PhysioJEPA / docs /PAPERS.md
guychuk's picture
Upload folder using huggingface_hub
31e2456 verified

PAPERS.md β€” PhysioJEPA Reference Index

Oz Labs β€” April 2026 Covers every paper referenced across the full conversation and all project documents.


How to use this file

Three things per entry:

  1. What to use it for β€” the specific task or decision the agent needs this paper for
  2. Key numbers β€” exact figures the agent must not get wrong in code or prose
  3. Location β€” where to fetch the PDF

Read the tier before writing any code in that tier's domain. Do not cite a number that isn't in this file without fetching the source first.


Tier 1 β€” Implement from these

Read before writing any training code. Contains exact equations, hyperparameters, architecture details.


[T1-1] Weimann & Conrad β€” ECG-JEPA

arXiv: 2410.13867 Β· arxiv.org/pdf/2410.13867 Code: github.com/kweimann/ECG-JEPA ← fork this

Use for: This is the codebase we fork. Before writing any encoder code, read Section 2 (architecture), Section 3 (data), Appendix A (hyperparameters).

  • Patch tokenisation: 2D over (12 leads Γ— time), patch size = 25 time steps at 500 Hz
  • Masking: multi-block contiguous, 50% ratio, 4 target blocks
  • EMA: Ο„ starts 0.996, cosine-annealed to 0.9999 over training
  • Loss: L1 in latent space β€” no pixel decoder
  • ViT-S: 12 layers, d=256, 8 heads, MLP ratio=4

Key numbers: PTB-XL all-statements AUC 0.945 β€” this is Baseline A in the experiment matrix. Training time ~26h on RTX 3090.


[T1-2] Assran et al. β€” I-JEPA

arXiv: 2301.08243 Β· arxiv.org/pdf/2301.08243 Code: github.com/facebookresearch/ijepa

Use for: The masking strategy foundation. Why multi-block contiguous > random masking (forces semantic prediction, not texture interpolation). The stop-gradient / EMA target encoder design justification. The predictor should be narrower than the encoder β€” this prevents shortcutting through the predictor.

Key numbers: ViT-H/14 ImageNet β€” scale reference only, not a target for us.


[T1-3] Bardes et al. β€” V-JEPA (Revisiting Feature Prediction)

arXiv: 2404.08471 Β· arxiv.org/pdf/2404.08471

Use for: Spatiotemporal tube masking β€” how to mask contiguous blocks across both spatial and temporal axes simultaneously. Template for PPG 1D+time representation. Two-encoder EMA recipe at scale. Why predicting in latent space beats pixel reconstruction for noisy signals β€” core justification for JEPA over MAE.

Key numbers: SSv2 top-1 77.3%.


[T1-4] Balestriero & LeCun β€” LeJEPA

arXiv: 2511.08544 Β· arxiv.org/pdf/2511.08544

Use for: Ablation A3 only (SIGReg). Do not implement SIGReg without reading this first.

  • Theorem 1: isotropic Gaussian is the optimal JEPA embedding distribution
  • SIGReg: K=128 random 1D projections w~N(0,I), KL(zΒ·w || N(0,1)) per projection, sum. O(Kd).
  • Ξ» range: [0.01, 0.1]; start at 0.05
  • Apply to pooled global representation only β€” not per-patch tokens
  • ~50 lines of PyTorch

Key numbers: 79% ImageNet ViT-H/14 with only 2 loss terms.


[T1-5] Kim β€” CroPA-ECG-JEPA

arXiv: 2410.08559 Β· arxiv.org/pdf/2410.08559 Code: github.com/sehunfromdaegu/ECG_JEPA

Use for: Second ECG-JEPA implementation for debugging. Cross-Pattern Attention (CroPA) = inter-lead masked attention = inspiration for cardiac phase encoding in ablation A2. Also: 1D PE for predictor vs 2D for encoders β€” different from Weimann, compare before finalising.

