Основні поняття
The proposed FOG-Transformer deep neural network architecture, which combines convolutional and transformer blocks, can significantly improve the detection of freezing of gait episodes in Parkinson's disease patients using a single waist-worn triaxial accelerometer.
Анотація
The study aimed to improve the performance of automatic freezing of gait (FOG) detection in Parkinson's disease (PD) patients using a single waist-worn triaxial accelerometer and a novel deep learning approach based on transformer networks.
Key highlights:
The study used the REMPARK-FOG dataset, which contains recordings from 21 PD patients performing activities of daily living in home settings, with over 1,000 FOG episodes recorded.
Several feature extraction methods were evaluated, including Mazilu features, raw signals, and spectral representations using the Fast Fourier Transform (FFT).
Deep learning approaches were assessed, including convolutional neural networks (CNNs), CNN-LSTM, and the proposed FOG-Transformer architecture, which combines CNN and transformer blocks.
The FOG-Transformer achieved the best performance, with a sensitivity and specificity of 0.891, an AUC of 0.957, and a 10.9% equal error rate, outperforming the baseline and other deep learning methods.
The use of transformer blocks in the FOG-Transformer allowed improved modeling of the temporal information from previous spectral windows, leading to enhanced FOG detection compared to recurrent neural network-based approaches.
The study also proposed a post-processing methodology to analyze the detection of FOG episodes and clusters, providing insights into the practical implementation of the system for long-term monitoring in home settings.
Статистика
The dataset contains over 1,000 FOG episodes recorded from 21 Parkinson's disease patients during activities of daily living in home settings.
The total signal duration corresponds to 18 hours, with 9.1 hours in the OFF state and 8.9 hours in the ON state.
10.5% of the accelerometer signals correspond to FOG episodes, while 89.5% correspond to non-FOG activities.
Цитати
"Freezing of gait (FOG) is one of the most disabling symptoms in Parkinson's disease (PD), characterized by brief episodes of inability to step or the presence of short steps when initiating gait, on turning while walking or when experiencing stressful situations."
"Wearable devices have shown high potential for the development of systems for FOG detection in both laboratory and home settings."
"Transformer networks (Vaswani et al., 2017) were introduced for sequence to-sequence learning and offer improved performance at long-range sequential modelling over RNNs."