Automatic Detection and Assessment of Freezing of Gait Manifestations

Abstract

Freezing of gait (FOG) is an episodic and highly disabling symptom of Parkinson's disease (PD). Although described as a single phenomenon, FOG is not univocal and can express as different manifestations, such as trembling in place or complete akinesia. We aimed to analyze the utility of deep learning trained on inertial measurement unit data to classify FOG into both manifestations. We developed a temporal convolutional neural network, which we compared to three state-of-the-art FOG detection algorithms that were adapted to the FOG manifestation detection task. Next, we investigated its performance in distinguishing between the two manifestations and other forms of movement cessation (e.g., volitional stopping and sitting) based on gold-standard video annotations. Experiments were conducted on a dataset of twelve PD patients with FOG that completed a FOG-provoking protocol, including the timed-up-and-go and 360-degree turning-in-place tasks during ON and OFF anti-Parkinsonian medication. The results showed that our model enables accurate detection of FOG manifestations with an 11.43% higher F1 score than the second-best model. Assessment of FOG manifestation severity was moderately strong for trembling in place (Intra-class Correlation Coefficient (ICC)=0.64, [0.16,0.88]) and strong for complete akinesia (ICC=0.87, [0.63,0.96]). Remarkably, our results show that complete akinesia can be distinguished from volitional stopping. In conclusion, we established that FOG manifestations could be accurately detected and assessed with deep learning. Future work should establish whether these results hold firm for a more extensive and varied verification cohort.