Deep Learning-Based Prediction of Freezing of Gait in Parkinson's Disease With the Ensemble Channel Selection Approach

Abstract

Purpose: A debilitating and poorly understood symptom of Parkinson's disease (PD) is freezing of gait (FoG), which increases the risk of falling. Clinical evaluations of FoG, relying on patients' subjective reports and manual examinations by specialists, are unreliable, and most detection methods are influenced by subject-specific factors.

Method: To address this, we developed a novel algorithm for detecting FoG events based on movement signals. To enhance efficiency, we propose a novel architecture integrating a bottleneck attention module into a standard bidirectional long short-term memory network (BiLSTM). This architecture, adaptable to a convolution bottleneck attention-BiLSTM (CBA-BiLSTM), classifies signals using data from ankle, leg, and trunk sensors.

Finding: Given three movement directions from three locations, we reduce computational complexity in two phases: selecting optimal channels through ensemble learning followed by feature reduction using attention mapping. In FoG event detection tests, performance improved significantly compared to control groups and existing methods, achieving 99.88% accuracy with only two channels.

Conclusion: The reduced computational complexity enables real-time monitoring. Our approach demonstrates substantial improvements in classification results compared to traditional deep learning methods.

Keywords: Parkinson's disease; bidirectional long short‐term memory; bottleneck attention module; channel selection; ensembling; freezing of gait.

FIGURE 1

The diagram shows how FoG events are detected and then used to classify signals as PD. biLSTM, bidirectional long short‐term memory network; PD, Parkinson's disease.

FIGURE 2

In our method of ensembling, three metrics are used to find the most advantageous channels, with the channel weight increasing the classification accuracy of the training data. CV, cross‐validation; ROC, receiver operating characteristic; SNR, signal‐to‐noise ratio.

FIGURE 3

This figure outlines the concept of channel attention. The aforementioned structure encompasses three distinct processes, namely, global average pooling (F_gap ), fully connected (F_f ), and batch normalization.

FIGURE 4

An illustration of spatial attention. It is mostly composed of three operations: batch normalization, 3 × 3 dilated convolution F_dc , and 1 × 1 convolution ${\mathrm{F}}_{{\mathrm{f}}^{1\times 1}}$.

FIGURE 5

This figure depicts the bottleneck‐LSTM framework. The aforementioned architecture, constructed upon a conventional BiLSTM model, possesses the capability to deduce channel attention (M_c ) and spatial attention (M_s ) maps.

FIGURE 6

The provided structure offers a comprehensive outline of CBA‐BiLSTM, a model that integrates convolutional, max‐pooling, and softmax layers. CBA‐biLSTM, convolution bottleneck attention–bidirectional long short‐term memory network; FoG, freezing of gait.

FIGURE 7

A detection of PD was performed by observing the accelerations of ankle, thigh, and trunk movements during gait movements.

FIGURE 8

A detection of PD was performed by observing the accelerations of ankle, thigh, and trunk movements during gait movements. FoG, freezing of gait.

FIGURE 9

This figure compares the performance of two strong approaches with the proposed method based on ROC curves. AUC, area under the curve; biLSTM, bidirectional long short‐term memory network; CBA, convolution bottleneck attention; CNN, convolutional neural network; ROC, receiver operating characteristic.

FIGURE 10

On the basis of RMSE and LOSS, this figure compares the almost real‐time error of monitoring FoG for three effective channels with two DL structures. biLSTM, bidirectional long short‐term memory network; CBA, convolution bottleneck attention; CNN, convolutional neural network; RMSE, root mean square error.

FIGURE 11

When using three unseen signals in channel selection, the ensemble (Ensm) model performs better than other similar approaches. ROC, receiver operating characteristic; SNR, signal‐to‐noise ratio.