Results

1. FOG Detection

Using the signal model FOG detection algorithm , FOG instances are identified as shown in the plots below. The FOG instances are shown in the last plot.

FOG Detection

FOG Detection

2. FOG Classification

Among the full feature sets shown in Method tab, a feature importance ranking is done using random forest feature permutation. As shown in the table below, only the top 8 features are used for classification.

Ranking Feature Weight
1 ZuPT Entropy 0.254
2 Yaw Angle Amplitude 0.199
3 FI 0.133
4  Pitch angular velocity 0.111
5  x direction acceleration 0.108
6  z direction acceleration 0.094
7 Average contact time 0.094
8 Cycle length 0.081
9 Average stride length small
10  Average foot velocity
11  y direction acceleration
12  Roll angular velocity

For FOG classification, Naïve Bayes, Random Forests and SVM are evaluated using a dataset of 328 FOG instances. 50% of data is used for training and 50% is used for testing.Table below presents the classification accuracy of the three algorithms. The Random Forests and SVM demonstrate high classification accuracy.

Classification Accuracy Naïve Bayes Random Forests SVM
Initiation Freeze 20.0% 100.0% 100%
Walking Freeze 70.0% 97.8% 95.5%
Turning Freeze 70.7% 83.3% 85.2%
Overall 65.3% 90.6% 92.2%

3. FOG Prediction

For FOG prediction, only the binary prediction of FOG occurrence is considered in the scope of this project. Table below shows the prediction accuracy of feature space approach using random forests and time series approach using neural network. The prediction algorithm does not show promising performance and is the focus for future work.

Prediction Accuracy Feature Space Time Series
FOG occur 43% 52%
No FOG occur 58% 57%