Source-linked AI summary
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
TL;DR
Electrode discomfort motivates estimating ECG from chest-worn IMU signals for continuous heart-rate monitoring. TinyCardioUNet combines all six IMU axes through graph-based bottleneck modeling and tensor-decomposition compression, achieving competitive benchmark results with a compact model and comparatively robust noise performance.
Problem
Long-term ECG electrode wear can cause discomfort, while prior IMU-to-ECG methods either use large models or select a fixed subset of axes.
Source: §1 · p. 1Method
TinyCardioUNet uses all six IMU axes, models their dependencies with a bottleneck graph neural network, and compresses parameters using tensor decomposition.
Source: §1 · p. 1, §2.3 · p. 2, §2.4 · p. 2Results
Against WaveNet, TinyCardioUNet lowers no-noise RMSE by 31% and raises PCC from 0.597 to 0.677; against CGAN, UNet, and WaveUNet, it reduces no-noise RMSE by 51–57%.
Source: §3.2 · p. 3Takeaways & Limitations
With 36.0k parameters, TinyCardioUNet's PCC declines from 0.677 without noise to 0.606 at SNR 10 dB, supporting its potential for resource-constrained multichannel signal translation.
Source: §4 · p. 4, §5 · p. 4The evaluation used healthy males rather than cardiac patients, and the graph structure may not be optimal.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Estimating electrocardiography (ECG) from a chest-worn inertial measurement unit (IMU) enables continuous heart rate (HR) monitoring without the discomfort of electrodes. We propose TinyCardioUNet, a lightweight UNet that uses all six IMU axes without prior channel selection, refines its bottleneck with a graph neural network that encodes inter-axis dependencies, and employs tensor decomposition with automatic variational Bayesian rank selection for parameter reduction. On a public dataset, TinyCardioUNet achieves an RMSE of $0.098$ and a Pearson correlation coefficient of $0.677$ with only $36.0$k parameters and remains comparatively robust to additive noise, demonstrating accurate ECG reconstruction with a compact model.
1. INTRODUCTION
Continuous HR monitoring matters for stress management and early heart-disease prevention, but electrode discomfort motivates chest-worn IMU alternatives. TinyCardioUNet uses all six IMU axes, graph-based inter-axis modeling, and tensor decomposition to target compact, noise-robust ECG reconstruction.
Motivation
- Continuous HR monitoring supports stress management and early heart-disease prevention, while long-term ECG electrode wear can cause discomfort and skin irritation.ECG is the clinical gold standard for HR monitoring; chest-worn IMUs offer an electrode-free alternative by recording cardiac vibrations and rotational dynamics.Source: §1 · p. 1
Research gap
- Prior IMU-to-ECG approaches either used a CGAN with approximately 13.6M generator parameters or selected only two of six IMU axes.The fixed-axis approach may be less robust when signal quality varies across subjects or measurement conditions.Source: §1 · p. 1
Proposed approach
- TinyCardioUNet models dependencies among all six IMU axes as a graph and refines the UNet bottleneck with a graph neural network.The proposed design aims to let physically coupled axes complement information from a noise-corrupted axis while reducing model size through tensor decomposition.Source: §1 · p. 1
2. MATERIALS AND METHODS
The methods use a public dataset of simultaneous chest-worn IMU and Lead II ECG recordings, with fixed volunteer-level train/test separation and signal preprocessing. TinyCardioUNet combines an encoder–graph bottleneck–decoder architecture with post-training tensor decomposition and fine-tuning.
2.1. Dataset
- The dataset contains simultaneous SCG, GCG, and Lead II ECG recordings from 29 healthy male volunteers, totaling roughly 260 minutes.Recordings were acquired at 800 Hz in supine and left and right lateral decubitus positions.Source: §2.1 · p. 1
2.2. Data preprocessing
- Volunteers 11, 21, and 24 form the test set, while the remaining 26 form the training set.Signals were downsampled to 256 Hz, filtered and baseline-corrected, then segmented into 1024-sample windows with 50% overlap and normalized.Source: §2.2 · p. 2
2.3. TinyCardioUNet architecture
- The UNet encoder uses four Conv1D blocks with 16, 32, 64, and 128 output channels, reducing temporal length to N/8 before the bottleneck.The graph module applies a mean-aggregator GraphSAGE layer; its first six nodes represent the IMU axes and form a complete graph.Source: §2.3 · p. 2
- Three decoder upsampling stages restore temporal resolution, and skip connections link decoder features to corresponding encoder features.A terminal 1 × 1 convolution outputs the reconstructed ECG; training minimizes MSE between reconstructed and ground-truth ECG segments.Source: §2.3 · p. 2, Figure 1 · p. 2
2.4. Model compression using tensor decomposition
- VBMF automatically selects ranks for Tucker-2 convolutional and Tucker-1 GraphSAGE weight decompositions, replacing layers with smaller factorized maps.Ranks are capped at 90% of original dimensions; trained weights are decomposed after convergence and fine-tuned for five or ten epochs.Source: §2.4 · pp. 2–3
3. RESULTS
TinyCardioUNet outperforms the benchmark models on no-noise waveform metrics while using fewer parameters and, versus WaveNet, fewer FLOPs. Graph refinement improves PCC across noise levels, and tensor decomposition reduces model cost while fine-tuning recovers or surpasses uncompressed accuracy.
