EcoScale - Net: A Lightweight Multi - Kernel Network for Long - Sequence 12-lead ECG Classification

Dong-Hyeon Kang, Ju-Hyeon Nam, Sang-Chul Lee · 2025

Read EcoScale - Net: A Lightweight Multi - Kernel Network for Long - Sequence 12-lead ECG Classification by Dong-Hyeon Kang, Ju-Hyeon Nam, Sang-Chul Lee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-24748.txt.

Introduction

Dong-Hyeon Kang, Ju-Hyeon Nam, and Sang-Chul Lee built a lean net to sort long twelve lead heart traces. Hand reads of such traces can err, and past nets had to pick a fixed view width that may not fit all trace lengths. A past all scale net fixed that by trying many widths, but at huge cost.

Their net, EcoScaleNet, caps the width at each stage to what is still needed after the trace shrinks. Tight squeeze steps keep the channel count in check and join traits across scales. On a large trace set, it cut weights by most of the past net's count and cost by far more, while its score rose.

Catalog record

CREATORDong-Hyeon Kang, Ju-Hyeon Nam, Sang-Chul Lee
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-24748

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

EcoScale - Net: A Lightweight Multi - Kernel Network for Long - Sequence 12-lead ECG Classification Dong-Hyeon Kang, Ju-Hyeon Nam, Sang-Chul Lee Abstract Accurate interpretation of 12 - lead electrocardiograms (ECGs) is critical for early detection of cardiac abnormalities, yet manual reading is error - prone and existing CNN - based classifiers struggle to choose receptive - field sizes that generalize to the long sequences typical of ECGs. Omni - Scale CNN (OS - CNN) addresses this by enumerating prime - sized kernels inspired by Goldbach’s conjecture to cover every scale, but its exhaustive design explodes computational cost and blocks deeper, wider models. We present E fficient C onvolutional O mni- Scale Net work ( EcoScale-Net ) , a hierarchical variant that retains full receptive - field coverage while eliminating redundancy. At each stage, the maximum kernel length is capped to the scale still required after down - sampling, and 1 × 1 1\times 1 bottleneck convolutions inserted before and after every Omni - Scale block curtail channel growth and fuse multi - scale features. On the large - scale CODE - 15% ECG dataset, EcoScale-Net reduces parameters by 90% and FLOPs by 99% compared with OS - CNN, while raising macro - averaged F1 - score by 2.4%. These results demonstrate that EcoScale-Net delivers state - of - the - art accuracy for long - sequence ECG classification at a fraction of the computational cost, enabling real - time deployment on commodity hardware. Our EcoScale-Net code is available in GitHub Link . 1 Introduction Cardiovascular disease is the leading global cause of death, and 12 - lead electrocardiograms (ECGs) remain the most widely used non - invasive modality for detecting cardiac abnormalities. Manual interpretation, however, is error - prone owing to noise and inter - observer variability [ 1 ] . These constraints have accelerated interest in automated ECG analysis based on deep learning, a trend already validated in other medical - vision domains [ 2 ] . Early ECG classifiers combined convolutional neural networks (CNNs) with recurrent units to exploit both local morphology and long - range rhythm cues, achieving super - human accuracy in some tasks [ 3 ] . Subsequent work expanded to multi - lead, multi - label settings [ 4 ] and even image - based ECG interpretation [ 5 ] . Yet these architectures still depend on manually tuned kernel sizes or dilations, which scale poorly to the long sequences characteristic of ECG signals. Omni - Scale CNN (OS - CNN) [ 6 ] sidesteps manual tuning by enumerating prime - sized kernels—via Goldbach’s conjecture—to cover every receptive - field length. Unfortunately, this exhaustive design inflates parameters and FLOPs quadratically with input length, throttling model depth and width; follow - ups such as OSGAN [ 7 ] and TF - Net [ 8 ] inherit the same bottleneck. To address these issues, we introduce E fficient C onvolutional O mni- Scale Net work ( EcoScale-Net ) , a hierarchical reformulation of OS - CNN that preserves full receptive - field coverage while eliminating redundancy. This straightforward architectural modification enables the proposed model to leverage Omni-Scale Convolution within a hierarchical