Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Flow-matching models are the strongest default for generating real-world sensor time series, and signal-aware design plus moderate synthetic data make the outputs useful beyond looking realistic.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 20:40 UTC pith:3FIMH5FD

load-bearing objection Useful multi-domain bake-off with real artifacts; flow-matching-as-default is a protocol-level finding, not a settled paradigm ranking. the 4 major comments →

arxiv 2607.04245 v1 pith:3FIMH5FD submitted 2026-07-05 cs.LG cs.AI

Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

Zitao Shuai , Zongzhe Xu , Yuntian Wu , Sirui Li , Tianhong Li , Yuzhe Yang This is my paper
classification cs.LG cs.AI
keywords sensor time seriesgenerative modelsflow matchingdiffusion modelssynthetic dataphysiological signalsevaluation protocols
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most real-world systems are recorded as continuous, noisy sensor streams rather than text or images, yet generative work on those streams has been scattered across one modality or one task at a time. This paper builds SensorGen, a shared study of 14 generation settings spanning four real sensing domains, seven public datasets, and twelve modalities, then trains and evaluates representative models from five modern generative families under one protocol. The central claim is that flow-matching models give the strongest overall baseline across most of those settings, while two signal-aware choices improve the hard cases: carefully normalized demographic covariates help long-horizon physiological generation, and time-frequency conditioning helps high-frequency channels such as EEG. The paper further argues that the outputs are not only visually plausible: more training and larger capacity improve quality, and moderate amounts of synthetic ECG improve both supervised disease classification and zero-shot transfer under data scarcity. A sympathetic reader cares because the work turns fragmented sensor-generation recipes into concrete, reusable design and evaluation lessons for physiology, wearables, and clinical monitoring.

Core claim

Across SensorGen's 14 standardized settings, flow-matching models (instantiated by SiT) provide the strongest overall generation performance among five major families, while demographic covariates improve longitudinal generation only when properly normalized, time-frequency modeling improves high-frequency fidelity, scaling improves quality, and moderate synthetic data improves downstream performance under scarcity.

What carries the argument

SensorGen: a unified pipeline that standardizes 14 real-world sensor generation settings (semantic-to-signal, interpolation/extrapolation, channel translation, and editing) across heterogeneous modalities and then compares five generative families under matched data, training budgets, and evaluation protocols.

Load-bearing premise

That a small set of off-the-shelf family instantiations, trained under a deliberately light shared protocol without heavy per-method tuning, is enough to rank entire generative paradigms for highly heterogeneous sensor signals.

What would settle it

If stronger, equally budgeted implementations of diffusion, autoregressive, normalizing-flow, or hierarchical models reverse the category-level rankings and outperform flow matching on the same SensorGen translation, forecasting, and semantic-to-signal suites, Takeaway 1 fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces SensorGen, a unified large-scale empirical study of generative modeling for real-world sensor time series spanning 14 settings across 4 domains, 7 datasets, and 12 modalities, organized into semantic-to-signal, interpolation/extrapolation, translation, and editing tasks. Under a shared preprocessing, training, and evaluation protocol, it compares representative models from five generative families (DiT, SiT, MAR, TarFlow, FractalGen/Imagen). The main claims are that (1) flow-matching (SiT) is a strong overall baseline across most settings, (2) signal-aware design choices matter—normalized demographic covariates help longitudinal generation and time-frequency conditioning helps high-frequency signals—and (3) generated signals have practical utility via scaling and moderate synthetic-data augmentation on downstream ECG tasks. The manuscript also analyzes normalization, long-sequence behavior, evaluation metrics, and failure modes, and releases code, models, and a website.

