Research map / Label-efficient and robust learning / Self-supervised and contrastive
Parallel-Learning of Invariant and Tempo-variant Attributes of Single-Lead Cardiac Signals: PLITA
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PLITA, a BYOL-based self-supervised method for single-lead ECG that adds a tempo-variant loss so that temporally closer segments have closer representations, alongside an invariant loss across recordings of the same subject.
Problem
Existing self-supervised methods for single-lead ECG capture only invariant (subject-level) attributes and neglect tempo-variant patterns that evolve over a recording, which matter for tasks like AFib detection and sleep staging.
Approach
Combines a BYOL-style invariant loss (positives from different recordings of the same subject) with a tempo-variant loss that matches cosine-distance matrices among time-sorted windows of the same record to an ideal |i-j|/(N-1) temporal distance matrix, using separate projectors for the two attribute kinds.
Key results
- On AFib classification with SHHS pretraining, PLITA reaches 80.7% accuracy and 78.4 F1, versus 76.6% accuracy for BYOL.
- On sleep stage classification with SHHS pretraining, PLITA reaches 75.3% accuracy and 0.81 AUC, versus 73.8% and 0.79 for Mix-Up.
- On gender identification PLITA is competitive but not best (76.5% accuracy vs 76.7% for BYOL with SHHS pretraining).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the tempo-variant loss to the invariant-learning objective (ablation) | PLITA · Tempo-variant-related downstream tasks | performance | Helped |
| Use PLITA instead of BYOL | ViT ECG representations · AFib classification, SHHS pretraining | accuracy: 76.6 -> 80.7 | Helped |
| Use PLITA instead of BYOL | ViT ECG representations · AFib classification, Icentia pretraining | accuracy: 75.3 -> 80.0 | Helped |
| Use PLITA instead of Mix-Up | ViT ECG representations · Sleep stage classification, SHHS pretraining | accuracy: 73.8 -> 75.3 | Helped |
| Use PLITA instead of BYOL | ViT ECG representations · Sleep stage classification, Icentia pretraining | accuracy: 73.3 -> 74.8 | Helped |
| Use PLITA instead of BYOL | ViT ECG representations · Gender identification, SHHS pretraining | accuracy: 76.7 -> 76.5 | Hurt |
| Use PLITA instead of BYOL | ViT ECG representations · Gender identification, Icentia pretraining | accuracy: 66.7 -> 66.5 | Hurt |
| Use N=3 rather than N=4 inputs per recording (ablation) | PLITA · AFib classification, SHHS pretraining | accuracy: 80.7 -> 79.5 | Hurt |
| Use N=5 rather than N=4 inputs per recording (ablation) | PLITA · AFib classification, SHHS pretraining | accuracy: 80.7 -> 77.3 | Hurt |
| Use W=90 rather than W=120 seconds (ablation) | PLITA · AFib classification, SHHS pretraining | accuracy: 80.7 -> 76.6 | Hurt |
| Use W=150 rather than W=120 seconds (ablation) | PLITA · AFib classification, Icentia pretraining | accuracy: 80.0 -> 79.4 | Hurt |
| Use W=150 rather than W=120 seconds (ablation) | PLITA · Sleep stage classification, Icentia pretraining | accuracy: 74.8 -> 72.9 | Hurt |
Surprising findings
- Gender classification, a purely invariant task, shows PLITA not outperforming BYOL and PCLR, which the authors attribute to the invariant objective being shared.
Limitations
- Evaluated only with a ViT architecture.
- Only BYOL is used for the invariant objective; other non-contrastive frameworks such as DINO or VICReg are not tested.
- Icentia pretraining yields lower performance on some tasks, attributed to noise in the data.
Experimental setup
Datasets: SHHS Icentia MIT-AFIB MIT-PSG Cinc2017
Benchmarks: MIT-AFIB MIT-PSG SHHS Cinc2017
Metrics: accuracy F1 score AUC
Baselines: PCLR CLOCS Mixing-Up Ti-MAE Siamese Masked Autoencoders BYOL
Builds on: BYOL PCLR CLOCS Siamese Image Embeddings ViT
Closest papers by meaning
- From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining (ICML 2025 · similarity 0.88)
- Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram (ICLR 2024 · similarity 0.88)
- Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model (ICLR 2025 · similarity 0.86)
- GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images (NeurIPS 2025 · similarity 0.85)
- Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners (ICML 2025 · similarity 0.84)
- Leveraging an ECG Beat Diffusion Model for Morphological Reconstruction from Indirect Signals (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.