Parallel-Learning of Invariant and Tempo-variant Attributes of Single-Lead Cardiac Signals: PLITA

TL;DR

PLITA tackles the limitation of prior self-supervised ECG learning by explicitly encoding both invariant and tempo-variant attributes in single-lead signals. It achieves this with parallel learning: a BYOL-based invariant objective and a novel tempo-variant loss $\mathcal{L}_{tv}$ that enforces a time-ordered spatial structure via the ideal matrix $\mathbf{M}^{tv}$ and its prediction $\hat{\mathbf{M}}^{tv}$. The method uses separate projection heads to avoid conflicting objectives and demonstrates superior performance on AFib classification and Sleep Stage detection, with competitive gender classification, supported by representation analyses (disentangling and SHAP). Findings indicate tempo-variant information carries unique, task-relevant signals, motivating broader adoption of tempo-aware objectives in time-series SSL and potential extension to other modalities and architectures.

Abstract

Wearable sensing devices, such as Holter monitors, will play a crucial role in the future of digital health. Unsupervised learning frameworks such as Self-Supervised Learning (SSL) are essential to map these single-lead electrocardiogram (ECG) signals with their anticipated clinical outcomes. These signals are characterized by a tempo-variant component whose patterns evolve through the recording and an invariant component with patterns that remain unchanged. However, existing SSL methods only drive the model to encode the invariant attributes, leading the model to neglect tempo-variant information which reflects subject-state changes through time. In this paper, we present Parallel-Learning of Invariant and Tempo-variant Attributes (PLITA), a novel SSL method designed for capturing both invariant and tempo-variant ECG attributes. The latter are captured by mandating closer representations in space for closer inputs on time. We evaluate both the capability of the method to learn the attributes of these two distinct kinds, as well as PLITA's performance compared to existing SSL methods for ECG analysis. PLITA performs significantly better in the set-ups where tempo-variant attributes play a major role.

Figures (6)

• Figure 1: A pair ECG strips from distinct subjects are shown. The signal morphology accommodates a strong stationary component with visible differences between the subjects.
• Figure 2: Considering time-sorted inputs equally spaced in time ($X_1 \dots X_4$), representations of nearby inputs in time are expected to be closer than time-distant ones.
• Figure 3: PLITA illustrated. Built on top of BYOL, PLITA includes both a student and a teacher network. For the sake of clarity, the teacher network is not included in the illustration. The losses are computed between representation triplets from both networks that process data equally. While $\mathcal{L}_{iv}$ is computed between a set of $N$ time series representations belonging to different records (displayed in black and blue colors), $\mathcal{L}_{tv}$ is computed between representations belonging to the same record. All inputs belong to the same subject. The encoder ($\mathcal{F}(X)$) is saved at the end of the training procedure and used for downstream tasks.
• Figure 4: The features that play an important role in the gender classification task (displayed in Figure \ref{['fig:static_importance']}), are highlighted in green in Figure \ref{['fig:static_overlap']}. The feature that accounts for the AFib classification task (Figure \ref{['fig:dynamic_importance']}) is displayed in purple.
• Figure 5: The informative features in the AFib classification are displayed in Figure \ref{['fig:dynamic_importance']} and highlighted in purple in Figure \ref{['fig:dynamic_overlap']}.