Significance. If the empirical picture holds, SensorGen is a substantial contribution to a fragmented area: it consolidates heterogeneous sensor-generation tasks under one protocol, covers unusually broad sequence lengths, frequencies, and time spans, and provides actionable design lessons beyond single-modality generators. Strengths include open code and model release, a clear task taxonomy with UnifiedBatch, setting-level tables in the appendix, and useful ablations (demographics, spectrogram conditioning, min–max vs z-score, scaling, synthetic augmentation on PTB-XL). The downstream utility experiments and the metric critique (Finding 4) are particularly valuable for the community. Even if the paradigm-ranking claim is softened, the resource and controlled analyses would remain a useful foundation for future sensor generative work.

major comments (4)
  1. [§4.1, Tables 2–6, Takeaway 1] §4.1, Takeaway 1, Abstract finding (1), and category-level Tables 2–6: the claim that flow-matching is the strongest overall default is only partially supported. Setting-level results already show non-uniform transfer—Imagen dominates training-required editing/super-resolution (Tables 5, 13), FractalGen is competitive on Peripheral-to-EEG (Table 10), MAR is competitive on cardiac forecasting (Table 11), and DiT/Imagen are close or better on text-to-ECG FID (Table 15). Category averages mix heterogeneous metrics (MSE/MAE/SMSE/PSNR/SSIM vs FID/Precision/Recall) without a stated aggregation rule or uncertainty. Please either (i) report multi-seed means±std and a transparent ranking procedure (e.g., per-setting ranks then average ranks), or (ii) reframe Takeaway 1 as “strong competitive baseline under a shared light protocol,” consistent with the paper’s own caveat that no method dominates e
  2. [§3.3, §4 Training Setup, Table 23] §3.3 and §4 Training Setup: paradigm-level conclusions rest on one lightly tuned representative per family (up to 50K steps or until convergence, limited hyperparameter search, capacity matching “when possible,” Table 23). This is a load-bearing premise for ranking entire families. Capacities already span ~143M–547M, and hierarchical models (Imagen/FractalGen) are used only on some tasks. The manuscript should either add at least one stronger family-specific training check on a subset of settings where rankings are close, or explicitly scope Takeaway 1 to “under a matched shared protocol with the chosen instantiations,” so readers do not over-generalize to all diffusion/flow/AR/NF/hierarchical methods.
  3. [Tables 2–6, Appendix C] Main result tables (Tables 2–6, 10–16) report point estimates only, with no multi-seed variance, bootstrap intervals, or significance tests. Several ranking margins are small (e.g., SiT vs DiT on Translation MSE 0.214 vs 0.224; SiT vs MAR on cardiac forecasting). Without uncertainty, the category-level “best overall” narrative is hard to trust. At minimum, re-run a representative subset of settings with ≥3 seeds and report mean±std; ideally mark statistically supported wins.
  4. [§5.1, Fig. 7, Tables 17–18] §5.1 / Finding 1 and Fig. 7: downstream utility is demonstrated only for text-to-ECG synthetic data on PTB-XL (supervised and zero-shot). The abstract and Finding 1 generalize to “synthetic sensor signals” and “downstream performance under data scarcity.” Please either extend utility experiments to at least one non-ECG modality/task in SensorGen, or narrow the claim to ECG/text-conditioned generation. Also clarify how synthetic labels are obtained and whether quality filtering was applied, given the observed degradation at 100% augmentation (Table 17).
minor comments (6)
  1. [Table 1, §3.3] Table 1 claims coverage of five model families and four task types; ensure the main text consistently says “five” vs “six” models (DiT, SiT, MAR, TarFlow, FractalGen, Imagen).
  2. [Fig. 6, §4.3] Fig. 6 and long-sequence analysis: report exact metric values in the main figure or caption (appendix Tables 20–21 help, but the main figure is hard to read quantitatively).
  3. [§3.2, Appendix B.1] Notation: C is used both for channel count and for the condition set; consider distinct symbols to avoid confusion in §3.2 and Appendix B.1.
  4. [Appendix D, Introduction] Typos/clarity: “glocuse” in Appendix D; “Generationsettings” spacing in the introduction; “Wedemonstratethat” and similar missing spaces in contribution bullets.
  5. [§4 Evaluation Protocol, Appendix E.2] SMSE definition (Eq. 1 in Appendix E.2) is useful; consider promoting a short definition to the main evaluation protocol so readers understand spectrogram-domain MSE without the appendix.
  6. [§6 Discussion] Clinical disclaimer in §6 is appropriate; also state clearly that public datasets may not cover rare conditions or deployment noise, so rankings may shift outside SensorGen’s coverage.

