基于深度LSTM残差网络的帕金森症诊断方法

Diagnosis of Parkinson's disease based on deep LSTM residual network

语音特征分类下的帕金森症诊断方法具有无创、高效、准确、远程与成本低等特点,本研究提出一种基于深度长短期记忆网络(LSTM)残差网络的帕金森症诊断方法。分析帕金森症语音信号特点和LSTM残差模型,基于深度LSTM残差网络的帕金森症诊断模型分成3个部分:语音信号预处理网络、深度LSTM残差语音诊断网络和GAP-ELM帕金森症分类网络。该模型能实现帕金森语音信号的深层特征提取,通过LSTM结构的遗忘门和记忆门得到帕金森语音信号随时间变化的状态,最后通过帕金森症元音集完成帕金森症诊断测试。结果表明本文方法在各类信噪比环境中的帕金森症识别准确度均较高,并可在较少的轮次中完成训练,达到较优的稳定性和较小的损失值。

Parkinson's disease(PD)diagnosis method based on speech feature classification has the characteristics of noninvasive,efficient,accurate,remote and low-cost.A PD diagnosis method based on deep long short-term memory(LSTM)residual network is proposed.The characteristics of speech signal in PD patients and deep LSTM residual model are analyzed,and the PD diagnosis model based on deep LSTM residual network is divided into 3 parts,namely speech signal pre-processing network,deep LSTM residual network for speech diagnosis and GAP-ELM network for PD classification.The proposed model can realize the extraction of deep features of PD speech signal,and obtain the time-changing state of PD speech signal through the forgetting gate and memory gate of LSTM structure.PD diagnostic test is completed using PD vowel set.The experimental results demonstrate the proposed method has a higher recognition accuracy for PD in various signal-to-noise ratio environments,and it can complete the training in fewer epochs,and achieve better stability and smaller loss value.