Classification of mild Parkinson's disease: data augmentation of time-series gait data obtained via inertial measurement units

Abstract

Data-augmentation methods have emerged as a viable approach for improving the state-of-the-art performances for classifying mild Parkinson's disease using deep learning with time-series data from an inertial measurement unit, considering the limited amount of training datasets available in the medical field. This study investigated effective data-augmentation methods to classify mild Parkinson's disease and healthy participants with deep learning using a time-series gait dataset recorded via a shank-worn inertial measurement unit. Four magnitude-domain-transformation and three time-domain-transformation data-augmentation methods, and four methods involving mixtures of the aforementioned methods were applied to a representative convolutional neural network for the classification, and their performances were compared. In terms of data-augmentation, compared with baseline classification accuracy without data-augmentation, the magnitude-domain transformation performed better than the time-domain transformation and mixed-data augmentation. In the magnitude-domain transformation, the rotation method significantly contributed to the best performance improvement, yielding accuracy and F1-score improvements of 5.5 and 5.9%, respectively. The augmented data could be varied while maintaining the features of the time-series data obtained via the sensor for detecting mild Parkinson's in gait; this data attribute may have caused the aforementioned trend. Notably, the selection of appropriate data extensions will help improve the classification performance for mild Parkinson's disease.

Figure 1

Data-processing pipeline used in this study. (a) Preprocessing of the input dataset for the CNN-based deep learning model comprises the filtering process using a fourth-order Butterworth filter (bandpass: 0.25–35 Hz) and the slicing process, which sliced the measured raw time-series data into 1024 slices of uniform data length. (b) The training flowchart for the CNN-based classification sequentially consists of the input data obtained via the slicing process, specific data augmentation, short-time Fourier transform (STFT), and the representative CNN-based classification model. (c) The representative CNN-based classification model is structured using 2D convolution layers with batch normalization (BN), the Rectified Linear Unit (ReLU), and maxpooling layers (Maxpool), outputting the parameters of the Softmax function for the mild Parkinson’s disease patient and healthy elderly participant.

Figure 2

Examples of data augmentation methods explored in this study. The blue line shows the original time-series not converted via data augmentation. The orange line indicates the converted time-series based on a specific data augmentation method as follows: (a) Left graph with a blue line presents the original time-series of shank-worn IMU acceleration in a human-gait axis. Right graph with an orange line shows the time-series converted based on the rotation method in data augmentation, with the blue line depicting the original time-series, allowing a result comparison before and after augmentation. (b) Left graph shows the converted time-series based on the jittering method. Right graph plots the time-series of noise added to the original time-series. (c) Left graph shows the converted time-series based on the scaling method. Right graph illustrates the change over time in the static scaling-parameter value as an example. (d) Left graph displays the converted time-series based on the magnitude warping method. Right graph shows the change over time in the value of a dynamic scaling parameter as an example. (e) Left graph depicts the converted time-series based on the permutation method. Right graph presents the segmentation parts and exchanging relationship among the segmented parts. (f) Left graph shows the converted time-series based on the time-warping method. In the right graph, the orange line indicates the value of a dynamic scaling parameter of the time-warping method compared with the timestamp of the original time-series shown as the blue straight line. (g) Left graph shows the converted time-series based on the cropping method. The last 10% length of time-series was excluded. (h) An example of a mixed data-augmentation method using the rotation and permutation methods. Left graph plots the original time-series. The center graph shows the converted time-series obtained via the rotation method from the original time-series. Right graph depicts the time-series converted via the permutation method for the time-series converted through the rotation method. This research experimentally investigated the following four types of mixed data-augmentation methods: combinations of rotation–scaling, rotation–jittering, jittering–permutation, and rotation–scaling–magnitude warping, respectively.

Figure 3

Performance evaluation results of data augmentation methods (bar-charts). The data augmentation methods explored are four methods of magnitude-domain transformation, including rotation (R), jittering (J), scaling (S), and magnitude-warping (M) methods, three methods of time-domain transformation, including permutation (P), time-warping, and cropping methods, and four methods featuring mixtures of the aforementioned methods, which are R & S, R & J, J & P, and R & S & M. The evaluated performances are reported based on the indices, which are the F1-score, accuracy, recall, and precision. The baseline of performance is also evaluated, without featuring any data augmentation methods.