framework, progressively reducing the required coverage range at deeper layers. This modification facilitates more comprehensive feature extraction and allows for a deeper and wider design than the OS-CNN. Evaluated on the large - scale CODE - 15% dataset [ 9 ] , EcoScale - Net cuts parameters by 90% and FLOPs by 99% relative to OS - CNN, while boosting macro - averaged F1 - score by 2.4%—the best among all compared models. The contributions of this study are as follows: • We introduce a novel structure ( EcoScale-Net ) aimed at efficiently configuring the kernel size for processing long signals such as ECG. Compared to OS-CNN, our method allows for a deeper and wider design using a hierarchical framework, leading to a reduction in parameters by over tenfold and FLOPs by more than 99%, while achieving a 2.4% performance enhancement. • We demonstrate that applying convolution with a kernel size of 1 before and after each OS stage effectively prevents an explosive increase in channel dimensions while modeling the relationships of each receptive field. • Despite having lower FLOPs and fewer parameters compared to most other models, it demonstrates the best performance in terms of the average F1-score among the various models we compared. 2 Method 2.1 Preliminary: Omni - Scale Convolution Omni - Scale Convolution (OS - Conv) was introduced to avoid manual kernel - size tuning in long signals. Its key idea is to enumerate prime–sized kernels so that every possible receptive - field (RF) length can be obtained by adding three 1 - D convolutions. Kernel set. Let \mathbb{P}^{(j)} be the kernel set at OS stage j ( j=1,2,3 ): \mathbb{P}^{(j)}=\begin{cases}\{\,p\mid p\ \text{is prime},\;1\leq p\leq p_{k}\}&(j=1,2),\\[2.0pt] \{1,2\}&(j=3),\end{cases} (1) where p_{k} is the largest prime allowed at the first two stages. Receptive - field coverage. For p^{(j)}\!\in\!\mathbb{P}^{(j)} , the RF length produced by the three - stage stack is S=p^{(1)}+p^{(2)}+p^{(3)}-2. (2) Because any even number e (\leq 2p_{k}) can be written as the sum of two primes ( p^{(1)}{+}p^{(2)} ) by Goldbach’s conjecture, ( 2 ) yields \mathbb{S}=\bigl\{e,\;e-1\;\big|\;e\in\mathbb{E}\bigr\},\qquad\mathbb{E}=\{2,4,\dots,2p_{k}\}, so all even and adjacent odd RF lengths up to 2p_{k} are covered. Computational bottleneck. Enumerating every prime up to p_{k} incurs O\!\bigl(\sum_{p\leq p_{k}}p\bigr)=O(p_{k}^{2}) parameters and FLOPs per stage; thus complexity grows quadratically with p_{k} and linearly with sequence length. For long ECG sequences ( \sim 4 k samples) this leads to impractical memory and time costs, preventing deeper or wider models. The remainder of this paper presents a hierarchical reformulation that keeps full RF coverage while eliminating this redundancy. Figure 1: (a) Overview of EcoScale - Net. (b) EcoScale Block which is main module for proposed approach. 2.2 Overall Architecture Our main goal is to efficiently configure the kernel size to cover the receptive field for signals with arbitrary lengths, leveraging the key advantage of OS convolution . To achieve main goals and address such impractical issues for long-range 12-ECG, we propose the E fficient C onvolutional O mni- Scale Net work ( EcoScale-Net ) , a lightweight reformulation of OS - CNN for multi - label ECG classification (Fig. 1 ). EcoScale - Net replaces OS - CNN’s exhaustive kernel list with a hierarchical design and dual 1\times 1 bottlenecks, preserving full receptive - field (RF) coverage while cutting parameters and FLOPs. Note that we select same kernel selection strategy with OS-CNN (Algorithm 1 ). In this paper, we used ResNet-34 [ 10 ] as the backbone for the hierarchical architecture. Let \mathbf{X}=\{x_{1},x_{2},\dots,x_{12}\}\in\mathbb{R}^{12\times T} be the input 12-lead ECG signals where T is the length of the ECG signals. Then, we extract the fundamental features z_{0} from the input ECG signal \mathbf{X} as follows: z_{0}=s(\mathbf{X})\in\mathbb{R}^{C_{0}\times L_{0}} (3) where s(\cdot) denotes the stem layer, which consists of a convolution with a kernel size of 7 and a stride of 2, followed by batch normalization, ReLU, and max pooling for downsampling. And, C_{0} and L_{0} are the numbers of channels and time length of the stem feature z_{0} , respectively. Subsequently, the stem feature z_{0} is processed through a feature extractor composed of

Read the complete text