Impact of Data Preparation in Freezing of Gait Detection Using Feature-Less Recurrent Neural Network

Abstract

Many studies showed the feasibility of detecting Freezing of Gait (FOG) of Parkinson patients by using several numbers of inertial sensors worn at the body and back-end computing power. This work uses machine learning approaches analyzing the data of one single body-worn inertial sensor system to classify and detect FOG. Long-Short-Term-Memory (LSTM) is employed as the FOG detection algorithm and the Daphnet (FOG and normal gait) dataset provides the data for model training and testing in this paper. The model considers raw data from three channels of the acceleration sensor mounted on the patient's shank and ignores all other data from other sensors. The model is patient dependent and sensitivity and specificity metrics are used to evaluate the model's performance. In this paper, we propose a novel padding method that is applied to the windows of FOG and non-FOG with zero overlaps on the training set and adapts the padding to the individual regions. This method produces windows containing only one type of data and label. The proposed padding method reduces the padding amount by two orders of magnitude compared to bigger batch-sizes in the sequence splitting method offered by MATLAB 2019a. The padding amount is independent of the batch-size. Raw data is fed to the model in the testing mode without any pre-processing or data transformation. The standard rolling window generates fixed-size windows for the test set without overlap and the higher amount of FOG or Normal walking data defines the label of the individual window. The model for one-second long windows applied in this work outperformed the literature results with a sensitivity of 92.57% and a specificity of 95.62% compared to 82% and 94% reported by Masiala et al.

Recent trends in wearable device used to detect freezing of gait and falls in people with Parkinson’s disease: A systematic review

“…For FOG detection, 72 papers investigated the usage of wearable devices to access FOG in PD ( Table 2 ; Mazilu et al, 2015 , 2016 ; Zach et al, 2015 ; Capecci et al, 2016 ; Ahn et al, 2017 ; Kita et al, 2017 ; Saad et al, 2017 ; Camps et al, 2018 ; Samà et al, 2018 ; Borzì et al, 2019 ; Chomiak et al, 2019 ; Pierleoni et al, 2019 ; San-Segundo et al, 2019 ; Ayena and Otis, 2020 ; Kleanthous et al, 2020 ; Li et al, 2020 ; Tang et al, 2020 ; Dvorani et al, 2021 ; El-Attar et al, 2021 ; Esfahani et al, 2021 ; Ghosh and Banerjee, 2021 ; Halder et al, 2021 ; Prado et al, 2021 ; Shalin et al, 2021 ; Naghavi and Wade, 2022 ). The number of subjects used to test the validity of the FOG detection system varied significantly between studies, from 1 ( O’day et al, 2020 ) to 131 ( Borzì et al, 2019 ) (MED = 12).…”

“…To improve the accuracy of FOG and fall detection, machine learning algorithms, including SVM ( Tzallas et al, 2014 ; Ahlrichs et al, 2016 ; Iakovakis et al, 2016 ; Rodríguez-Martín et al, 2017 ; Aich et al, 2018 ; Arami et al, 2019 ; Borzì et al, 2019 , 2021 ; Kleanthous et al, 2020 ; Reches et al, 2020 ; Dvorani et al, 2021 ; El-Attar et al, 2021 ; Ghosh and Banerjee, 2021 ; Mesin et al, 2022 ), k-NN ( Aich et al, 2018 ; Borzì et al, 2019 ; Demrozi et al, 2020 ; Halder et al, 2021 ; Mesin et al, 2022 ), decision trees ( Aich et al, 2018 ; Borzì et al, 2019 ; Pardoel et al, 2021 , 2022 ), hidden Markov model ( Tzallas et al, 2014 ; San-Segundo et al, 2019 ), neural network ( Cole et al, 2011 ; Iakovakis et al, 2016 ; Ly et al, 2017 ; Saad et al, 2017 ; Kim et al, 2018 ; Arami et al, 2019 ; Borzì et al, 2019 ; Mikos et al, 2019 ; Kleanthous et al, 2020 ; O’day et al, 2020 , 2022 ; Shi et al, 2020 , 2022 ; Sigcha et al, 2020 ; Ashfaque Mostafa et al, 2021 ; El-Attar et al, 2021 ; Prado et al, 2021 ; Naghavi and Wade, 2022 ), random forest ( San-Segundo et al, 2019 ; Kleanthous et al, 2020 ; Ghosh and Banerjee, 2021 ) and LSTM ( Li et al, 2020 ; Ashfaque Mostafa et al, 2021 ; Esfahani et al, 2021 ; Shalin et al, 2021 ; Guo et al, 2022 ), were used extensively in recent studies. Data were collected from sensors, and a training period is necessary for machine learning.…”

Insights into Parkinson’s Disease-Related Freezing of Gait Detection and Prediction Approaches: A Meta Analysis

“…It is often cited by researchers, making it appropriate for machine learning model evaluation and training. The extensive use of the Daphnet dataset in research enables insightful comparisons and promotes improvements in FOG prediction techniques [ 40 , 60 , 80 , 82 , 94 , 106 , 123 ]. Even though the Daphnet dataset is widely used, it is important to recognize its limitations.…”

“… The utilization of the datasets in the literature. (Ferster et al, 2015): [ 35 ]; (Mazilu et al, 2015): [ 42 ]; (Zhang et al, 2022): [ 44 ]; (Huang et al, 2024): [ 54 ]; (Xia et al, 2024): [ 55 ]; (Khosla et al, 2024): [ 56 ]; (Sun et al, 2024): [ 57 ]; (Dimoudis et al, 2023): [ 60 ]; (Borzi et al, 2023): [ 65 ]; (Halder et al, 2021): [ 76 ]; (Esfahani et al, 2021): [ 80 ]; (Bikias et al, 2021): [ 81 ]; (Basaklar et al, 2021): [ 82 ]; (Suppa et al, 2017): [ 83 ]; (Ghosh et al, 2021): [ 84 ]; (Li et al, 2020): [ 85 ]; (Demrozi et al, 2020): [ 87 ]; (Kleanthous et al, 2020): [ 90 ]; (San-Segundo et al, 2019): [ 93 ]; (Naghavi et al, 2019): [ 94 ]; (Guo et al, 2019): [ 96 ]; (Arami et al, 2019): [ 98 ]; (Orphanidou et al, 2018): [ 100 ]; (Pham et al, 2017): [ 106 ]; (Palmerini et al, 2017): [ 107 ]; (Rezvanian et al, 2016): [ 114 ]; (Mazilu et al, 2012): [ 123 ]. …”

“…Some, on the other hand, placed more emphasis on specificity and sensitivity [ 46 , 110 , 122 ], concentrating on the model’s accuracy in identifying positive occurrences (FOG events) or negative instances (non-FOG events). Alternative measures, such as Area Under the Curve (AUC), geometric mean, and Mean Average Precision (MAP) score, were utilized by a selection of studies [ 59 , 80 , 89 , 92 ]. These measures provide more detailed information about the performance of the model than just accuracy, which may not be sufficient in some situations, particularly when there is a dominance of one class in the dataset.…”