Objective: Validate a novel software platform to measure freezing of gait using a modular build-up from 1) basic algorithm, to 2) integration of wearable sensor signals, to 3) personalized freezing detection algorithm.
Background: Freezing of Gait (FOG) is a devastating and heterogeneous symptom in Parkinson’s disease (PD). The most important bottleneck to advance research is the difficulty to measure FOG, especially in the home setting. Clinical expert-rated videos of FOG are currently the gold standard outcome, yet labor intensive. Until now, sensor-based algorithms were validated in-lab and lack personalization. Here, we present the first results of the FOG-IT project intended to address these drawbacks.
Method: Based on a multi-stage graph convolution network method, a deep learning FOG-IT model was computed on existing 3D marker-based motion capture (MoCap) data from 14 PD patients with FOG (age: 68.6±7.4 years; disease duration: 9.0±4.8years), 14 without FOG (age: 66.7±7.4 years; disease duration: 7.8±4.8 years) and 14 controls (age: 65.2±6.8 years). MoCap data were collected OFF medication. The protocol consisted of straight walking and making left and right 360° and 180° turns (Figure 1). Only 7 PD patients experienced FOG (56 events). A leave-one-subject-out validation process was used to test the model’s ability to detect FOG and percentage time frozen (%TF) during the protocol.
Results: Compared with expert raters, the FOG-IT model had a moderate correlation for FOG events (r=0.74; 95%CI[0.53-0.86]) and a very strong correlation for %TF (r=0.95; 95%CI[0.91-0.98]). The model had an overall error of 9.54±10.1% to determine %TF. However, on a case-by-case basis, results were more variable with error rates ranging from 0.2%-25.6%, highlighting the variability of FOG. Figure 2 shows 5/35 examples of expert versus model FOG detection illustrating this variable pattern.
Conclusion: We demonstrated a high accuracy of the initial model to detect FOG on a small dataset forming the foundation of the FOG-IT project. Next steps will include collecting a prospective dataset using a freezing-provoking protocol in the lab and the home with a wearable sensor-based system. For future enhancement of the episode detection, expert rater input will be used to complement deep learning iterations to obtain patient-specific algorithms, constituting the final goal of the FOG-IT project.
To cite this abstract in AMA style:B. Filtjens, N. Ghosh, P. Slaets, B. Vanrumste, A. Nieuwboer, P. Ginis. FOG-IT – Towards Personalized Freezing of Gait detection using Artificial Intelligence [abstract]. Mov Disord. 2021; 36 (suppl 1). https://www.mdsabstracts.org/abstract/fog-it-towards-personalized-freezing-of-gait-detection-using-artificial-intelligence/. Accessed December 3, 2023.
« Back to MDS Virtual Congress 2021
MDS Abstracts - https://www.mdsabstracts.org/abstract/fog-it-towards-personalized-freezing-of-gait-detection-using-artificial-intelligence/