Circularity Check

0 steps flagged

Empirical bake-off with held-out metrics and external downstream tasks; no derivation reduces to its inputs by construction.

full rationale

SensorGen is a large-scale empirical comparison of generative model families on standardized sensor tasks, not a first-principles derivation. Takeaway 1 (flow matching as a strong default) and the category tables rest on training models from scratch under a shared protocol and scoring them on held-out conditions with metrics (MSE/MAE/SMSE/PSNR/SSIM, FID/Precision/Recall via MIRA) that are distinct from the training objectives. Ablations (demographic encoding, spectrogram conditioning, normalization, scaling, sequence length) and synthetic-data utility (PTB-XL supervised and zero-shot transfer) are likewise measured against external or held-out targets. Self-citations (e.g., HEARTS, OSF, SleepLM, SensorLM, SimPer) appear as background on sensor regimes or as co-authored model instantiations (MAR, FractalGen) used as family representatives; none is a uniqueness theorem or fitted constant that forces the comparative claims. Uneven per-setting wins and light hyperparameter search may weaken the strength of the ranking, but that is a fairness/generalization concern, not circularity. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

As an empirical systems/benchmark paper, the load-bearing content is not a short formal derivation but a stack of modeling and evaluation choices: which tasks count as valid sensor generation, which model checkpoints represent whole families, which normalizations and feature spaces define quality, and which training budgets make rankings fair. Those choices, not free physical constants, are what the central claims rest on.

free parameters (6)
  • training budget (up to 50K steps / until convergence)
    Shared compute budget used to compare families; rankings could shift with longer training or family-specific schedules.
  • model capacity targets (~143M–547M params by family)
    Parameter matching is approximate and architecture-native; capacity differences remain free design choices affecting family comparisons.
  • learning rates / EMA / warmup per family
    Table 23 lists family-specific optimization settings chosen by convention rather than exhaustive search.
  • patch length P for 1-D tokenization
    Transformer adaptations depend on non-overlapping 1-D patches; P is a design hyperparameter of the shared interface.
  • synthetic augmentation ratios (10%/50%/100%)
    Downstream utility conclusions depend on these hand-chosen mix ratios for real+synthetic ECG training.
  • noise scale σ in ECG denoising construction
    Editing/denoising targets are synthesized with additive Gaussian noise; σ shapes task difficulty and model rankings on that setting.
axioms (5)
  • domain assumption The four task categories and three validity criteria (evidence-grounded, application value, generative suitability) adequately represent real sensor-generation bottlenecks.
    §3.2 uses these criteria to construct the taxonomy; breadth claims depend on this coverage assumption.
  • ad hoc to paper One representative model per major family (DiT, SiT, MAR, TarFlow, FractalGen/Imagen) is enough to compare generative paradigms.
    §3.3 explicitly selects family instantiations rather than surveying many architectures per family.
  • domain assumption MIRA embeddings provide a valid feature space for FID/precision/recall on healthcare time series.
    §4 Evaluation Protocol uses MIRA for semantic-to-signal distribution metrics.
  • ad hoc to paper Per-window min–max normalization to [-1,1] is an appropriate shared target for generative training across modalities.
    §4 Data setup and Finding 2 treat fixed-range scaling as the default preprocessing choice after observing z-score issues.
  • domain assumption Sample-level MSE/MAE/PSNR/SSIM/SMSE on normalized signals measure generation quality for paired tasks.
    Used throughout Tables 2–6 and appendices; later analysis partially questions raw waveform metrics, but main rankings still rely on them.
invented entities (2)
  • SensorGen unified study/benchmark independent evidence
    purpose: Provide a shared multi-domain, multi-task protocol for comparing sensor generative models and design choices.
    The paper’s primary constructed object; independent evidence is the public code/data release and reusable task definitions, not an external physical discovery.
  • UnifiedBatch = (x, c1, c2, meta) task interface no independent evidence
    purpose: Map heterogeneous sensor tasks into one conditional-generation batch format for all model families.
    Appendix A introduces this abstraction to make cross-task training comparable; it is an engineering construct of the paper.

