Real-Time Freezing of Gait Prediction and Detection in Parkinson's Disease

Abstract

Freezing of gait (FOG) is a walking disturbance that can lead to postural instability, falling, and decreased mobility in people with Parkinson's disease. This research used machine learning to predict and detect FOG episodes from plantar-pressure data and compared the performance of decision tree ensemble classifiers when trained on three different datasets. Dataset 1 (n = 11) was collected in a previous study. Dataset 2 (n = 10) included six new participants and four participants from Dataset 1 who were re-tested (approximately 2 years later), and Dataset 3 (n = 21) combined Datasets 1 and 2. The prediction model trained on Dataset 3 had a 2.28% higher sensitivity and 3.09% lower specificity compared to the models trained on Dataset 1. The model trained on Dataset 3 identified 86.84% of the total FOG episodes compared to 74.31% from the model trained on Dataset 1. Also, the model using Dataset 3 identified the FOG episodes 0.3 s earlier than the model developed with Dataset 1. The model trained using Dataset 3 showed improved performance in sensitivity, identification time, and FOG identification. The improvements using the expanded dataset (Dataset 3) in this study compared to the previous model reinforce the validity and generalizability of the original model. The model was able to predict and detect FOG well and is, therefore, ready to be implemented in a FOG prevention device.

Keywords: Parkinson’s disease; freezing of gait; machine learning; plantar pressure; prediction; wearable sensors.

Figure 1

Example of data windowing and target class compositions: (a) Windows W1–W3 contain non-freezing of gait (FOG) data only; W4–W8: non-FOG and pre-FOG data; W9–W13: pre-FOG data only; W14–W18: pre-FOG and FOG data; W19: FOG data only; and W20: FOG and non-FOG data, (b) prediction model class composition [20].

Figure 2

Model trigger decision (MTD) example. Three consecutive windows (W1–W3) classified as the target class (Figure 1) result in a MTD, where the MTD instant corresponds to the end of the third window. FOG is successfully identified (before or after onset) if a MTD instant occurs within the MTD target zone. The time difference between FOG onset and MTD instant is the identification delay (ID). The period between the beginning of the MTD target zone and the FOG onset is the prediction target zone. (Figure and caption adapted from [20]).

Figure 3

Flowchart of the data processing steps in this study.

Figure 4

Normalized number of initial FOG identifications at each time interval plotted for each model. Normalization was performed by dividing the number of FOG episode identifications at each point for a dataset by the total number of FOG episode identifications for that dataset. The histogram was plotted with a fixed interval of 1 s starting at −6 s.

Figure 5

Total number of FOG occurrences for each participant. P07 shows a significantly higher number of FOG episodes compared to other participants in the dataset. P04, P05, P10, and P11 experienced no FOG episodes during testing.

Figure 6

Change in sensitivity (left) and specificity (right) with each modification on Datasets 1 and 3.