Autism Spectrum Disorders are associated with atypical movements, of whichstereotypical motor movements (SMMs) interfere with learning and socialinteraction. The automatic SMM detection using inertial measurement units (IMU)remains complex due to the strong intra and inter-subject variability,especially when handcrafted features are extracted from the signal. We proposea new application of the deep learning to facilitate automatic SMM detectionusing multi-axis IMUs. We use a convolutional neural network (CNN) to learn adiscriminative feature space from raw data. We show how the CNN can be used forparameter transfer learning to enhance the detection rate on longitudinal data.We also combine the long short-term memory (LSTM) with CNN to model thetemporal patterns in a sequence of multi-axis signals. Further, we employensemble learning to combine multiple LSTM learners into a more robust SMMdetector. Our results show that: 1) feature learning outperforms handcraftedfeatures; 2) parameter transfer learning is beneficial in longitudinalsettings; 3) using LSTM to learn the temporal dynamic of signals enhances thedetection rate especially for skewed training data; 4) an ensemble of LSTMsprovides more accurate and stable detectors. These findings provide asignificant step toward accurate SMM detection in real-time scenarios.
展开▼
