TLDR: TS-P2CL is a novel framework that improves medical time series classification by transforming 1D physiological signals into 2D pseudo-images, allowing it to leverage pre-trained vision models. It uses a dual contrastive learning strategy—intra-modal consistency and cross-modal alignment—to learn robust, domain-invariant representations. The method is plug-and-play, as it uses frozen vision models, and has shown superior generalization across various medical datasets in both subject-dependent and subject-independent settings, addressing the critical challenge of cross-individual heterogeneity in intelligent healthcare.
Medical time series (MedTS) classification is a crucial area in intelligent healthcare, involving the analysis of physiological signals like electroencephalography (EEG) and electrocardiography (ECG) to enable real-time monitoring, early intervention, and personalized medicine. However, a significant hurdle in its clinical application is the poor generalization of models across different individuals. This is due to the vast differences in signal patterns, amplitudes, and noise profiles from one person to another, making it difficult for models to learn universally applicable representations.
Existing approaches, such as advanced neural network architectures and transfer learning techniques, have made progress but still face limitations. Many rely on biases specific to a particular type of data, which can prevent them from learning features that are truly universal. Transferring knowledge from language models also presents a challenge, as it involves bridging a fundamental gap between continuous physiological signals and discrete text, potentially distorting the original temporal dynamics.
Introducing TS-P2CL: A Vision-Guided Approach
To address these challenges, researchers have proposed a novel framework called TS-P2CL: Plug-and-Play Dual Contrastive Learning for Vision-Guided Medical Time Series Classification. This innovative method draws inspiration from the remarkable ability of pre-trained vision models to recognize universal patterns in images. The core idea is to transform one-dimensional physiological signals into two-dimensional ‘pseudo-images,’ effectively creating a bridge to the visual domain. This transformation allows the system to tap into the rich semantic knowledge that vision models have already learned from natural images.
The intuition behind this approach is that time series and natural images share fundamental structural commonalities. For instance, trends in a time series can resemble edges in an image, periodic patterns might look like textures, and sudden changes correspond to discontinuities. This concept is not only theoretically sound but also clinically intuitive, as medical professionals frequently diagnose conditions by visually examining waveforms.
Dual Contrastive Learning for Robust Representations
Within this unified visual space, TS-P2CL employs a sophisticated ‘dual contrastive learning’ strategy to learn robust and generalizable representations. This strategy involves two complementary objectives:
-
Intra-modal consistency: This component ensures that the model learns stable and discriminative representations by enforcing temporal coherence. It does this by comparing different augmented versions of the same time series, treating them as positive pairs and maximizing their similarity while pushing them away from other, negative samples.
-
Cross-modal alignment: This crucial part directly aligns the time-series embeddings with the features extracted from a *frozen* pre-trained vision model. By keeping the vision model frozen, TS-P2CL becomes a ‘plug-and-play’ solution, efficiently transferring universal visual knowledge without requiring any fine-tuning of the vision model itself. This alignment helps to mitigate individual-specific biases and learn features that are more robust and independent of the specific data domain.
The framework also incorporates a ‘progressive joint learning’ strategy, which dynamically balances self-supervised contrastive objectives with supervised classification. This allows the model to first discover robust underlying structures and then refine its predictive accuracy for specific tasks.
Demonstrated Effectiveness
Extensive experiments were conducted on six diverse medical time series datasets, including EEG, ECG, and sEEG. TS-P2CL consistently outperformed fourteen other methods in both subject-dependent (where samples are randomly split across subjects) and, more importantly, subject-independent settings (where all data from a subject stays in one split). The strong performance in subject-independent scenarios highlights the method’s ability to generalize well to new individuals, which is critical for real-world clinical deployment.
The results showed superior accuracy, precision, recall, F1 score, and AUROC across various datasets. For instance, on the ADFD dataset, TS-P2CL achieved 98.36% accuracy in subject-dependent classification. In subject-independent settings, it demonstrated consistent F1 dominance, indicating a strong balance between precision and recall. This balance is vital in healthcare, as it minimizes false positives (avoiding unnecessary treatments) and false negatives (preventing delayed interventions).
Ablation studies further confirmed the effectiveness of both intra-modal and cross-modal contrastive learning components, showing their complementary contributions to performance gains. The method also proved to be model-agnostic, improving performance across different vision models like VanillaNet and ViT-Tiny. Visualizations using t-SNE demonstrated that the combined dual contrastive loss leads to clearer and more discriminative embeddings.
Also Read:
- UNIPHY+: A New AI Framework for Continuous Health Monitoring from ICU to Home
- SegReg: A Segmentation-Driven Approach for Precise Medical Image Alignment
Conclusion
TS-P2CL represents a significant advancement in medical time series classification. By innovatively bridging the gap between 1D physiological signals and 2D visual representations, and leveraging dual contrastive learning with frozen pre-trained vision models, it offers a plug-and-play framework that learns robust, domain-invariant features. This enhances cross-subject generalization, making it a promising solution for improving intelligent healthcare applications. For more detailed information, you can refer to the full research paper: TS-P2CL: Plug-and-Play Dual Contrastive Learning for Vision-Guided Medical Time Series Classification.
