基于长短时记忆网络(LSTM)的帕金森步态识别

Detection of Freezing of Gait for Parkinson’s Disease Patients Based on LSTM Network

长短期记忆网络是一种时间递归神经网络。与循环神经网络相比,其遗忘门、输入门和输出门的结构在预测时间序列上解决了时间序列中间间隔较长的重要时间被遗忘的问题。针对检测步态冻结这种具有时间序列特性的任务,提出一种基于LSTM的特征学习方法。通过DAPHNet数据集对该方法进行测试,结果表明,该方法可有效检测步态冻结的发生,在治疗经常出现步态冻结的帕金森症患者时有一定适用价值。

The Long Short-Term Memory is a time recurrent neural network. Compared with the cyclic neural network, the structure of the forgetting gate, the input gate and the output gate solves the problem that the important time interval between the time series is forgotten in the predicted time series. Aiming at detecting the task of gait freezing with time series characteristics, a feature learning method based on LSTM is proposed. The method was tested by DAPHNet dataset. The results showed that the method can effectively detect the occurrence of gait freezing and has certain applicability in the treatment of Parkinson’s disease patients with frequent gait freezing.