pith-pipeline@v1.1.0-grok45 · 30542 in / 3424 out tokens · 42433 ms · 2026-07-11T20:40:06.046679+00:00 · methodology

0 comments
read the original abstract

Generative models have changed how machine learning represents complex data distributions, especially in language and vision, yet many real-world systems are observed instead as continuous, high-dimensional, and noisy sensor time series. Existing generative modeling of sensor data, however, remains fragmented across modalities, datasets, and task formulations, limiting a systematic understanding of when, how, and why generative models succeed or fail in real-world settings. To address this gap, we introduce SensorGen, a large-scale study of sensor-signal generation spanning 14 settings across 4 domains, 7 datasets, and 12 signal modalities. Leveraging SensorGen, we systematically evaluate generative models from five major families and uncover three key findings: (1) flow-matching models provide strong overall performance across most settings; (2) signal properties matter, with demographic covariates improving longitudinal generation and time-frequency modeling improving high-frequency signal generation; and (3) generated signals have practical utility beyond visual realism, with scaling improving generation quality and synthetic data improving downstream performance. Together, SensorGen establishes a broader understanding of design choices, evaluation protocols, and failure modes in real-world sensor data generation.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 14 linked inside Pith

  1. [1]

    Flowecg: Using flow matching to create a more efficient ecg signal generator.arXiv preprint arXiv:2509.10491, 2025

    Vitalii Bondar, Serhii Semenov, Vira Babenko, and Dmytro Holovniak. Flowecg: Using flow matching to create a more efficient ecg signal generator.arXiv preprint arXiv:2509.10491, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  3. [3]

    Capture-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition.Scientific Data, 11(1):1135, 2024

    Shing Chan, Yuan Hang, Catherine Tong, Aidan Acquah, Abram Schonfeldt, Jonathan Gershuny, and Aiden Doherty. Capture-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition.Scientific Data, 11(1):1135, 2024

    2024

  4. [4]

    Versatile cardiovascularsignalgenerationwithaunifieddiffusiontransformer.NatureMachineIntelligence, 8(1):6–19, 2026

    Zehua Chen, Yuyang Miao, Liyuan Wang, Luyun Fan, Danilo P Mandic, and Jun Zhu. Versatile cardiovascularsignalgenerationwithaunifieddiffusiontransformer.NatureMachineIntelligence, 8(1):6–19, 2026

    2026

  5. [5]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  6. [6]

    Self-supervised learning for biomedical signal processing: A systematic review on ecg and ppg signals.medRxiv, pages 2024–09, 2024

    Cheng Ding and Chenwei Wu. Self-supervised learning for biomedical signal processing: A systematic review on ecg and ppg signals.medRxiv, pages 2024–09, 2024

    2024

  7. [7]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024

    2024

  8. [8]

    Ppgflowecg: Latent rectified flow with cross-modal en- coding for ppg-guided ecg generation and cardiovascular disease detection.arXiv preprint arXiv:2509.19774, 2025

    Xiaocheng Fang, Jiarui Jin, Haoyu Wang, Che Liu, Jieyi Cai, Yujie Xiao, Guangkun Nie, Bo Liu, Shun Huang, Hongyan Li, et al. Ppgflowecg: Latent rectified flow with cross-modal en- coding for ppg-guided ecg generation and cardiovascular disease detection.arXiv preprint arXiv:2509.19774, 2025

    arXiv 2025

  9. [9]

