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.LGSep 24, 2026

TinyCardioUNet: IMU-to-ECG Translation with Graph-Encoded Inter-Axis Dependencies and Tensor Decomposition-Based Parameter Reduction

Authors: Seungwoo Han, Ingon Chanpornpakdi, Motoi Noda, Puwadej Leelasiri, Ibuki Hiruma, Toshihisa Tanaka

Organizations: Department of Electrical Engineering and Computer Science, Tokyo University of Agriculture and Technology, Japan

Abstract

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.

Figures & tables

Block name	Size $(D,S,T)$	Structure	Output shape	TD
Encoder1	$(3,6,16)$	Conv1D–BN–ReLU	( $B$ , 16, $N$ )	–
Encoder2	$(3,16,32)$	Pool–Conv1D–BN–ReLU	( $B$ , 32, $N/2$ )	–
Encoder3	$(3,32,64)$	Pool–Conv1D–BN–ReLU	( $B$ , 64, $N/4$ )	–
Encoder4	$(3,64,128)$	Pool–Conv1D–BN–ReLU	( $B$ , 128, $N/8$ )	✓
GraphSAGE	$(-,128,128)$	GraphSAGE	( $B$ , 128, $N/8$ )	✓
UpConv1	$(3,128,64)$	Upsample–Conv1D–BN–ReLU	( $B$ , 64, $N/4$ )	✓

Table 1: TinyCardioUNet parameters and structure.

Model / Metric	No noise			SNR 20			SNR 15			SNR 10
Model / Metric	MAE $\downarrow$	RMSE $\downarrow$	PCC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	PCC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	PCC $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	PCC $\uparrow$
CGAN [ 20 ]	0.156 $\pm$ 0.047	0.200 $\pm$ 0.050	0.190 $\pm$ 0.059	0.162 $\pm$ 0.046	0.207 $\pm$ 0.049	0.179 $\pm$ 0.059	0.173 $\pm$ 0.043	0.222 $\pm$ 0.047	0.158 $\pm$ 0.057	0.207 $\pm$ 0.038	0.264 $\pm$ 0.043	0.117 $\pm$ 0.052
UNet [ 18 ]	0.205 $\pm$ 0.092	0.230 $\pm$ 0.082	0.555 $\pm$ 0.186	0.203 $\pm$ 0.091	0.227 $\pm$ 0.081	0.551 $\pm$ 0.181	0.199 $\pm$ 0.086	0.226 $\pm$ 0.076	0.532 $\pm$ 0.170	0.210 $\pm$ 0.069	0.239 $\pm$ 0.060	0.455 $\pm$ 0.151
WaveNet [ 17 ]	0.118 $\pm$ 0.065	0.143 $\pm$ 0.067	0.597 $\pm$ 0.169	0.123 $\pm$ 0.063	0.150 $\pm$ 0.063	0.567 $\pm$ 0.161	0.132 $\pm$ 0.057	0.159 $\pm$ 0.058	0.521 $\pm$ 0.151	0.160 $\pm$ 0.063	0.187 $\pm$ 0.061	0.388 $\pm$ 0.119
WaveUNet [ 21 ]	0.203 $\pm$ 0.101	0.219 $\pm$ 0.096	0.528 $\pm$ 0.222	0.213 $\pm$ 0.107	0.229 $\pm$ 0.102	0.507 $\pm$ 0.227	0.238 $\pm$ 0.120	0.254 $\pm$ 0.115	0.459 $\pm$ 0.230	0.296 $\pm$ 0.108	0.310 $\pm$ 0.101	0.343 $\pm$ 0.196
TinyCardioUNet (baseline, ours)	0.093 $\pm$ 0.055	0.123 $\pm$ 0.056	0.622 $\pm$ 0.208	0.101 $\pm$ 0.058	0.128 $\pm$ 0.058	0.616 $\pm$ 0.208	0.118 $\pm$ 0.064	0.144 $\pm$ 0.063	0.584 $\pm$ 0.215	0.194 $\pm$ 0.074	0.215 $\pm$ 0.069	0.453 $\pm$ 0.197

Table 2: Performance comparison across noise conditions (No noise, SNR

\{20,15,10\}

dB). Values are mean

\pm

standard deviation. All results under our experimental setup.

Model	#Param. $\downarrow$	FLOPs $\downarrow$
CGAN [ 20 ]	13.6M ${}^{\text{A}}$	614.11M ${}^{\text{A}}$
UNet [ 18 ]	10.8M	2,545.68M
WaveNet [ 17 ]	48,257	98.17M
WaveUNet [ 21 ]	3.4M	293.81M
TinyCardioUNet (baseline, ours)	131,361	68.89M
TinyCardioUNet w/o graph module (ours)	98,465	56.28M

Table 3: Computational cost comparison.

Explore similar work

ELF: A Family of Encoder-Free ECG-Language Models
Jan 5, 2026William Han, Tony Chen, Chaojing Duan +6Electrocardiography Cnn-Lstm
Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA +1
How Do Electrocardiogram Models Scale?
May 17, 2026Jiawei Li, Fabio Bonassi, Ming Jin +4Electrocardiogram Foundation Models Medical World Model
Uppsala University · Griffith University
ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment
Jul 6, 2026Xiaocheng Fang, Haoyu Wang, Jieyi Cai +14Adult-To-Pediatric Electrocardiogram Transfer Cardiac
National Institute of Health Data Science, Peking University, Beijing, China · School of Intelligence Science and Technology, Peking University, Beijing, China · Institute of Medical Technology, Peking University Health Science Center, Beijing, China +6

Jan 5, 2026, cs.MMJ/K move · Enter open · S save

ELF: A Family of Encoder-Free ECG-Language Models

William Han, Tony Chen, Chaojing Duan, Xiaoyu Song+5

Carnegie Mellon University, USA · Allegheny Health Network, USA · University of California, Los Angeles, USA+1

Electrocardiography Cnn-Lstm Emotion Recognition

Open paper AskPDF arXiv

ECG-Language Models (ELMs) extend recent advances in Multimodal Large Language Models (MLLMs) to automated ECG interpretation. However, most existing ELMs inherit Vision-Language Model (VLM) design choices and rely on pretrained ECG encoders, introducing substantial architectural and training complexity. Inspired by encoder-free VLMs, we introduce ELF, a family of three encoder-free ELMs that remain competitive with, and often outperform, prior state-of-the-art ELMs across two datasets despite substantially simpler architectures and training pipelines. All code and data are available at github.com/ELM-Research/ECG-Language-Models.

Abstract

Explore similar work

ELF: A Family of Encoder-Free ECG-Language Models

How Do Electrocardiogram Models Scale?

ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment