IT Parkinson’s Disease Diagnostics Based on the Freezing of Gait Analysis Using Long Short Term Memory Neural Network

Abstract

An analysis of methods for processing data from gait deceleration sensors for detecting Parkinson’s disease and a description of the development of a Parkinson’s recognition system based on neural networks with long short term memory (LSTM) are performed. The data used was a publicly available dataset of gait deceleration scores of patients with Parkinson’s disease, obtained using three wearable sensors to collect data from different parts of the body. The research was carried out using machine learning using an LSTM neural network. First, the DAPHNet datasets were segmented using a fixed sliding window algorithm. The wavelet algorithm was then used to extract features from the data set: wavelet entropy and energy, wavelet waveform length, variance and standard deviation of wavelet coefficient. Next, a data enhancement algorithm was used to balance the number of samples in the data sets. To train the model, an LSTM neural network was built with a six-layer network structure: input layer, LSTM layer, reLU layer, fully connected layer, Softmax layer and output layer. After training the model for 1000 iterations, the LSTM neural network algorithm achieved 96.3 % accuracy, 96.05 % precision, 96.5 % sensitivity, and 96.24 % average F1 score for recognizing Parkinson’s disease based on test datasets. Similar studies conducted by other scientific organizations achieved a maximum accuracy of 91.9 % for the same data sets.