帕金森病是一种常见的神经退行性疾病，基于肢体运动信号、语音信号、深部脑电信号等神经-运动生理信号的帕金森病运动状态评估技术，为自动化、精确化的帕金森病诊疗提供了重要的工具。基于机器学习的疾病表征方法作为其中的关键环节，能够有效地表征相关疾病信息。受限于领域知识和数据规模的不足，现有的疾病表征方法存在表达能力不足、泛化性能有限等问题。本文从帕金森病的神经机制出发，深入开展了生理信号的疾病表征方法研究，从而实现了更加可靠、快速且通用的帕金森病运动状态评估，并将运动状态评估方法成功应用于帕金森病的症状量化、疾病诊断和生理状态评估等问题。本文的主要贡献有： 一、实现了基于肢体运动信号的帕金森病患者下肢运动障碍精确评估。本文从临床评估的先验知识出发，提出了下肢运动疾病表征方法，有效评估了下肢运动状态，并分析了脑深部电刺激对下肢运动的影响。针对临床评分精细度不足的问题，本文进一步提出基于离散步态障碍标签的连续值预测模型。二、实现了在日常场景下基于语音信号的帕金森病诊断评估。语音信号作为一种便于采集的生理信号，有望被用于快速地表征患者的疾病状况。本文从发声机制出发，分析了帕金森病患者与健康人的发声器官运动差异。进一步优化基于深度神经网络的帕金森病表征学习方法，针对深度表征模型在推广到大规模数据集时泛化能力不足的问题，本文首次引入自监督学习方法，通过融合同模态外部数据，实现了对临床数据的高效利用。该方法在疾病表征的表达能力和泛化性能上超过了现有方法，可用于对帕金森病患者的诊断评估。 三、实现了基于深部脑电信号的帕金森病患者睡眠-觉醒状态评估。本文以语音信号与深部脑电信号的相似性为出发点，提出了从语音到深部脑电信号的跨模态迁移学习方法，该方法提高了对帕金森病患者睡眠-觉醒状态的表征能力，实现了精确的睡眠-觉醒状态评估，为脑深部电刺激闭环治疗方案提供了支持。
Parkinson‘s Disease, a commonplace neurodegenerative disorder, utilizes assessment methodologies premised on neuro-motor physiological signals such as limb motion, speech, and deep brain signals, thus providing indispensable instruments for an automated and precise diagnostic and treatment framework. Disease representation methods based on machine learning, being integral components thereof, effectually represent pertinent disease information. Limited by domain knowledge and insufficient data volume, current disease representation methodologies grapple with limited generalization capabilities. This paper, stemming from the neurological mechanisms of Parkinson‘s disease, delves into research on physiological signal disease characterization methods, thereby achieving a more reliable, swift, and universal assessment of Parkinson‘s disease motor status. It has been successfully implemented for symptom quantification, disease diagnosis, and physiological status evaluation of Parkinson‘s Disease. The principal contributions of this paper include:1. Realizing precise evaluation of lower limb motor impairment in patients with Parkinson‘s disease based on limb motion signals. Derived from the prior knowledge of clinical evaluation, this paper proposes a disease representation method for lower limb movement, effectively assessing lower limb motor status and analyzing the impact of deep brain stimulation on lower limb movement. To address the issue of insufficient granularity in clinical scoring, a continuous prediction model based on discrete gait disorder labels has been further proposed. 2. Implementing Parkinson‘s disease diagnosis evaluation based on speech signals in daily scenarios. As an easily collectible physiological signal, speech signals hold the potential for swiftly characterizing patients‘ disease conditions. Premised on the vocalization mechanism, this paper analyzes the difference in vocal organ movement between Parkinson‘s disease patients and healthy individuals. Further optimizing the Parkinson‘s disease representation learning method based on deep neural networks, the paper introduces a self-supervised learning method. By integrating same-modality external data, it realizes efficient utilization of clinical data. This method surpasses existing methods in expressibility and generalization capabilities of disease representation, thus applicable for the diagnostic evaluation of Parkinson‘s disease patients. 3. Achieving sleep-awake status evaluation of Parkinson‘s disease patients based on deep brain signals. Starting from the similarity between speech signals and deep brain signals, this paper proposes a cross-modality transfer learning method from speech to deep brain signals. This method enhances the representation ability of the sleep-awake status of Parkinson‘s disease patients, achieving precise sleep-awake status evaluation, thereby providing support for the closed-loop treatment plan of deep brain stimulation.
