基于深度卷积神经网络的帕金森步态识别

发布时间：2018-02-26 09:18

本文关键词： 帕金森疾病步态冻结识别深度学习卷积神经网络　出处：《中国生物医学工程学报》2017年04期 　论文类型：期刊论文

【摘要】：步态冻结(FOG)是晚期帕金森病患者最常见的症状,FOG的突然发作会造成患者的行走障碍。为患者佩戴FOG检测可穿戴设备助手是一种有效可行的治疗途径,当检测到FOG发作时,可穿戴设备助手为患者提供一段有节奏的听觉信号刺激患者恢复行走。针对FOG检测,提出一种系统式的特征学习方法。该方法采用一个基于深度学习的卷积神经网络,对原始输入信号自动地进行特征学习。采用监督式学习方法利用标签信息,使学习到的特征更具识别能力。在整个网络模型中,特征学习和分类互相加强使整个网络更加稳定,更具智能化。通过DAPHNet数据集进行验证,结果表明,该方法可以自动地进行特征学习并识别出步态冻结。与以往的阈值法实验结果相比,平均正确率提高到91.43%,灵敏性提高到85.58%,特异性提高到了93.63%。该方法可以在一定程度上代替人工干预,在处理频繁出现FOG症状的帕金森患者的治疗中具有重要意义。
[Abstract]:Gait freezing is the most common symptom in patients with advanced Parkinson's disease. The sudden onset of fog can cause walking disorders. Wearing a wearable device assistant for patients with FOG is an effective therapeutic approach when the onset of FOG is detected. Wearable device assistants provide a rhythmic auditory signal to stimulate patients to resume walking. For FOG detection, a systematic feature learning method is proposed, which uses a convolution neural network based on deep learning. The feature learning of the original input signal is carried out automatically. The supervised learning method is used to utilize the label information to make the features more recognizable. In the whole network model, the feature learning and classification reinforce each other to make the whole network more stable. The results show that the method can automatically learn features and recognize gait freezing. Compared with the experimental results of the previous threshold method, the proposed method is more intelligent. The average accuracy rate was increased to 91.43 and the sensitivity to 85.58 and the specificity to 93.63.The method can replace artificial intervention to some extent and is of great significance in the treatment of Parkinson's patients with frequent symptoms of FOG.
【作者单位】：燕山大学信息科学与工程学院;
【基金】：国家自然科学基金(61473339) 河北省青年拔尖人才支持计划资助项目([2013]17) 中国博士后科学基金(2014M561202)
【分类号】：R742.5;TP183

【相似文献】