of Health Data Science, Peking University, Beijing, China 5Institute of Medical Technology, Health Science Center of Peking University, Beijing, China
Pseudocode Yes Algorithm 1: Self-supervised training procedure of ISL Input: Pre-training dataset P = {X1, X2, ..., Xp}, augmentation set G, pretraining iterations Max iter, number of frames N. Parameter: ISL encoder Fenc(θe), ISL discriminator D(θd). Output: Well-trained Fenc(ˆθe).
Open Source Code No The paper does not include an unambiguous statement that the authors are releasing their code for the work described in this paper, nor does it provide a direct link to a source-code repository for their method.
Open Datasets Yes To benchmark the performance of our proposed ISL model and to ensure reproducibility of our results, we pick three of the largest publicly available real-world ECG datasets for cardiac arrhythmias classification. Chapman (Zheng et al. 2020), CPSC (Liu et al. 2018), PTB-XL (Wagner et al. 2020).
Dataset Splits Yes We split each dataset into 60%, 20%, 20% in subject-wise for training, validation and testing. Table 1 shows description of each pre-processed dataset. Dataset Train Validation Test Categories Chapman 6,352 2,113 2,123 4 CPSC 5,612 1,870 1,818 9 PTB-XL 13,104 4,361 4,370 71
Hardware Specification Yes The experiments were conducted using Py Torch 1.8 (Paszke et al. 2019) on NVIDIA Ge Force Tesla V100 GPU
Software Dependencies Yes The experiments were conducted using Py Torch 1.8 (Paszke et al. 2019) on NVIDIA Ge Force Tesla V100 GPU
Experiment Setup Yes The model is optimized using Adam optimizer (Kingma and Ba 2015) with a learning rate of 3e-3 and weight decay of 4e-4. We use a hard-stop of 40 epochs and a batch size of 232 for both pre-training and downstream tasks, as the training loss does not further decrease.