    Mimic-iv-ecg: Diagnostic electrocardiogram matched subset.Type: dataset, 2023

    Brian Gow, Tom Pollard, Larry A Nathanson, Alistair Johnson, Benjamin Moody, Chrystinne Fernandes, Nathaniel Greenbaum, Jonathan W Waks, Parastou Eslami, Tanner Carbonati, et al. Mimic-iv-ecg: Diagnostic electrocardiogram matched subset.Type: dataset, 2023

    2023

  10. [10]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 12 Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

    2020

  11. [11]

    Personalized heart disease detection via ecg digital twin generation.arXiv preprint arXiv:2404.11171, 2024

    Yaojun Hu, Jintai Chen, Lianting Hu, Dantong Li, Jiahuan Yan, Haochao Ying, Huiying Liang, and Jian Wu. Personalized heart disease detection via ecg digital twin generation.arXiv preprint arXiv:2404.11171, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Ppg2abp: Translating photoplethysmogram (ppg) signals to arterial blood pressure (abp) waveforms

    Nabil Ibtehaz, Sakib Mahmud, Muhammad EH Chowdhury, Amith Khandakar, Muhammad Salman Khan, Mohamed Arselene Ayari, Anas M Tahir, and M Sohel Rahman. Ppg2abp: Translating photoplethysmogram (ppg) signals to arterial blood pressure (abp) waveforms. Bioengineering, 9(11):692, 2022

    2022

  13. [13]

    Cofe: A framework generating counterfactual ecg for explainable cardiac ai-diagnostics

    Jong-Hwan Jang, Junho Song, and Yong-Yeon Jo. Cofe: A framework generating counterfactual ecg for explainable cardiac ai-diagnostics. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3471–3475. IEEE, 2026

    2026

  14. [14]

    Deep imputation of missing values in time series health data: A review with benchmarking.Journal of biomedical informatics, 144:104440, 2023

    Maksims Kazijevs and Manar D Samad. Deep imputation of missing values in time series health data: A review with benchmarking.Journal of biomedical informatics, 144:104440, 2023

    2023

  15. [15]

    Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Diffusets: 12-lead ecg generation conditioned on clinical text reports and patient-specific information.Patterns, 6(10), 2025

    Yongfan Lai, Jiabo Chen, Qinghao Zhao, Deyun Zhang, Yue Wang, Shijia Geng, Hongyan Li, and Shenda Hong. Diffusets: 12-lead ecg generation conditioned on clinical text reports and patient-specific information.Patterns, 6(10), 2025

    2025

  17. [17]

    Vitaldb, a high-fidelity multi-parameter vital signs database in surgical patients.Scientific Data, 9(1):279, 2022

    Hyung-Chul Lee, Yoonsang Park, Soo Bin Yoon, Seong Mi Yang, Dongnyeok Park, and Chul-Woo Jung. Vitaldb, a high-fidelity multi-parameter vital signs database in surgical patients.Scientific Data, 9(1):279, 2022

    2022

  18. [18]

    Mira: Medical time series foundation model for real-world health data.arXiv preprint arXiv:2506.07584, 2025

    Hao Li, Bowen Deng, Chang Xu, Zhiyuan Feng, Viktor Schlegel, Yu-Hao Huang, Yizheng Sun, Jingyuan Sun, Kailai Yang, Yiyao Yu, et al. Mira: Medical time series foundation model for real-world health data.arXiv preprint arXiv:2506.07584, 2025

    arXiv 2025

  19. [19]

    HEARTS: Benchmarking llm reasoning on health time series.arXiv preprint arXiv:2603.06638, 2026

