Enhancing tremor classification: Transformer-based analysis of biomechanics patterns for Parkinson's and essential tremor

IF 1.4 3区医学 Q4 ENGINEERING, BIOMEDICAL

Clinical Biomechanics Pub Date : 2025-06-20 DOI:10.1016/j.clinbiomech.2025.106599

Muhammad Izzuddin Mahali , Jenq-Shiou Leu , Cries Avian , Jeremie Theddy Darmawan , Muhamad Faisal , Nur Achmad Sulistyo Putro , Setya Widyawan Prakosa

{"title":"Enhancing tremor classification: Transformer-based analysis of biomechanics patterns for Parkinson's and essential tremor","authors":"Muhammad Izzuddin Mahali , Jenq-Shiou Leu , Cries Avian , Jeremie Theddy Darmawan , Muhamad Faisal , Nur Achmad Sulistyo Putro , Setya Widyawan Prakosa","doi":"10.1016/j.clinbiomech.2025.106599","DOIUrl":null,"url":null,"abstract":"<div><h3>Background</h3><div>Differentiating Essential Tremor and Parkinson's Disease is challenging due to overlapping tremor characteristics, including similar frequency ranges (4–8 Hz) and kinetic manifestations that defy conventional clinical differentiation. This study aimed to develop a multiclass differentiation system for Essential Tremor, Parkinson's Disease, and Healthy Controls by employing deep learning to decode distinct biomechanical patterns from multi-sensor movement data during dynamic motor tasks.</div></div><div><h3>Methods</h3><div>Tremor severity was assessed using accelerometers positioned on the thumb, index finger, metacarpal, and wrist during four protocols: two static (rest, postural) and two dynamic (free motion, motion with object). We employed a Transformer-based model with multi-head attention to capture spatiotemporal movement patterns. Two analytical approaches were compared: (1) feature extraction followed by Transformer processing, and (2) direct Transformer processing of raw signals.</div></div><div><h3>Findings</h3><div>The feature-based approach achieved perfect classification accuracy (100 %) for postural holding (utilizing integrated absolute value and other derived features) and free motion (employing mean power and additional features). The raw signal approach similarly attained 100 % accuracy in classifying free motion (200-sample window) and motion with object (200- and 300-sample windows). Integration of multi-protocol dynamic tasks (free motion and motion with object) yielded 99.32 % overall accuracy. Crucially, dynamic protocols demonstrated consistent superiority over static protocols in diagnostic performance.</div></div><div><h3>Interpretation</h3><div>The Transformer model with multi-head attention effectively identified disease-specific biomechanical patterns. Its high accuracy in distinguishing Essential Tremor, Parkinson's Disease, and Healthy Control participants, particularly during dynamic tasks, positions it as a promising tool for enhancing clinical decision-making and artificial intelligence-assisted monitoring of neurodegenerative disorders.</div></div>","PeriodicalId":50992,"journal":{"name":"Clinical Biomechanics","volume":"127 ","pages":"Article 106599"},"PeriodicalIF":1.4000,"publicationDate":"2025-06-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Clinical Biomechanics","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S026800332500172X","RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"ENGINEERING, BIOMEDICAL","Score":null,"Total":0}

引用次数: 0

Abstract

Background

Differentiating Essential Tremor and Parkinson's Disease is challenging due to overlapping tremor characteristics, including similar frequency ranges (4–8 Hz) and kinetic manifestations that defy conventional clinical differentiation. This study aimed to develop a multiclass differentiation system for Essential Tremor, Parkinson's Disease, and Healthy Controls by employing deep learning to decode distinct biomechanical patterns from multi-sensor movement data during dynamic motor tasks.

Methods

Tremor severity was assessed using accelerometers positioned on the thumb, index finger, metacarpal, and wrist during four protocols: two static (rest, postural) and two dynamic (free motion, motion with object). We employed a Transformer-based model with multi-head attention to capture spatiotemporal movement patterns. Two analytical approaches were compared: (1) feature extraction followed by Transformer processing, and (2) direct Transformer processing of raw signals.

