MULTI-HEAD CROSS-ATTENTIONAL PPG AND MOTION SIGNAL FUSION FOR HEART RATE ESTIMATION Panagiotis KasnesisLazaros ToumanidisAlessio Burrello

2025-05-02 0 0 1.84MB 5 页 10玖币

侵权投诉

MULTI-HEAD CROSS-ATTENTIONAL PPG AND MOTION SIGNAL FUSION FOR HEART

RATE ESTIMATION

Panagiotis Kasnesis§†,Lazaros Toumanidis§†,Alessio Burrello*,

Christos Chatzigeorgiou§†,Charalampos Z. Patrikakis§†,

§ThinGenious PC, Maroussi, Greece

†Department of Electrical and Electronics Engineering, University of West Attica, Greece

*Department of Electrical, Electronic and Information Engineering, University of Bologna, Italy

ABSTRACT

Nowadays, Hearth Rate (HR) monitoring is a key feature of

almost all wrist-worn devices exploiting photoplethysmogra-

phy (PPG) sensors. However, arm movements affect the per-

formance of PPG-based HR tracking. This issue is usually ad-

dressed by fusing the PPG signal with data produced by iner-

tial measurement units. Thus, deep learning algorithms have

been proposed, but they are considered too complex to deploy

on wearable devices and lack the explainability of results. In

this work, we present a new deep learning model, PULSE,

which exploits temporal convolutions and multi-head cross-

attention to improve sensor fusion’s effectiveness and achieve

a step towards explainability. We evaluate the performance

of PULSE on three publicly available datasets, reducing the

mean absolute error by 7.56% on the most extensive available

dataset, PPG-DaLiA. Finally, we demonstrate the explainabil-

ity of PULSE and the beneﬁts of applying attention modules

to PPG and motion data.

Index Terms—Deep Learning, Sensor Fusion, Heart

Rate Monitoring, Attention, Photoplethysmography

1. INTRODUCTION

In recent years, wrist-worn devices (i.e., smartwatches) en-

able a 24h-monitoring of the subject’s vital conditions thanks

to miniaturized sensors, becoming increasingly popular in

personalized health care and medical IoT applications [1].

One of the most important indices to monitor is Heart Rate

(HR). Compared to ﬁrst-generation monitoring devices,

which exploit a simple 1-3 leads ECG connected through

a chest strip, modern ones use photoplethysmographic (PPG)

sensors, allowing HR monitoring to be integrated into the

smartwatches [2]. However, a limitation of PPG-based HR

monitoring is given by the presence of motion artifacts (MA).

These are caused by variations of sensor position on the wrist

or ambient light leaking in the gap between the sensor and

the wrist. In literature, this problem has been ﬁrst tackled uti-

lizing ﬁltering approaches. They use the correlation between

acceleration data and the PPG signal to cancel the noise and

remove the MAs. Then, the HR is extrapolated from the

cleaned signal [3, 4]. The critical limitation of these ap-

proaches is the high number of free hyper-parameters, which

often limits their generalization.

Deep learning approaches have been proposed to improve

generalization, bringing promising results on different public

datasets [5, 6, 7, 8]. On the other hand, these models lack

explainability, since acceleration and PPG data are fused in a

black box. Until now, little attention has been posed to recent

Transformers, given the usual high number of parameters re-

quired to train them. These models are based on the so-called

Attention Modules, which correlate different tensors.

In this paper, we demonstrate that combining feature maps

of convolutions with attention modules leads to improved ac-

curacy in the PPG-based HR monitoring and allows to inter-

pret the connection between acceleration and PPG data. The

main contributions of this work are summarized as follows:

• We introduce a new state-of-the-art, yet lightweight (around

130M parameters), deep neural network to fuse PPG and

motion signals for precise heart rate estimation. The model

includes both temporal convolutional and multi-head cross-

attention modules.

• We evaluate the effectiveness of the produced model on

three publicly available datasets. On the largest one, PPG-

DaLiA [5], we improve the mean absolute error (MAE) to

4.03 beats per minute (BPM), outperforming the best state-

of-the-art model (a pure CNN) by 0.33 BPM.

• We demonstrate the explainability of the developed model

and the beneﬁts of applying attention modules to PPG and

motion data by showcasing examples of attentional maps.

2. BACKGROUND

Temporal Convolutional Networks: TCNs are 1D-Convolutional

Neural Networks (CNNs), with the insertion of dilation in

convolutional layers [9, 10]. The dilation is a ﬁxed gap d

inserted between input samples before being convolved with

arXiv:2210.11415v1 [eess.SP] 14 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MULTI-HEADCROSS-ATTENTIONALPPGANDMOTIONSIGNALFUSIONFORHEARTRATEESTIMATIONPanagiotisKasnesis§,LazarosToumanidis§,AlessioBurrello*,ChristosChatzigeorgiou§,CharalamposZ.Patrikakis§,§ThinGeniousPC,Maroussi,GreeceDepartmentofElectricalandElectronicsEngineering,UniversityofWestAttica,Greece*Departmen...

展开>> 收起<<

MULTI-HEAD CROSS-ATTENTIONAL PPG AND MOTION SIGNAL FUSION FOR HEART RATE ESTIMATION Panagiotis KasnesisLazaros ToumanidisAlessio Burrello.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

MULTI-HEAD CROSS-ATTENTIONAL PPG AND MOTION SIGNAL FUSION FOR HEART RATE ESTIMATION Panagiotis KasnesisLazaros ToumanidisAlessio Burrello

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: