SOTER: A Generative Time-Series Foundation Model for Wearable Human Physiological Signals
Fangke Chen, Fangke Chen, Sirry Chen, Sirry Chen, Wei Chen, Wei Chen, Zhongyu Wei
cs.LG · cs.AI
Quick verdict
SOTER claims zero-shot wins for wearable-signal foundation modeling over generalist TSFMs; evidence is broad and ablated but lacks significance tests.
Summary
SOTER is a generative time-series foundation model built specifically for wearable physiological signals (ECG, PPG, EDA, etc.). Its architecture combines three components matched to the authors' three stated properties of physiological data: lower backbone layers process each channel independently with continuous-time rotary positional encodings while a single top layer applies cross-channel attention; a mixture-of-experts feed-forward layer routes tokens via a non-learned, deterministic rule based on a causal prefix-DFT spectral statistic (top-2 of band-tied experts plus a shared expert); and a Neural CDE decoder evolves the latent state to arbitrary query timestamps, reducing exactly to a time-augmented Neural ODE without a control path (proven in Appendix D). The model is pre-trained for one epoch on roughly 226 billion time points from five public datasets (49.79M parameters, 20.29M active; 8×H200, ~312h). On six out-of-distribution wearable datasets under leave-one-subject-out splits, SOTER achieves the best RMSE on 4/6 and MAE on 5/6 in zero-shot forecasting (losing MMASH to TimesFM 2.5 and IEEE PPG RMSE to TimeMoE-Large), the highest average Macro-AUROC (80.79) in frozen linear-probe classification, and the lowest imputation error on all six datasets at 75% missingness. A corpus-matched comparison against MIRA-Large supports that the corpus alone does not explain the gains. Limitations: no significance testing, corpus matching restricted to MIRA on two forecasting datasets, and 6.09 s/sample inference latency.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single frozen SOTER backbone transfers zero-shot across forecasting, linear-probe classification, and causal imputation on six out-of-distribution wearable datasets.
- A non-learned PSD-guided routing rule (causal prefix DFT over latent representations) ties experts to fixed spectral bands and beats a learned router with auxiliary balancing loss (Table 3).
- The corpus alone does not explain SOTER's advantage: corpus-matched MIRA-Large (continued) still trails on both forecasting datasets tested (Table 4).
- SOTER is robust to additive acquisition noise, matching or beating clean-input baselines at the strongest corruption β=0.5 (Figure 4).
Strengths
Exact CDE→ODE reduction with formal proof
Appendix D proves the forecasting decoder is an algebraic identity of the imputation decoder under zero control (Picard–Lindelöf well-posedness argued), and candidly reclassifies the mechanism as control-as-input rather than bilinear Neural CDE — unusually careful terminology for a TSFM paper.
Corpus-matched control experiment
Section 4.5 / Table 4 re-trains MIRA-Large on the same 226B-point corpus (scratch and continued) under the same budget; continued pre-training improves MIRA (MIT-BIH RMSE 2.20→1.79) yet still trails SOTER (1.29), partially separating corpus effects from architecture.
Thorough ablation and sensitivity work
Honest reporting of failure cases
Above-average reproducibility detail
Weaknesses
No statistical significance testing
Corpus-vs-architecture control is narrow
Inference latency undercuts the efficiency framing
Corpus accounting does not close
AUROC emphasis sidesteps weak operational F1
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SOTER (arXiv 2609.16804) proposes a domain-specialized generative time-series foundation model for wearable physiological signals, pre-trained for one epoch on ~226B time points from five public datasets (MIMIC-III Waveform, Sleep-EDF, PTB-XL, WESAD, Chapman-ECG; Appendix A). Three central claims: (1) an architecture jointly modeling cross-channel coupling, spectral specialization, and continuous-time dynamics outperforms generalist and medical TSFMs in zero-shot transfer; (2) a single frozen backbone covers forecasting, linear-probe classification, and causal imputation; (3) component ablations plus a corpus-matched MIRA comparison attribute the gains to architecture rather than corpus alone.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SOTER and what is it designed for?
SOTER is a generative time-series foundation model (49.79M parameters, 20.29M active per step) pre-trained on about 226 billion time points from five public physiological datasets (MIMIC-III Waveform, Sleep-EDF, PTB-XL, WESAD, Chapman-ECG). It targets wearable signals such as ECG, PPG, and EDA, and supports zero-shot forecasting, frozen-encoder classification, and causal online imputation from one backbone, using a channel-independent backbone with one cross-channel attention layer, a deterministic PSD-guided mixture-of-experts router, and a Neural CDE/ODE decoder.
How does SOTER compare to TimesFM, Chronos-2, or TimeMoE on wearable data?
In the paper's zero-shot forecasting benchmark on six wearable datasets (Table 1), SOTER has the best RMSE on 4 of 6 and best MAE on 5 of 6 datasets. It loses to TimesFM 2.5 on MMASH (0.08 vs 1.26 scaled RMSE) and to TimeMoE-Large on IEEE PPG RMSE (1.77 vs 2.19). TimeMoE-Large is generally its strongest competitor; SOTER wins most head-to-heads while activating far fewer parameters, but needs 6.09 s per sample due to its ODE solver.
Is SOTER's code and training data publicly available?
The paper states source code is available at github.com/SII-fkchen/SOTER, and all five pre-training datasets plus six evaluation datasets are public (PhysioNet, Monash archive, etc.). Evaluation uses leave-one-subject-out splits with normalization fitted only on training splits, so a released checkpoint plus the described protocol should allow third-party re-evaluation.
What are SOTER's main limitations?
Per this review: (1) no statistical significance testing — the forecasting table's error bars reflect window configurations, not seeds; (2) the corpus-vs-architecture control covers only MIRA-Large on two forecasting datasets, while the ECG-heavy pre-training corpus overlaps in modality with two evaluation sets; (3) inference is slow (6.09 s/sample); (4) the corpus size (226B time points) is not itemized per dataset; (5) it underperforms on the low-complexity MMASH dataset and on Macro-F1 for imbalanced classification.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2–7 could not be inspected (image content); imputation and noise-robustness claims were cross-checked against the numeric Tables 12–14 and in-text statements instead.
- The GitHub repository's contents and checkpoint availability could not be verified from the text; codeAvailable reflects the paper's stated link only.
- Per-dataset itemization of the ~226B time points is absent, so the corpus-scale claim is inferred (≈3.3 channels/sample) rather than verified.
- The metadata abstract was truncated but the full extracted text, including appendices and tables, was available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.