Transforming label-efficient decoding of healthcare wearables with self-supervised learning and “embedded” medical domain expertise

Figures

• 

  Fig. 1. Pipelines of “old-school” domain feature engineering and “new-school” self-supervised contrastive learning (SSCL) for healthcare wearables. Take the electrocardiogram (ECG) as an example, visualizations illustrate how different ECG patterns are clustered in each pipeline. a In the “old-school” pipeline, domain features such as heart rate variability (HRV) and P-wave amplitude (P Amp) are extracted from raw electrocardiogram (ECG) signals using biomedical signal processing tools. They can be used to distinguish abnormal samples against normal samples. b In the “new-school” pipeline, two augmented views (Aug 1 and Aug 2) are generated from each ECG input and passed through a shared encoder. The model learns to minimize the distance between the augmented views (positive pair) while maximizing it relative to other samples (negative pairs). ECG electrocardiogram, HRV heart rate variability, P Amp P-wave amplitude.

• 

  Fig. 2. Overview of our domain-knowledge-guided self-supervised contrastive learning (SSCL) framework for decoding wearable time series data. a Offline domain feature extraction and clustering. Before training, domain features are extracted from raw input signals and clustered using k-means to assign a prototype index to each sample. b Augmentation and feature extraction. During training, two augmented views (Aug 1 and Aug 2) are generated per sample and passed through a shared encoder to obtain deep features. c Instance-level contrast. In addition to treating paired augmented views as a positive pair (marked in black), the nearest neighbor in the domain feature space is selected, and its two augmented views (marked in purple) are included as additional positive. d Prototype-level contrast. Each sample is encouraged to align with its domain-informed prototype representation, which is updated online using the encoded features of all assigned samples.

• 

  Fig. 3. Performance comparison in terms of class-average F1, Accuracy (Acc), Area Under the Receiver Operating Characteristic Curve (AUC), Recall, and Precision (Prec). We compared the performance on the test subset, based on the model with the best F1 score on the validation subset. The values are normalized by the deviation from the best performance for each respective metric.

• 

  Fig. 4. Grad-CAM-based waveform saliency map visualizations for two representative samples. A fully trained classifier, appended to each self-supervised learning backbone, was used to generate the saliency maps. These maps highlight the most important waveform segments that contribute to the model’s predictions. The saliency values, normalized between 0 and 1 for each sample, allow for improved visual comparison. Our method assigns higher saliency to clinically relevant regions, such as ST-segment elevation, thereby improving both interpretability and prediction accuracy.

• 

  Fig. 5. t-SNE visualizations of domain features, features extracted from end-to-end supervised learning, from SimCLR, and from our method, respectively. Each point corresponds to a sample from the Capture24 dataset. Each column represents the same plots, with points colorized by different manners. In the top panel, each sample points were colored by data split (train/validation/test), whilst in the bottom, shaped by class label.

• 

  Fig. 6. Performance comparison of pretrained models fine-tuned using varying proportions of labeled data under a semi-supervised learning setting on ECG datasets. Fine-tuning was performed using a supervised loss computed only on the labeled samples. Model performance is reported in terms of class-average F1 score.

• 

  Fig. 7. Comparison of feature group importance for direct logistic regression versus our framework. The heatmap on the left illustrates which combinations of features (AF-Feature, Morphological, RR Interval, and Similarity) were used, where the purple cubes indicate feature utility. The bar plot on the right shows the class-average F1 score for each feature combination, with direct logistic regression represented in blue (hatched) and our self-supervised contrastive learning method in green.