Key numbers: Recovers HR and QRS duration from frozen representations without supervised training β€” target behaviour for PTT.


[T1-6] Botman et al. β€” Laya (LeJEPA for EEG)

arXiv: 2603.16281 Β· arxiv.org/pdf/2603.16281

Use for: Most direct prior to PhysioJEPA. Read before implementing ablation A3.

  • SIGReg with aggressive Ξ» destabilises training on impulsive signals (QRS-like spikes in EEG)
  • Mitigation: lower Ξ» (0.001–0.01), aggressive gradient clipping, apply to pooled global rep only
  • Latent prediction outperforms reconstruction on EEG clinical tasks

Key numbers: Outperforms reconstruction baselines on EEG-Bench with 10% of pretraining data.


Tier 2 β€” Baseline numbers and comparisons

Read to correctly report comparison numbers. Getting baselines wrong is a rejection risk.


[T2-1] Nie et al. β€” AnyPPG

arXiv: 2511.01747 Β· arxiv.org/pdf/2511.01747

Use for: Primary contrastive baseline (Baseline C in experiment matrix).

  • Exact loss: symmetric InfoNCE with learnable temperature Ο„
  • CRITICAL: ECGFounder encoder is FROZEN during AnyPPG training. ECG is a fixed supervisory signal. AnyPPG is not a jointly trained dual-encoder model.
  • Architecture: Net1D (PPG branch), ECGFounder frozen (ECG branch)
  • Trained on >100,000 hours

Key numbers: PPG→ECG retrieval R@1=0.736, R@5=0.906, R@10=0.935. AF detection AUC ~0.90. Mean 9.1% AUC improvement over non-ECG-guided baselines.


[T2-2] Wagner et al. β€” PTB-XL

arXiv: 2004.13701 Β· arxiv.org/pdf/2004.13701

Use for: ECG evaluation benchmark. Task definitions, train/test/val splits, and label hierarchy. Must replicate Weimann's exact split for comparison.

Key numbers: Weimann ECG-JEPA AUC 0.945 all-statements = Baseline A target.


[T2-3] Charlton et al. β€” Towards Ubiquitous BP Monitoring via PTT (review)

URL: pmc.ncbi.nlm.nih.gov/articles/PMC4515215/

Use for: Before writing E4 rollout coherence physiological consistency checks. PTT definition, normal range, PTT–BP and HR–PTT relationships. Per-patient calibration required for absolute BP β€” do not claim uncalibrated absolute BP from PTT.

Key numbers: Normal PTT 100–400ms (ICU adults). Within-patient tracking ~10 mmHg MAE with calibration.


[T2-4] Assran et al. β€” V-JEPA 2 (including V-JEPA 2-AC)

arXiv: 2506.09985 Β· arxiv.org/pdf/2506.09985

Use for: Architecture D future work template. Two-stage recipe: action-free pretraining β†’ action-conditioned fine-tuning with frozen encoder.

Key numbers: <62 hours of robot interaction data for Stage 2. SSv2 top-1 77.3%.


Tier 3 β€” Related work framing

Read to correctly describe prior work and differentiate PhysioJEPA.


[T3-1] Sarkar & Etemad β€” CardioGAN

arXiv: 2010.00104 Β· arxiv.org/pdf/2010.00104 Code: github.com/pritamqu/ppg2ecg-cardiogan

Use for: First major cross-modal ECG-PPG paper (AAAI 2021).

  • Uses CycleGAN backbone with attention-based generators and dual time/frequency discriminators
  • NOT reconstruction/L1, NOT InfoNCE β€” adversarial + cycle consistency loss
  • t=0 alignment β€” discards lag. Do NOT call this "pixel reconstruction."

[T3-2] Liu, Wang & Wang β€” TSTA-Net

PMLR: proceedings.mlr.press/v278/liu25d.html

Use for: Hierarchical contrastive ECG-PPG baseline (PMLR 2025).

