Prediction and detection of freezing of gait in Parkinson’s disease from plantar pressure data using long short-term memory neural-networks

Abstract

Background Freezing of gait (FOG) is a walking disturbance in advanced stage Parkinson’s disease (PD) that has been associated with increased fall risk and decreased quality of life. Freezing episodes can be mitigated or prevented with external intervention such as visual or auditory cues, activated by FOG prediction and detection systems. While most research on FOG detection and prediction has been based on inertial measurement unit (IMU) and accelerometer data, plantar-pressure data may capture subtle weight shifts unique to FOG episodes. Different machine learning algorithms have been used for FOG detection and prediction; however, long short-term memory (LSTM) deep learning methods hold an advantage when dealing with time-series data, such as sensor data. This research aimed to determine if LSTM can be used to detect and predict FOG from plantar pressure data alone, specifically for use in a real-time wearable system. Methods Plantar pressure data were collected from pressure-sensing insole sensors worn by 11 participants with PD as they walked a predefined freeze-provoking path. FOG instances were labelled, 16 features were extracted, and the dataset was balanced and normalized (z-score). The resulting datasets were classified using long short-term memory neural-network models. Separate models were trained for detection and prediction. For prediction models, data before FOG were included in the target class. Leave-one-freezer-out cross validation was used for model evaluation. In addition, the models were tested on all non-freezer data to determine model specificity. Results The best FOG detection model had 82.1% (SD 6.2%) mean sensitivity and 89.5% (SD 3.6%) mean specificity for one-freezer-held-out cross validation. Specificity improved to 93.3% (SD 4.0%) when ignoring inactive state data (standing) and analyzing the model only on active states (turning and walking). The model correctly detected 95% of freeze episodes. The best FOG prediction method achieved 72.5% (SD 13.6%) mean sensitivity and 81.2% (SD 6.8%) mean specificity for one-freezer-held-out cross validation. Conclusions Based on FOG data collected in a laboratory, the results suggest that plantar pressure data can be used for FOG detection and prediction. However, further research is required to improve FOG prediction performance, including training with a larger sample of people who experience FOG.

Time series analysis and modelling of the freezing of gait phenomenon

“…Specifically, our analytical estimate 1 1−λ1 (equation 5.7) for the time it takes a patient to transition into freezing given a segment of stepping time series data immediately preceding a FE could be used in future algorithms for online FE prediction that could be transformative for the quality of life of patients with Parkinson's disease. Although our data set does not contain information on the step length, the use of insoles in [20,23] allows the force generated during forward walking to be measured. In this regard the temporal characteristics evaluated here should still apply to forward walking.…”

“…A variety of mathematical modelling and data analysis approaches have been applied in the context of Parkinson's gait and motor control more generally as reviewed in [22]. The phenomenon of FOG has been evaluated in terms of data-driven prediction and detection based on wearable sensors data [19], stepping in place force platform data [18], and very recently plantar pressure data [20,23]. A stochastic model of gait consisting of a random walk on a chain has been also proposed and applied to describe alterations in gait dynamics from childhood to adulthood [2].…”

Detection and assessment of Parkinson's disease based on gait analysis: A survey

“…Demrozi et al ( 2019 ) directly leveraged PCA to select informative raw data segments and reached the accuracy of 94.1% with the conventional KNN classifier. Using pressure insoles, Shalin et al ( 2021 ) extracted COP and GRF and fed them into LSTM to predict pre-FOG events with a successful rate of 72.5%, which is inferior to those using inertial sensors. With the high-precision kinematic features captured by a Mocap system, Filtjens et al ( 2021 ) used a CNN model to precede the FOG episodes, and they proposed layer-wise relevance propagation to enhance the explainability of the deep model, where the pre-FOG events can be successfully detected with a rate of 98.7%.…”

Technological support for people with Parkinson’s disease: a narrative review

“…The device was able to detect and report motor blocks. Shalin et al 44 also used a pressure-sensitive insole to collect enrolled participants' data through sensors to predict FoG episodes. Participants, wearing the insoles, had to face a path that had the aim of triggering FoG episodes through a double-task test (e.g.…”