3.1. Evaluation metrics
- The evaluation reports MAE, RMSE, and PCC for reconstructed versus reference ECG, plus HR agreement and model parameters and FLOPs.Metrics are computed per test segment and reported as mean±standard deviation; PCC is computed at zero lag, while HR agreement compares average HR over four-second windows. Test inputs also include additive noise at SNR 20, 15, and 10 dB.Source: §3.1 · p. 3
3.2. Ablations and benchmark comparison
- Against WaveNet, TinyCardioUNet (TD + FT@10epoch) lowers no-noise RMSE by 31% and raises PCC from 0.597 to 0.677; against CGAN, UNet, and WaveUNet, it reduces no-noise RMSE by 51–57%.The comparisons also report fewer parameters than all four benchmarks and roughly 3.5× fewer FLOPs than WaveNet.Source: §3.2 · p. 3
- Graph refinement raises PCC across all noise levels and lowers MAE and RMSE except at SNR 10 dB, with Table 2 covering no noise and SNR 20, 15, and 10 dB.The compared models share an encoder–decoder, and the reported improvement is attributed to the GraphSAGE layer and temporal ℓ2 normalization enabling inter-axis information exchange.Source: §3.2 · p. 3, Table 2 · p. 4
- Across a representative segment, TinyCardioUNet recovers QRS complexes and their timing at SNR 20, 15, and 10 dB, while the other models show waveform degradation or invalid generation.Figure 2 shows the reconstructed signals alongside the three-axis SCG and GCG inputs and the reference ECG.Source: §3.2 · p. 3, Figure 2 · p. 4
- 72.6% fewer parameters and 59% fewer FLOPs result from VBMF-ranked Tucker-2 and SVD decomposition of the full model.Fine-tuning the decomposed model recovers and surpasses uncompressed accuracy across noise levels; 10 epochs gives lowest waveform error in clean and moderate-noise conditions, while five epochs gives highest PCC at every noise level and least heavy-noise degradation.Source: §3.2 · p. 3, Table 3 · p. 4
4. DISCUSSION
TinyCardioUNet combines a 36.0k-parameter model with comparatively modest PCC degradation across tested noise levels. The discussion compares its no-noise results with Tapotee et al. and notes limits in dataset diversity and graph design.
4. DISCUSSION
- With 36.0k parameters, TinyCardioUNet's PCC (TD + FT@10epoch) declines only from 0.677 to 0.606 between no noise and SNR 10 dB, while larger models degrade more.The compared models are CGAN, UNet, WaveNet, and WaveUNet.Source: §4 · p. 4
- Under no noise, TinyCardioUNet reports MAE 0.068±0.046, RMSE 0.098±0.047, and PCCHR 0.814, versus Tapotee et al. 's 0.054±0.026, 0.080 ± 0.034, and 0.941.The respective metric pairs are MAE, RMSE, and PCCHR.Source: §4 · p. 4
- Unlike Tapotee et al., TinyCardioUNet uses all six IMU axes without channel selection; their selected inputs were not evaluated under controlled varying-noise conditions.Tapotee et al.'s channel combination was selected based on test-set performance and noise characteristics.Source: §4 · p. 4
- The evaluation used healthy males rather than cardiac patients, and the proposed graph structure may not be optimal.The authors identify dataset diversity and domain-specific graph design as areas for further work.Source: §4 · p. 4
5. CONCLUSION
TinyCardioUNet is a lightweight UNet that uses all six IMU axes and models their inter-axis dependencies with a bottleneck GraphSAGE layer. Tensor decomposition and brief fine-tuning reduce parameters and computational cost while retaining competitive reconstruction accuracy.
5. CONCLUSION
- TinyCardioUNet uses all six IMU axes without prior channel selection and models inter-axis dependencies via a bottleneck GraphSAGE layer.The conclusion describes the architecture as a lightweight UNet.Source: §5 · p. 4
- Tensor decomposition with brief fine-tuning substantially reduces parameters and computational cost while retaining competitive reconstruction accuracy.Source: §5 · p. 4
- The results highlight the potential of graph-based inter-axis modeling for resource-constrained multichannel physiological signal translation.Source: §5 · p. 4