Predicting Scores of Repetitive Movement Measurements using Image Classification
To test the hypothesis that image classification by a convolutional neural network accurately classifies the transformed images of accelerometer-based signals produced by repetitive movements, an automated method predicted motor impairment scores of repetitive movement measurements using image classification. This method could augment ratings generated by visual observations of trained raters. The ability of the image classification network to identify and classify images containing pathognomonic signatures of Parkinson’s disease (PD) and parkinsonian syndromes can also be investigated. A dataset containing repetitive movement measurements from patients with PD and age- and sex-matched control participants was used (Harrigan et al., 2020). A low-cost, accelerometer-based protocol (Goetz, et al., 2008; McKay, et al., 2019) was administered to obtain the movement measurements from the extremities of participants. The technologist began recording on the instrumentation prior to the participant receiving instruction to perform each item. An examiner scored the movements live and the instrumentation recorded the movements. Participants with PD completed a single test session (0002, 0005, 0007-0009, 0012, 0017-0018, and 0021), a test and a retest session (0001, 0003, 0006, 0010, 0011, 0013, 0015, 0019, 0022-0023), or a test and two retest sessions (0014). Control participants completed test and retest sessions (0020, 0024-0030). A participant with MSA-P (0004) completed a test session. Another participant (0016) consented to the study, however, no ratings were completed. Data from the instrumentation was saved as WinDaq files (Dataq Instruments, Inc., Akron, Ohio) and converted into Excel files (McKay, et al., 2019) using the WinDaq Waveform Data Browser (Dataq Instruments, Inc., Akron, Ohio) (Brasic, et al., 2017, 2018, 2019, 2020; Harrigan, et al., 2017, 2018, 2019); Harrigan, Hwang, et al., 2020; Harrigan, Syed, et al., 2020; Hwang, et al., 2017; Ziegelman, et al., 2020). For each trial of a repetitive movement, ten-second segments of accelerometer data were selected. Only data collected from instrumentation on the index finger for the upper extremity and the big toe for the lower extremity were used and instrumentation data from the wrist and ankle were omitted. The selected data were converted into continuous wavelet transforms (CWTs) and short-term Fourier transforms (STFTs) and the bone colormap in MATLAB was used for visualization (The Math Works, Inc., Natick, Massachusetts). The generated images were sorted into five folders (0-4) corresponding to the score for that trial from the examiner’s live rating. A higher number corresponded to a greater impairment score.
Steps to reproduce
Image classification was performed using GoogLeNet in MATLAB using the Deep Learning Toolbox (The Math Works, Inc., Natick, Massachusetts). The CWT images and STFT images were divided into two classes: low (corresponding to ratings 0-1 suggesting no or slight impairment) and high (corresponding to ratings 3-4 suggesting moderate to severe impairment). An equal number of images was selected randomly from each class and used for classification. An 80-20 split for training and testing images was selected. A custom dropout layer was created to prevent overfitting. Training options were customized so that a stochastic gradient descent with momentum optimizer was used. The execution environment was set to CPU for reproducibility. The network’s ability to correctly classify the testing images is used to determine an accuracy score.