Abstract

Inertial measurement sensing technology with the capability of capturing disease-relevant data has a great potential for improving the current clinical assessments and enhancing the quality of life in patients with neuro-developmental and neuro-degenerative diseases such as autism spectrum disorders (ASD) and Parkinson's disease (PD). The current clinical assessments can be improved by developing objective tools for the disease diagnosis and continuous monitoring of patients in out of clinical settings. To this end, it is necessary to develop automatic abnormal movement detection methods with the capability of adjusting on new patients' data in real-life settings. However, achieving this goal is challenging mainly because of the inter and intra-subject variability in acquired signals and the lack of labeled data. The research presented in this thesis investigates the application of deep neural networks to address these challenges of abnormal movement detection using inertial measurement unit (IMU) sensors with case studies on stereotypical motor movements in ASD and freezing of gait in PD patients. In this direction, this thesis provides four main contributions: i) A convolutional neural network (CNN) architecture is proposed to learn discriminative features which are sufficiently robust to inter and intra-subject variability. It is further shown how the proposed CNN architecture can be used for parameter transfer learning to enhance the adaptability of the abnormal movement detection system to new data in a longitudinal study. ii) An application of recurrent neural networks and more specifically long short-term memory (LSTM) in combination with CNN is proposed in order to incorporate more the temporal dynamics of IMU signals in the process of feature learning for abnormal movement detection. iii) An ensemble learning approach is proposed to improve the detection accuracy and at the same time to reduce the variance of models. iv) In the normative modeling framework, the problem of abnormal movement detection is redefined in the context of novelty detection and it is shown how a probabilistic denoising autoencoder can be used to learn the distribution of the normal human movements. The resulting deep normative model then is used in a novelty detection setting for unsupervised abnormal movement detection. The experimental results on three benchmark datasets collected from ASD and PD patients illustrate the high potentials of deep learning paradigm to address the crucial challenges toward real-time abnormal movement detection systems using wearable technologies.