TinyCardioUNet: IMU-to-ECG Translation with Graph-Encoded Inter-Axis Dependencies and Tensor Decomposition-Based Parameter Reduction
Seungwoo Han, Ingon Chanpornpakdi, Motoi Noda, Puwadej Leelasiri, Ibuki Hiruma, Toshihisa Tanaka
cs.LG
Quick verdict
Promising 36k-param IMU-to-ECG model with graph bottleneck and tensor compression, but the graph ablation is confounded and the test set spans only 3 subjects.
Summary
The paper proposes TinyCardioUNet, a lightweight UNet that reconstructs ECG from all six axes of a chest-worn IMU (three-axis accelerometer SCG plus three-axis gyroscope GCG) without prior channel selection. A single GraphSAGE layer at the UNet bottleneck treats channels as graph nodes — six of them connected as a complete graph representing the IMU axes — to exchange inter-axis information, followed by temporal ℓ2 normalization. Post-training compression applies Tucker-2 decomposition to selected conv layers and SVD to the graph weights, with ranks chosen automatically by variational Bayesian matrix factorization, then brief fine-tuning. On the public 'Mechanocardiograms with ECG Reference' dataset (29 healthy males; test set of only 3 volunteers, following Tapotee et al.'s split), the compressed model reaches RMSE 0.098 and PCC 0.677 with 36.0k parameters (72.6% parameter and 59% FLOPs reduction), and degrades gracefully under calibrated additive noise (PCC 0.677→0.606 at SNR 10 dB). It outperforms CGAN, UNet, WaveNet, and WaveUNet baselines on RMSE and parameter count. However, the graph ablation confounds the graph layer with ℓ2 normalization, the no-graph variant paradoxically improves as noise increases (unexplained), bottleneck channels no longer map to physical axes after conv mixing, absolute fidelity trails prior work on the same data (Tapotee RMSE 0.080, PCCHR 0.941 vs 0.814), and no code is released.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
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 bottleneck GraphSAGE layer encoding the six IMU axes as a complete graph lets a UNet use all six axes without channel selection, raising PCC from 0.380 to 0.622.
- VBMF-ranked Tucker-2/SVD tensor decomposition compresses the model by 72.6% in parameters and 59% in FLOPs (to 36,004 parameters), and brief fine-tuning recovers and slightly surpasses uncompressed accuracy across noise levels.
- The 36k-parameter compressed model degrades only from PCC 0.677 to 0.606 between clean input and SNR 10 dB, showing greater noise robustness than much larger CGAN, UNet, WaveNet, and WaveUNet baselines.
- TinyCardioUNet outperforms all four baselines on RMSE and parameter count under a shared training protocol on the public mechanocardiography dataset.
Strengths
Extreme parameter efficiency with competitive accuracy
After Tucker-2+VBMF compression the model runs at 36,004 parameters and 27.99M FLOPs (Table 3) with clean RMSE 0.098 / PCC 0.677 (Table 2), beating the 13.6M-parameter CGAN (RMSE 0.200) and the 48,257-parameter WaveNet (RMSE 0.143) on fidelity — roughly two orders of magnitude fewer parameters than the CGAN baseline.
Honest positioning against prior work on the same dataset
The Discussion openly reports that Tapotee et al. [10] achieve better fidelity on the same data (MAE 0.054, RMSE 0.080, PCCHR 0.941 vs. this paper's 0.068/0.098/0.814) and identifies [10]'s channel selection as test-set-informed — a real methodological flaw — rather than hiding the unfavorable comparison.
Rigorous noise protocol and metric hygiene
Channel-selection-free design targets a real generalization gap
Transparent, internally consistent compression accounting
Weaknesses
Ablation conflates the graph layer with temporal ℓ2 normalization
Unexplained non-monotonic noise behavior in the no-graph variant
Bottleneck channel–axis correspondence is conceptually unjustified
Three-subject test set with weak statistics
Modest absolute fidelity and ambiguous model selection
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This four-page conference-style paper (Tokyo University of Agriculture and Technology; arXiv 2609.29322) proposes **TinyCardioUNet**, a lightweight UNet translating chest-worn IMU signals — three-axis seismocardiography (SCG) plus three-axis gyrocardiography (GCG) — into a Lead-II ECG waveform. Its three central claims: (1) a single GraphSAGE layer at the UNet bottleneck, encoding the six IMU axes as a complete graph, permits using **all six axes without prior channel selection** and lifts PCC from 0.380 to 0.622 (Table 2); (2) Tucker-2/SVD tensor decomposition with automatic VBMF rank selection compresses the model by **72.6% in parameters and 59% in FLOPs** (131,361 → 36,004 parameters; 68.89M → 27.99M FLOPs, Table 3), with brief fine-tuning recovering and slightly exceeding uncompressed accuracy; (3) the compressed model stays robust under calibrated additive noise (PCC 0.677 → 0.606 from clean to SNR 10 dB). Positioning against Tapotee et al. [10] on the same dataset is unusually honest: the Discussion concedes their better fidelity (RMSE 0.080 vs 0.098; PCCHR 0.941 vs 0.814) while arguing their channel choice was test-set-informed.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 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
How accurate is ECG reconstruction from a chest-worn IMU?
In this paper, a 36k-parameter model (TinyCardioUNet) reconstructs ECG from chest IMU signals with RMSE 0.098 and a Pearson correlation of 0.677 against ground truth on clean data from healthy male volunteers, dropping to 0.606 correlation when input noise equals the signal at SNR 10 dB. That is enough to track heart rate trends, but fidelity remains below larger prior models on the same dataset (e.g., RMSE 0.080, heart-rate correlation 0.941), so it is not a clinical-grade ECG replacement.
Can ECG be measured without electrodes?
Not directly, but a chest-worn inertial measurement unit captures cardiac vibrations: the accelerometer signal is called seismocardiography (SCG) and the gyroscope signal gyrocardiography (GCG). Deep learning models like the one in this paper translate these mechanical signals into an ECG-like waveform, enabling electrode-free, discomfort-free heart-rate monitoring; the electrical information itself is inferred, not measured.
How small is TinyCardioUNet compared to other ECG reconstruction models?
After tensor-decomposition compression it uses 36,004 parameters and about 28M FLOPs per signal segment — roughly 380 times fewer parameters than the 13.6M-parameter CGAN baseline and 72.6% fewer parameters (59% fewer FLOPs) than its own uncompressed 131k-parameter version. The target is edge hardware such as a Raspberry Pi Pico-class microcontroller.
What dataset was used, and is the code available?
The public 'Mechanocardiograms with ECG Reference' dataset (Kaisti et al., IEEE Sensors Journal 2019): simultaneous SCG, GCG, and Lead-II ECG from 29 healthy men, about 260 minutes recorded at 800 Hz in supine and side-lying positions; volunteers 11, 21, and 24 form the test set. The paper does not release code or the selected tensor ranks.
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.
- Figure 2 waveform overlays cannot be judged from extracted text; the visual claims about baseline drift and invalid generations in CGAN/UNet/WaveNet/WaveUNet rest on the authors' description alone.
- Per-layer VBMF-selected ranks, random seeds, and any supplementary details are absent from the extracted text, so reproducibility judgments rely on what is stated in §2.4.
- No external verification of cited prior-work numbers (e.g., Tapotee et al.'s RMSE 0.080 / PCCHR 0.941) was performed; this review takes the paper's reporting at face value.
- arXiv HTML extraction may have dropped minor typographic details; all quantitative claims above were cross-checked against the tables as extracted and are internally consistent.
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.