[Paper Review] Evaluating Contrastive Learning on Wearable Timeseries for Downstream Clinical Outcomes

Kevalee Shah, Dimitris Spathis+2 more|arXiv (Cornell University)|Nov 13, 2021

Machine Learning in Healthcare4 citations

TL;DR

This paper evaluates contrastive self-supervised learning (SimCLR and BYOL) on wearable time-series data for downstream clinical classification tasks, adapting data augmentation and architecture for temporal signals. It demonstrates that SimCLR outperforms both fully supervised and adversarial unsupervised methods across most disease prediction tasks, establishing contrastive learning as a powerful approach for clinical representation learning from unlabelled wearable data.

ABSTRACT

Vast quantities of person-generated health data (wearables) are collected but the process of annotating to feed to machine learning models is impractical. This paper discusses ways in which self-supervised approaches that use contrastive losses, such as SimCLR and BYOL, previously applied to the vision domain, can be applied to high-dimensional health signals for downstream classification tasks of various diseases spanning sleep, heart, and metabolic conditions. To this end, we adapt the data augmentation step and the overall architecture to suit the temporal nature of the data (wearable traces) and evaluate on 5 downstream tasks by comparing other state-of-the-art methods including supervised learning and an adversarial unsupervised representation learning method. We show that SimCLR outperforms the adversarial method and a fully-supervised method in the majority of the downstream evaluation tasks, and that all self-supervised methods outperform the fully-supervised methods. This work provides a comprehensive benchmark for contrastive methods applied to the wearable time-series domain, showing the promise of task-agnostic representations for downstream clinical outcomes.

Motivation & Objective

• To investigate whether contrastive self-supervised learning can extract meaningful representations from unlabelled wearable time-series data for downstream clinical prediction.
• To adapt vision-based contrastive methods (SimCLR and BYOL) to the temporal nature of wearable health signals.
• To benchmark these methods against fully supervised and adversarial unsupervised approaches on real-world clinical classification tasks.
• To evaluate performance across diverse conditions including sleep apnea, diabetes, hypertension, metabolic syndrome, and insomnia.

Proposed method

• Adapted SimCLR and BYOL frameworks for time-series by designing domain-specific data augmentations such as time reversal, segment shuffling, noise addition, and scaling.
• Used a two-branch neural network architecture with shared encoder weights to learn invariant representations from augmented views of the same time-series.
• Applied contrastive loss to pull positive pairs (augmentations of the same signal) closer and push negative pairs (from different signals) apart in the embedding space.
• Employed a linear classifier head on frozen representations for downstream disease classification, evaluating using F1-macro and F1-micro metrics.
• Selected a window size of 512 timesteps and batch size of 1024 to optimize contrastive learning performance.
• Compared results against a fully supervised CNN and an adversarial unsupervised method (day2vec) from prior work.

Experimental results

Research questions

• RQ1 Can contrastive self-supervised learning extract clinically relevant representations from unlabelled wearable time-series data?
• RQ2 How do SimCLR and BYOL compare to fully supervised and adversarial unsupervised methods in downstream classification of sleep, heart, and metabolic disorders?
• RQ3 Does the choice of data augmentation strategy significantly affect the quality of learned representations in the wearable time-series domain?
• RQ4 Is there a consistent performance advantage of self-supervised methods over fully supervised baselines across diverse clinical conditions?
• RQ5 How do evaluation metrics (macro vs. micro F1) influence the interpretation of model performance on imbalanced clinical datasets?

Key findings

• SimCLR outperformed both BYOL and the adversarial day2vec method on four out of five downstream classification tasks, including sleep apnea, metabolic syndrome, and hypertension.
• For metabolic syndrome classification—a novel task in this study—SimCLR achieved the highest F1-macro score, indicating strong generalization to new clinical conditions.
• BYOL achieved the best F1-micro score for hypertension, while SimCLR led in F1-macro, highlighting the impact of evaluation metric choice on model comparison.
• Both SimCLR and BYOL outperformed the fully supervised CNN baseline across all tasks, demonstrating the advantage of self-supervised pre-training on limited-label clinical data.
• The results show that self-supervised representations can surpass fully supervised models even when the latter are trained on the same downstream data, suggesting richer feature learning.
• The study establishes a benchmark for contrastive learning on wearable time-series, showing that task-agnostic representations are effective for diverse clinical outcomes.