Towards Self-Supervised Learning of ECG Signal Representation for the Classification of Acute Stress Types

We present a novel application of contrastive learning technique in learning the feature representation of ECG signal in a self-supervised manner for the classification of acute stress types. Acute stress types that occur for a very short period and are rapidly changing and alternating in nature are difficult to classify using conventional ECG features. This is because the change in conventional ECG features due to rapid and alternating acute stressors do not reflect instantaneously. We hypothesize that deep-learned features from ECG signals can better distinguish between the different stress types than conventional ECG features. Our proposed approach can generate distinct feature representations for the physical and mental stress task type using very short window lengths. Our results show that the deep-learned features perform better in terms of accuracy and F1 score in distinguishing between physical and mental stress task types. In the future, our proposed method can be used in a real world setting for understanding the dynamics of different stressors in a self-supervised fashion without the need for human labeling.