    Sirui Li, Shuhan Xiao, Mihir Joshi, Ahmed Metwally, Daniel McDuff, Wei Wang, and Yuzhe Yang. HEARTS: Benchmarking llm reasoning on health time series.arXiv preprint arXiv:2603.06638, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Fractal generative models.arXiv preprint arXiv:2502.17437, 2025

    Tianhong Li, Qinyi Sun, Lijie Fan, and Kaiming He. Fractal generative models.arXiv preprint arXiv:2502.17437, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Autoregressive image generation without vector quantization.Advances in Neural Information Processing Systems, 37:56424–56445, 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization.Advances in Neural Information Processing Systems, 37:56424–56445, 2024

    2024

  22. [22]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. InEuropean Conference on Computer Vision, pages 23–40. Springer, 2024

    2024

  23. [23]

    Scaling wearable foundation models.arXiv preprint arXiv:2410.13638, 2024

    Girish Narayanswamy, Xin Liu, Kumar Ayush, Yuzhe Yang, Xuhai Xu, Shun Liao, Jake Garrison, Shyam Tailor, Jake Sunshine, Yun Liu, et al. Scaling wearable foundation models.arXiv preprint arXiv:2410.13638, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Diffecg: A versatile probabilistic diffusion model for ecg signals synthesis

    Nour Neifar, Achraf Ben-Hamadou, Afef Mdhaffar, and Mohamed Jmaiel. Diffecg: A versatile probabilistic diffusion model for ecg signals synthesis. In2024 IEEE/ACIS 22nd International 13 Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series Conference on Software Engineering Research, Management and Applications (SERA), pages 182–

  25. [25]

    Climax: A foundation model for weather and climate.arXiv preprint arXiv:2301.10343, 2023

    TungNguyen, JohannesBrandstetter, AshishKapoor, JayeshKGupta, andAdityaGrover. Climax: A foundation model for weather and climate.arXiv preprint arXiv:2301.10343, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    A diffusion model for inertial based time series generation on scarce data availability to improve human activity recognition.Scientific Reports, 15(1):16841, 2025

    Heiko Oppel and Michael Munz. A diffusion model for inertial based time series generation on scarce data availability to improve human activity recognition.Scientific Reports, 15(1):16841, 2025

    2025

  27. [27]

    Phymer: physiological dataset for multimodal emotion recognition with personality as a context.IEEE Access, 11:107638–107656, 2023

    SudarshanPant, Hyung-JeongYang, EunchaeLim, Soo-HyungKim, andSeok-BongYoo. Phymer: physiological dataset for multimodal emotion recognition with personality as a context.IEEE Access, 11:107638–107656, 2023

    2023

  28. [28]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  29. [29]

    The sleep heart health study: design, rationale, and methods.Sleep, 20(12):1077–1085, 1997

    Stuart F Quan, Barbara V Howard, Conrad Iber, James P Kiley, F Javier Nieto, George T O’Connor, David M Rapoport, Susan Redline, John Robbins, Jonathan M Samet, et al. The sleep heart health study: design, rationale, and methods.Sleep, 20(12):1077–1085, 1997

    1997

  30. [30]

    Deep ppg: Large-scale heart rate estimation with convolutional neural networks.Sensors, 19(14):3079, 2019

    Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof Van Laerhoven. Deep ppg: Large-scale heart rate estimation with convolutional neural networks.Sensors, 19(14):3079, 2019

    2019

  31. [31]

    Photoreal- istic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photoreal- istic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    2022

  32. [32]

    Region-disentangled diffusion model for high-fidelity ppg-to-ecg translation

    Debaditya Shome, Pritam Sarkar, and Ali Etemad. Region-disentangled diffusion model for high-fidelity ppg-to-ecg translation. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 15009–15019, 2024

    2024

  33. [33]

    OSF: On pre-training and scaling of sleep foundation models.arXiv preprint arXiv:2603.00190, 2026

    Zitao Shuai, Zongzhe Xu, David Yang, Wei Wang, and Yuzhe Yang. OSF: On pre-training and scaling of sleep foundation models.arXiv preprint arXiv:2603.00190, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  35. [35]

    Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Generating realistic neurophysiological time series with denoising diffusion probabilistic models.Patterns, 5(9), 2024

    Julius Vetter, Jakob H Macke, and Richard Gao. Generating realistic neurophysiological time series with denoising diffusion probabilistic models.Patterns, 5(9), 2024

    2024

  37. [37]

    Ptb-xl, a large publicly available electrocardiography dataset

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset. Scientific data, 7(1):154, 2020

    2020

  38. [38]

    From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803, 2025

    Fuying Wang, Jiacheng Xu, and Lequan Yu. From token to rhythm: A multi-scale approach for ecg-language pretraining.arXiv preprint arXiv:2506.21803, 2025. 14 Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

    Pith/arXiv arXiv 2025

  39. [39]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600– 612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600– 612, 2004

    2004

  40. [40]

    Metabonet: The largest publicly available consolidated data set for type 1 diabetes management

    Miriam Kopperstad Wolff, Peter Calhoun, Eleonora Maria Aiello, Yao Qin, and Sam Fox Royston. Metabonet: The largest publicly available consolidated data set for type 1 diabetes management. Journal of Diabetes Science and Technology, page 19322968261441637, 2026

    2026

  41. [41]

    SleepLM: Natural-language intelligence for human sleep.arXiv preprint arXiv:2602.23605, 2026

    Zongzhe Xu, Zitao Shuai, Eideen Mozaffari, Ravi S Aysola, Rajesh Kumar, and Yuzhe Yang. SleepLM: Natural-language intelligence for human sleep.arXiv preprint arXiv:2602.23605, 2026

    Pith/arXiv arXiv 2026

  42. [42]

    Simper: Simple self-supervised learning of periodic targets

    Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  43. [43]

    Artificial intelligence-enabled detection and assessment of parkinson’s disease using nocturnal breathing signals.Nature Medicine, 28(10):2207–2215, 2022

    Yuzhe Yang, Yuan Yuan, Guo Zhang, Hao Wang, Ying-Cong Chen, Yingcheng Liu, Christopher G Tarolli, Daniel Crepeau, Jan Bukartyk, Mithri R Junna, et al. Artificial intelligence-enabled detection and assessment of parkinson’s disease using nocturnal breathing signals.Nature Medicine, 28(10):2207–2215, 2022

    2022

  44. [44]

    Normalizing flows are capable generative models.arXiv preprint arXiv:2412.06329, 2024

    Shuangfei Zhai, Ruixiang Zhang, Preetum Nakkiran, David Berthelot, Jiatao Gu, Huangjie Zheng, Tianrong Chen, Miguel Angel Bautista, Navdeep Jaitly, and Josh Susskind. Normalizing flows are capable generative models.arXiv preprint arXiv:2412.06329, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Thenationalsleepresearchresource: towardsa sleep data commons.Journal of the American Medical Informatics Association, 25(10):1351–1358, 2018

    Guo-Qiang Zhang, Licong Cui, Remo Mueller, Shiqiang Tao, Matthew Kim, Michael Rueschman, SaraMariani, DanielMobley, andSusanRedline. Thenationalsleepresearchresource: towardsa sleep data commons.Journal of the American Medical Informatics Association, 25(10):1351–1358, 2018

    2018

  46. [46]

    Sensorlm: Learning the language of wearable sensors.arXiv preprint arXiv:2506.09108, 2025

    Yuwei Zhang, Kumar Ayush, Siyuan Qiao, A Ali Heydari, Girish Narayanswamy, Maxwell A Xu, Ahmed A Metwally, Shawn Xu, Jake Garrison, Xuhai Xu, et al. Sensorlm: Learning the language of wearable sensors.arXiv preprint arXiv:2506.09108, 2025. 15 Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series A. Data Details A.1. Unified Da...

    Pith/arXiv arXiv 2025