Subject-Aware Contrastive Learning for Biosignals

30 Jun 2020  ·  Joseph Y. Cheng, Hanlin Goh, Kaan Dogrusoz, Oncel Tuzel, Erdrin Azemi ·

Datasets for biosignals, such as electroencephalogram (EEG) and electrocardiogram (ECG), often have noisy labels and have limited number of subjects (<100). To handle these challenges, we propose a self-supervised approach based on contrastive learning to model biosignals with a reduced reliance on labeled data and with fewer subjects. In this regime of limited labels and subjects, intersubject variability negatively impacts model performance. Thus, we introduce subject-aware learning through (1) a subject-specific contrastive loss, and (2) an adversarial training to promote subject-invariance during the self-supervised learning. We also develop a number of time-series data augmentation techniques to be used with the contrastive loss for biosignals. Our method is evaluated on publicly available datasets of two different biosignals with different tasks: EEG decoding and ECG anomaly detection. The embeddings learned using self-supervision yield competitive classification results compared to entirely supervised methods. We show that subject-invariance improves representation quality for these tasks, and observe that subject-specific loss increases performance when fine-tuning with supervised labels.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Person Identification EEG Motor Movement/Imagery Dataset SSL Accuracy 0.886 # 1
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-invariant Finetuned Subject-invariant SSL 1D ResNet Accuracy 0.498 # 5
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-invariant Finetuned Subject-invariant SSL 1D ResNet Accuracy 0.796 # 4
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-specific Finetuned Subject-specific SSL 1D ResNet Accuracy 0.539 # 1
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-specific Finetuned Subject-specific SSL 1D ResNet Accuracy 0.816 # 1
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-specific Finetuned Base SSL 1D ResNet Accuracy 0.526 # 2
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-invariant Supervised 1D ResNet Accuracy 0.44 # 7
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-invariant Supervised 1D ResNet Accuracy 0.763 # 7
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-specific Supervised 1D ResNet Accuracy 0.506 # 3
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-specific Supervised 1D ResNet Accuracy 0.81 # 2
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Random Forest Accuracy 0.801 # 3
Person Identification EEG Motor Movement/Imagery Dataset Subject-invariant SSL Embedding & Linear Classifier Accuracy 0.73 # 2
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-invariant SSL Accuracy 0.503 # 4
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-invariant SSL Accuracy 0.794 # 5
Person Identification EEG Motor Movement/Imagery Dataset Subject-specific SSL Accuracy 0.684 # 3
EEG Left/Right hand EEG Motor Movement/Imagery Dataset Subject-specific SSL Accuracy 0.772 # 6
EEG 4 classes EEG Motor Movement/Imagery Dataset Subject-specific SSL Accuracy 0.464 # 6

Methods


No methods listed for this paper. Add relevant methods here