[home] [Personal Program] [Help]

---

Session: Motion
Session starts: Thursday 26 January, 14:30
Presentation starts: 14:30
Room: Room 531

---

Irene Heijink (Biomedical Signals and Systems, University of Twente; Department of Neurology, Medisch Spectrum Twente.)
Emilie Klaver (Donders Institute; Department of Neurology, Medisch Spectrum Twente.)
Jeroen van Vugt (Department of Neurology, Medisch Spectrum Twente.)
Richard van Wezel (Donders Institute; Biomedical Signals and Systems, University of Twente)
Marleen Tjepkema-Cloostermans (Department of Neurology, Medisch Spectrum Twente; Clinical Neurophysiology Group, University of Twente.)

Abstract:
Introduction: Freezing of gait (FOG) is one of the debilitating symptoms experienced by patients with Parkinson’s Disease, and the most common cause of falls in these patients. FOG is described as feet being glued to the floor. Cueing, external stimuli like bars on the floor, can help to overcome FOG. In order to enhance the user experience of cueing devices and to diminish intrusiveness and habituation to cues, on-demand cueing is desired. In addition, objective FOG detection enables monitoring and objective assessment of therapy. In this research, the performance on the detection of FOG of three deep learning classifiers based on acceleration data is studied. Methods: Acceleration data from four studies with walking tasks ranging from straight walking, turning and a narrow pathway was combined. All experiments were recorded for video annotation of FOG, the gold standard. We evaluated the performance of three classification models for the detection of FOG: a Convolution Neural Network (CNN), MiniRocket and InceptionTime. Five fold cross-validation was applied to estimate an unbiased model performance. The models were evaluated using a receiver operating characteristic (ROC) curve. For the best model we made a comparison between different sensor combinations of acceleration data from the upper legs, lower legs and feet. The best model in combination with the best sensor selection was evaluated on the hold out set. Results: Seventy-one participants were included in this study, 8% of the data was labelled as FOG. The highest AUC-ROC was reached for the CNN trained on the acceleration data of the lower legs and feet with an AUC-ROC of 0.72, sensitivity of 73.7% (72.5 - 75.0%) and specificity of 60.8% (60.3 - 61.3%) on the test set. The mean AUC-ROC of MiniRocket was 0.10 smaller than the AUC-ROC of the CNN, and the mean AUC-ROC of InceptionTime was 0.04 smaller than the CNN. The difference in mean AUC-ROC for the sensor combinations was 0.01. Conclusion: The classification algorithm has potential to be implemented in on-demand cueing devices and home monitoring applications for objective FOG detection. Further research is needed to optimize the model and improve the performance.