Research map / Vision-language and multimodal / Cross-modal learning
Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning
Hao Zhou, Simon A. Lee, Cyrus Tanade, Keum San Chun, Juhyeon Lee, Migyeong Gwak, Megha Thukral, Justin Sung, Eugene Hwang, Mehrab Bin Morshed, Li Zhu, Viswam Nathan and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain PPG encoder with xMAE masked cross-modal ECG reconstruction instead of unimodal or exchangeable-view multimodal baselines | PPG encoder · 19 downstream tasks (linear probing) | AUROC / MAE: 15 of 19 tasks best | Improved |
| xMAE vs strongest baseline on hypertension (lab) | PPG encoder · Hypertension (lab) | AUROC: 56.8 -> 68.8 | Improved |
| xMAE vs strongest baseline on A1C | PPG encoder · A1C | AUROC: 52.5 -> 65.1 | Improved |
| xMAE on BMI (lab) and Deep sleep, where baselines were better | PPG encoder · BMI (lab); sleep Deep | MAE; AUROC: BMI 4.09 vs 3.79 (PaPaGei-P); Deep 55.9 vs 57.5 | Mixed |
| use continuous block ECG masking instead of random masking (ablation) | xMAE · Hypertension (lab) and Ectopic Beats | AUROC: 63.0 -> 68.8; 84.4 -> 87.8 | Improved |
| use curriculum masking ratio instead of a fixed ratio (ablation) | xMAE · Hypertension (lab) and Ectopic Beats | AUROC: 66.9 -> 68.8; 85.8 -> 87.8 | Improved |
| use cross-modal reconstruction instead of multimodal self-reconstruction (ablation) | xMAE · Hypertension (lab) and Ectopic Beats | AUROC: 65.8 -> 68.8; 83.4 -> 87.8 | Improved |
| remove directional cross-attention (ablation) | xMAE · ECG reconstruction error | mean absolute reconstruction error | Hurt |
| xMAE pretraining vs multimodal MAE baseline for ECG-PPG timing | multimodal MAE · ECG R-peak to PPG onset delay | median delay error: 21.5 ms median, 53.3% reduction | Improved |
| xMAE vs handcrafted PPG features (all and timing-only) | logistic regression on features · Hypertension (free-living) | AUROC: 58.4 -> 58.5 | No clear effect |
| xMAE (smaller model and less data) vs open-weight foundation models | open-source PPG models · Ectopic beats | AUROC: AnyPPG 89.3 vs xMAE 87.8 | Mixed |
Closest papers by meaning
- Unlocking Cross-Modal Biosignal Synthesis: A Temporally-Aware VAE-Diffusion Model (ICML 2026 · similarity 0.89)
- BIOT: Biosignal Transformer for Cross-data Learning in the Wild (NeurIPS 2023 · similarity 0.87)
- Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.85)
- Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners (ICML 2025 · similarity 0.85)
- Parallel-Learning of Invariant and Tempo-variant Attributes of Single-Lead Cardiac Signals: PLITA (AAAI 2025 · similarity 0.84)
- Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation (NeurIPS 2026 · similarity 0.83)
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.
