Self-Supervised Contrastive Learning for Medical Time Series: A Systematic Review

Xiang Zhang, Ziyu Liu, Azadeh Alavi, Minyi Li, Xiang Zhang 0012Sensors 2023 · DOI

TL;DR

This paper presents a systematic review of self-supervised contrastive learning (SSCL) applied to medical time series data. It highlights SSCL's potential to overcome the significant challenge of label scarcity in healthcare by learning meaningful data representations without explicit labels, paving the way for more efficient and accurate medical AI applications.

What Problem Does This Paper Address?

Medical time series data, such as ECG and EEG signals, are crucial for diagnosing and monitoring various health conditions. However, extracting valuable insights from this data using deep learning is severely hampered by the "label scarcity" problem. Manually annotating these time series requires extensive expert knowledge, is time-consuming, and often leads to inconsistencies. This bottleneck prevents the full utilization of the vast amounts of medical data being generated daily. Self-supervised contrastive learning (SSCL) has emerged as a promising solution, learning representations by comparing similar (positive) and dissimilar (negative) data samples without relying on explicit labels. This review systematically examines how SSCL is being applied to medical time series to address this critical data limitation.

Methods

This paper is a systematic review, following PRISMA guidelines, to analyze existing literature on SSCL for medical time series. The authors searched five major databases (IEEE, ACM, Scopus, Google Scholar, PubMed) and screened over 2100 papers, ultimately including 43 for detailed analysis. The review categorizes SSCL into three stages:

  • Pre-training: The core unsupervised learning phase where an encoder learns representations from unlabeled data by maximizing agreement between augmented versions of the same sample (positive pairs) and minimizing agreement with different samples (negative pairs).
  • Fine-tuning: The supervised phase where the pre-trained encoder is used with a smaller set of labeled data to train a downstream classifier for a specific medical task.
  • Testing: Evaluating the performance of the fine-tuned model on unseen data.

The review delves into key components of SSCL for medical time series, including: data augmentation techniques (transforming, masking, neighboring), pretext tasks (contrastive mapping, predictive coding, neighbor detection), pre-training encoder architectures (CNNs, RNNs like LSTM/GRU, ResNet), fine-tuning classifiers (logistic regression, MLP, linear layers), and contrastive loss functions (NT-Xent, NCE, InfoNCE, Triplet loss).

Datasets

The paper identifies and summarizes 51 public datasets used in the reviewed studies. The most prevalent data types are ECG, EEG, and ICU data. Notable datasets include:

  • ECG: PhysioNet 2017, CPSC 2018, PTB-XL, MIT-BIH, Chapman-Shaoxing.
  • EEG: Sleep-EDF, MASS, SEED, TUAB, EPILEPSIAE.
  • ICU: HiRID, eICU, PhysioNet 2012, MIMIC-III.

The paper notes that while many datasets exist, there's a demand for more diverse datasets beyond ECG, EEG, and ICU, and for a broader range of medical applications.

Key Insights and Ideas

  • SSCL is a powerful tool for label scarcity: The core message is that SSCL significantly mitigates the need for extensive expert annotations, which is a major bottleneck in medical AI.
  • Data augmentation is crucial but challenging: The effectiveness of SSCL heavily relies on how time series data is augmented. However, designing effective augmentations for the diverse temporal characteristics of medical data remains a significant challenge.
  • ECG and EEG dominate: The majority of studies focus on ECG and EEG data, likely due to the availability of large-scale public datasets.
  • Cardiovascular diseases and sleep status are primary applications: These areas, along with ICU-related tasks, represent the most common downstream applications for SSCL in medical time series.
  • Lack of unified framework for hierarchical data: Medical time series are often hierarchical (patient -> session -> trial -> sample). Current SSCL methods struggle to globally consider all these levels, limiting their potential.
  • Regression tasks are under-explored: While classification is common, SSCL's application to regression tasks (e.g., predicting future health trajectories) is less explored, despite its importance for early diagnosis.

Key Results

The review found 43 relevant publications applying SSCL to medical time series. The distribution of applications shows:

  • Cardiovascular diseases: 25.5%
  • Sleep status: 20%
  • ICU-related: 12.7%
  • Neurological disorder: 12.7%
  • Motor-Imagery classification: 5.45%
  • Maternal/Fetal health: 5.45%
  • Emotion recognition: 3.64%
  • COVID detection: 3.64%

The paper provides detailed tables summarizing the studies, including data types, applications, augmentations, pretext tasks, encoders, classifiers, loss functions, and datasets used.

Limitations and Future Work

The authors identify several limitations and suggest future research directions:

  • Less guidance for augmentation design: There's a need for more theoretical guidance on designing effective augmentations for time series, moving beyond empirical trial-and-error. Future work should explore more innovative and effective augmentations.
  • Lack of unified framework for hierarchical time series: Developing a global framework to handle the hierarchical nature of medical time series data is crucial.
  • Limited regression tasks: More research is needed to explore SSCL for regression tasks, which are vital for health trajectory monitoring and early diagnosis.
  • Lack of scalability: SSCL can be computationally expensive due to data augmentation and larger negative sample sets. Improving scalability is important.
  • Limited ability in multimodal time series: Current SSCL models primarily focus on univariate time series. Developing methods for effectively handling multimodal data is essential.
  • Lack of open-access diverse biomedical datasets: More diverse datasets beyond ECG, EEG, and ICU are needed to advance research across a wider range of medical applications.
  • Future work could explore generative representation learning: The authors suggest a future review focusing on generative SSCL models for medical time series.

Summary updated

BibTeX
@article{zhang2023self,
  author = {Xiang Zhang and Ziyu Liu and Azadeh Alavi and Minyi Li and Xiang Zhang},
  title = {{Self-Supervised Contrastive Learning for Medical Time Series: A Systematic Review}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23094221},
}
Figure