Detection and Prediction of Freezing of Gait in Parkinson’s Disease using Wearable Sensors and Machine Learning
MetadataShow full item record
Freezing of gait (FOG), is a brief episodic absence of forward body progression despite the intention to walk. Appearing mostly in mid-late stage Parkinson’s disease (PD), freezing manifests as a sudden loss of lower-limb function, and is closely linked to falling, decreased functional mobility, and loss of independence. Wearable-sensor based devices can detect freezes already in progress, and intervene by delivering auditory, visual, or tactile stimuli called cues. Cueing has been shown to reduce FOG duration and allow walking to continue. However, FOG detection and cueing systems require data from the freeze episode itself and are thus unable to prevent freezing. Anticipating the FOG episode before onset and supplying a timely cue could prevent the freeze from occurring altogether. FOG has been predicted in offline analyses by training machine learning models to identify wearable-sensor signal patterns known to precede FOG. The most commonly used sensors for FOG detection and prediction are inertial measurement units (IMU) that include an accelerometer, gyroscope and sometimes magnetometer. Currently, the best FOG prediction systems use data collected from multiple sensors on various body locations to develop person-specific models. Multi-sensor systems are more complex and may be challenging to integrate into real-life assistive devices. The ultimate goal of FOG prediction systems is a user-friendly assistive device that can be used by anyone experiencing FOG. To achieve this goal, person-independent models with high FOG prediction performance and a minimal number of conveniently located sensors are needed. The objectives of this thesis were: to develop and evaluate FOG detection and prediction models using IMU and plantar pressure data; determine if event-based or period of gait disruption FOG definitions have better classification performance for FOG detection and prediction; and evaluate FOG prediction models that use a single unilateral plantar pressure insole sensor or bilateral sensors. In this thesis, IMU (accelerometer and gyroscope) and plantar pressure insole sensors were used to collect data from 11 people with FOG while they walked a freeze provoking path. A custom-made synchronization and labeling program was used synchronize the IMU and plantar pressure data and annotate FOG episodes. Data were divided into overlapping 1 s windows with 0.2 s shift between consecutive windows. Time domain, Fourier transform based, and wavelet transform based features were extracted from the data. A total of 861 features were extracted from each of the 71,000 data windows. To evaluate the effectiveness of FOG detection and prediction models using plantar pressure and IMU data features, three feature sets were compared: plantar pressure, IMU, and both plantar pressure and IMU features. Minimum-redundancy maximum-relevance (mRMR) and Relief-F feature selection were performed prior to training boosted ensembles of decision trees. The binary classification models identified Total-FOG or Non-FOG states, wherein the Total-FOG class included windows with data from 2 s before the FOG onset until the end of the FOG episode. The plantar-pressure-only model had the greatest sensitivity, and the IMU-only model had the greatest specificity. The best overall model used the combination of plantar pressure and IMU features, achieving 76.4% sensitivity and 86.2% specificity. Next, the Total-FOG class components were evaluated individually (i.e., Pre-FOG windows, freeze windows, and transition windows between Pre-FOG and FOG). The best model, which used plantar pressure and IMU features, detected windows that contained both Pre-FOG and FOG data with 85.2% sensitivity, which is equivalent to detecting FOG less than 1 s after the freeze began. Models using both plantar pressure and IMU features performed better than models that used either sensor type alone. Datasets used to train machine learning models often generate ground truth FOG labels based on visual observation of specific lower limb movements (event-based definition) or an overall inability to walk effectively (period of gait disruption based definition). FOG definition ambiguity may affect FOG detection and prediction model performance, especially with respect to multiple FOG in rapid succession. This research examined the effects of defining FOG either as a period of gait disruption (merging successive FOG), or based on an event (no merging), on FOG detection and prediction. Plantar pressure and lower limb acceleration data were used to extract a set of features and train decision tree ensembles. FOG was labeled using an event-based definition. Additional datasets were then produced by merging FOG that occurred in rapid succession. A merging threshold was introduced where FOG that were separated by less than the merging threshold were merged into one episode. FOG detection and prediction models were trained for merging thresholds of 0, 1, 2, and 3 s. Merging had little effect on FOG detection model performance; however, for the prediction model, merging resulted in slightly later FOG identification and lower precision. FOG prediction models may benefit from using event-based FOG definitions and avoiding merging multiple FOG in rapid succession. Despite the known asymmetry of PD motor symptom manifestation, the difference between the more severely affected side (MSS) and less severely affected side (LSS) is rarely considered in FOG detection and prediction studies. The additional information provided by the MSS or LSS, if any, may be beneficial to FOG prediction models, especially if using a single sensor. To examine the effect of using data from the MSS, LSS, or both limbs, multiple FOG prediction models were trained and compared. Three datasets were created using plantar pressure data from the MSS, LSS, and both sides together. Feature selection was performed, and FOG prediction models were trained using the top 5, 10, 15, 20, 25 or 30 features for each dataset. The best models were the MSS model with 15 features, and the LSS and bilateral features with 5 features. The LSS model reached the highest sensitivity (79.5%) and identified the highest percentage of FOG episodes (94.9%). The MSS model achieved the highest specificity (84.9%) and the lowest false positive (FP) rate (2 FP/walking trial). Overall, the bilateral model was best. The bilateral model had 77.3% sensitivity, 82.9% specificity, and identified 94.3% of FOG episodes an average of 1.1 s before FOG onset. Compared to the bilateral model, the LSS model had a higher false positive rate; however, the bilateral and LSS models were similar in all other evaluation metrics. Therefore, using the LSS model instead of the bilateral model would produce similar FOG prediction performance at the cost of slightly more false positives. Given the advantages of single sensor systems, the increased FP rate may be acceptable. Therefore, a single plantar pressure sensor placed on the LSS could be used to develop a FOG prediction system and produce performance similar to a bilateral system.
Cite this version of the work
Scott Pardoel (2021). Detection and Prediction of Freezing of Gait in Parkinson’s Disease using Wearable Sensors and Machine Learning. UWSpace. http://hdl.handle.net/10012/17634