Findings

The feature-based approach achieved perfect classification accuracy (100 %) for postural holding (utilizing integrated absolute value and other derived features) and free motion (employing mean power and additional features). The raw signal approach similarly attained 100 % accuracy in classifying free motion (200-sample window) and motion with object (200- and 300-sample windows). Integration of multi-protocol dynamic tasks (free motion and motion with object) yielded 99.32 % overall accuracy. Crucially, dynamic protocols demonstrated consistent superiority over static protocols in diagnostic performance.

Interpretation

The Transformer model with multi-head attention effectively identified disease-specific biomechanical patterns. Its high accuracy in distinguishing Essential Tremor, Parkinson's Disease, and Healthy Control participants, particularly during dynamic tasks, positions it as a promising tool for enhancing clinical decision-making and artificial intelligence-assisted monitoring of neurodegenerative disorders.

查看原文本刊更多论文

增强震颤分类：基于变压器的帕金森病和特发性震颤生物力学模式分析

由于震颤特征重叠，包括相似的频率范围（4 - 8hz）和动力学表现，难以传统的临床鉴别，原发性震颤和帕金森病的鉴别具有挑战性。本研究旨在通过深度学习从动态运动任务中的多传感器运动数据中解码不同的生物力学模式，开发特发性震颤、帕金森病和健康对照的多类别分化系统。方法采用放置在拇指、食指、掌骨和手腕上的加速度计，在静态（休息、体位）和动态（自由运动、与物体运动）四种方案中评估震颤严重程度。我们采用了一个基于变形金刚的多头部注意模型来捕捉时空运动模式。比较了两种分析方法：(1)特征提取后进行Transformer处理，(2)原始信号直接进行Transformer处理。基于特征的方法在姿势保持（利用综合绝对值和其他衍生特征）和自由运动（利用平均力量和其他特征）方面实现了完美的分类精度（100%）。原始信号方法在分类自由运动（200个样本窗口）和带对象运动（200和300个样本窗口）方面同样达到了100%的准确率。多协议动态任务（自由运动和有目标运动）的集成总体准确率为99.32%。至关重要的是，动态协议在诊断性能上始终优于静态协议。具有多头关注的Transformer模型有效地识别了疾病特异性的生物力学模式。它在区分特发性震颤、帕金森氏病和健康控制参与者方面的高精度，特别是在动态任务中，使其成为加强临床决策和人工智能辅助监测神经退行性疾病的有前途的工具。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Clinical Biomechanics 医学-工程：生物医学

CiteScore

3.30

自引率

5.60%

发文量

189

审稿时长

12.3 weeks

期刊介绍： Clinical Biomechanics is an international multidisciplinary journal of biomechanics with a focus on medical and clinical applications of new knowledge in the field. The science of biomechanics helps explain the causes of cell, tissue, organ and body system disorders, and supports clinicians in the diagnosis, prognosis and evaluation of treatment methods and technologies. Clinical Biomechanics aims to strengthen the links between laboratory and clinic by publishing cutting-edge biomechanics research which helps to explain the causes of injury and disease, and which provides evidence contributing to improved clinical management. A rigorous peer review system is employed and every attempt is made to process and publish top-quality papers promptly. Clinical Biomechanics explores all facets of body system, organ, tissue and cell biomechanics, with an emphasis on medical and clinical applications of the basic science aspects. The role of basic science is therefore recognized in a medical or clinical context. The readership of the journal closely reflects its multi-disciplinary contents, being a balance of scientists, engineers and clinicians. The contents are in the form of research papers, brief reports, review papers and correspondence, whilst special interest issues and supplements are published from time to time. Disciplines covered include biomechanics and mechanobiology at all scales, bioengineering and use of tissue engineering and biomaterials for clinical applications, biophysics, as well as biomechanical aspects of medical robotics, ergonomics, physical and occupational therapeutics and rehabilitation.