  • Hierarchical contrastive learning β€” NOT raw InfoNCE
  • 9.3% higher AF F1 vs prior SSL methods
  • Still t=0 aligned

[T3-3] Fang et al. β€” PPGFlowECG

arXiv: 2509.19774 Β· arxiv.org/pdf/2509.19774

Use for: Two-stage generative translation baseline.

  • Stage 1: InfoNCE instance alignment (CardioAlign encoder, shared weights)
  • Stage 2: rectified flow generation from aligned latents
  • Figure 1 explicitly shows ECG precedes PPG temporally but the architecture does not exploit this
  • Do NOT describe as "rectified flow only" β€” InfoNCE is in Stage 1

[T3-4] Dong et al. β€” Brain-JEPA (NeurIPS 2024 Spotlight)

arXiv: 2409.19407 Β· arxiv.org/pdf/2409.19407 Code: github.com/hzlab/2024_Dong_Li_NeurIPS_Brain-JEPA

Use for: Cardiac phase encoding inspiration (ablation A2). Brain Gradient Positioning β†’ our cardiac phase PE. Hard phase boundaries fail during AF β€” use soft Gaussian encoding over cardiac landmarks.

Key numbers: NeurIPS 2024 Spotlight. UK Biobank 40k patients.


[T3-5] Hojjati et al. β€” EEG-VJEPA

arXiv: 2507.03633 Β· arxiv.org/pdf/2507.03633 Code: github.com/amir-hojjati/eeg-vjepa

Use for: V-JEPA adapted to 1D physiological signal β€” most direct predecessor. How to reshape multi-channel 1D signal into 3D tensor treated as "video." UMAP showing pathological clustering without labels.

Key numbers: TUH fine-tuned accuracy 85.8%, AUROC 88.5%. Frozen probe 83.3%.


[T3-6] Munim et al. β€” EchoJEPA

arXiv: 2602.02603 Β· arxiv.org/pdf/2602.02603

Use for: Strongest empirical evidence that JEPA > MAE for noisy medical signals. Use in intro to justify JEPA over MAE.

Key numbers: JEPA degrades 2% under perturbation vs 17% for VideoMAE. 79% accuracy at 1% labels. 20% LVEF improvement.


[T3-7] Wu, Lei et al. β€” SurgMotion

arXiv: 2602.05638 Β· arxiv.org/pdf/2602.05638

Use for: One-sentence citation alongside EchoJEPA: "JEPA's noise rejection under clinical signal artifacts has been validated in echocardiography [EchoJEPA] and surgical video [SurgMotion]."


[T3-8] LeCun β€” A Path Towards Autonomous Machine Intelligence (JEPA position paper)

URL: openreview.net/pdf?id=BZ5a1r-kVsf

Use for: One intro citation: "A world model should predict consequences of actions in abstract representation space [LeCun 2022]."


[T3-9] Abbaspourazad et al. β€” Apple Heart Study Foundation Model

arXiv: 2312.05409 Β· arxiv.org/pdf/2312.05409 Published: ICLR 2024

Use for: Prior art on wearable-scale PPG+ECG foundation models. InfoNCE + KoLeo, participant-level positives, Apple Watch data. Shows ECG more discriminative than PPG β€” context for why cross-modal training helps PPG.


Tier 4 β€” Evaluation methodology and datasets

Read when writing the evaluation harness code.


[T4-1] Pimentel et al. β€” BIDMC PPG and Respiration Dataset

PhysioNet: physionet.org/content/bidmc/1.0.0/

Use for: Fallback dataset if E0 fails.

  • WFDB format, 53 recordings Γ— 8 min, 125 Hz
  • Signals: Lead II ECG + fingertip PPG + impedance respiration
  • Labels: HR, RR, SpO2 β€” no AF labels (use for HR probe only)

Key numbers: 53 patients, ~7 hours total, 125 Hz.


[T4-2] Moody et al. β€” MIMIC-IV Waveform Database

PhysioNet: physionet.org/content/mimic4wdb/0.1.0/

Use for: Understanding HuggingFace mirror provenance.

  • v0.1.0: 200 records from 198 patients; upcoming release ~10,000 records
  • MIMIC-IV-ECG module: ~800k ECGs across ~160k patients, 500 Hz, 10s, 12-lead β€” AF label source candidate

[T4-3] Kachuee et al. β€” Cuffless BP Estimation Dataset (UCI)

UCI: archive.ics.uci.edu/dataset/340

Use for: E5a PTT probe evaluation.

  • 12,000 records, 942 patients β€” patient ID removed β€” population-level evaluation only
  • PPG + ABP at 125 Hz, derived from MIMIC-II

Key numbers: AAMI standard ≀5 mmHg mean Β± 8 mmHg SD.


[T4-4] Goldberger et al. β€” PhysioBank, PhysioToolkit, PhysioNet

DOI: 10.1161/01.CIR.101.23.e215

Use for: Required citation whenever using BIDMC, MIMIC waveforms, or any PhysioNet dataset. One line in methods: "Data obtained from PhysioNet [Goldberger et al., 2000]."


Tier 5 β€” Context and intellectual lineage

Do not read these to implement anything. One citation each.


[T5-1] Ha & Schmidhuber β€” World Models

arXiv: 1803.10122

Use for: Intro citation only. "World models learn a compressed latent representation and a transition function [Ha & Schmidhuber, 2018]."


[T5-2] Bardes et al. β€” VICReg

arXiv: 2105.04906

Use for: Related work only. "VICReg requires hand-crafted augmentations that JEPA avoids."


[T5-3] Ronan et al. β€” VICReg for Brugada ECG Detection

DOI: 10.1038/s41598-025-94130-x

Use for: One sentence. "VICReg-based SSL has been applied to ECG classification [Ronan et al., 2025] but requires augmentation engineering."


[T5-4] Johnson et al. β€” MIMIC-IV (clinical database paper)

DOI: 10.1038/s41597-022-01899-x

Use for: Required data citation whenever using MIMIC-IV derived data. "MIMIC-IV [Johnson et al., 2023], a freely accessible EHR database."


[T5-5] CLIMB multimodal clinical benchmark

arXiv: 2503.07667

Use for: ECG-JEPA performance in multimodal settings. "ECG-JEPA outperforms general time-series models like UniTS by 36.8% on ECG tasks [CLIMB, 2025]." One citation in intro.


Quick reference: numbers the agent must not get wrong

Claim Correct value Source
ECG-JEPA PTB-XL AUC 0.945 all-statements T1-1 Weimann
AnyPPG PPG→ECG R@1 0.736 T2-1 Nie
AnyPPG AUC improvement 9.1% over non-ECG baselines T2-1 Nie
AnyPPG ECGFounder FROZEN during training T2-1 Nie
EchoJEPA JEPA perturbation 2% degradation T3-6 Munim
EchoJEPA MAE perturbation 17% degradation T3-6 Munim
EchoJEPA 1% label accuracy 79% T3-6 Munim
Normal PTT range (ICU) 100–400ms T2-3 Charlton
BIDMC size 53 recordings Γ— 8 min @ 125 Hz T4-1 Pimentel
V-JEPA 2-AC interaction data <62 hours T2-4 Assran
EEG-VJEPA TUH AUROC 88.5% fine-tuned T3-5 Hojjati
CardioGAN objective CycleGAN adversarial β€” not reconstruction T3-1 Sarkar
TSTA-Net objective Hierarchical contrastive β€” not raw InfoNCE T3-2 Liu
PPGFlowECG Stage 1 InfoNCE alignment, then rectified flow T3-3 Fang
BP calibration requirement Per-patient calibration required for absolute values T2-3 Charlton

File locations in repo

docs/papers/*.pdf

This is the complete reference index. Fetch from arXiv if a PDF is missing. Never cite a number not in this file without verifying the source first.