Abstract
Freezing of gait (FOG) is a common symptom in Parkinson's disease (PD) patients that may result in falls and injuries. The accurate detection of FOG events is important for the clinical on-demand cueing and the treatment of patients. To realize FOG detection with high performance, we proposed methods to build FOG detection models using deep convolutional and recurrent networks. The acceleration signals included 237 FOG events captured from 10 PD patients who performed designed walking tasks with an accelerometer placed on their lower back. Raw acceleration patterns and spectrograms were extracted from successive gait cycles and used as input for FOG detection model training. The model performances were compared to evaluate different domain contributions in FOG detection model training. Deep convolutional layers combined with recurrent layers (DeepCNN-LSTM) were offline trained and then employed to detect FOG or non-FOG events. Using the acceleration spectrogram as input, we achieved an average FOG detection accuracy, sensitivity, specificity of 94.3%, 96.1%, and 95.5%, respectively, which was 2.1% higher in accuracy compared to using raw acceleration patterns as input. The proposed methods based on acceleration signals using the DeepCNN-LSTM classifier are able to detect FOG events with high accuracy, which could be transplanted to the wearable device for non-pharmacologic therapy for PD patients.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call