Date of Award
Daniel Felix Ritchie School of Engineering and Computer Science, Electrical and Computer Engineering
Mohammad Hosein Mahoor
Adam Olding Hebb
Bradley Davidson, Ph.D.
Mohammed Matin, Ph.D.
Daniel Linseman, Ph.D.
Deep brain stimulus, Parkinson's disease
Parkinson’s disease (PD) is a neurodegenerative condition and movement disorder that appears with symptoms such as tremor, rigidity of muscles and slowness of movements. Deep brain stimulation (DBS) is an FDA-approved surgical therapy for essential tremor and PD. Despite the fact that DBS substantially alleviates the motor signs of PD, it can cause cognitive side effects and speech malfunction mainly due to the lack of adaptivity and optimality of the stimulation signal to the patients’ current state. A behavior-adapted closed-loop DBS system may reduce the side effects and power consumption by adjusting the stimulation parameters to patients’ need.
Behavior recognition based on physiological feedbacks plays a key role in designing the next generation of closed-loop DBS systems. Hence, this dissertation is concentrated on: 1. Investigating the capability of local field potential (LFP) signals recorded from Subthalamic nucleus (STN) in identifying behavioral activities 2. Developing advanced machine learning algorithms to recognize behavioral activities using LFP signals 3. Investigating the effects of medication and stimulation pulse on the behavior recognition task as well as characteristics of the LFP signal.
STN-LFP is a great physiological signal candidate since the stimulation device itself can record it, eliminating the need for additional sensors. Continuous wavelet transform is utilized for time-frequency analysis of STN-LFPs. Experimental results demonstrate that different behaviors create different modulation patterns in STN within the beta frequency range.
A hierarchical classification structure is proposed to perform the behavior classification through a multi-level framework. The beta frequency components of STN-LFPs recorded from all contacts of DBS leads are combined through an MKL-based SVM classifier for behavior classification. Alternatively, the inter-hemispheric synchronization of the LFP signals measured by an FFT-based synchronization approach is utilized to pair up the LFP signals from left and right STNs. Using these rearranged LFP signals reduces the computational cost significantly while keeping the classification ability almost unchanged.
LFP-Net, a customized deep convolutional neural network (CNN) approach for behavior classification, is also proposed. CNNs learn different feature maps based on the beta power patterns associated with different behaviors. The features extracted by CNNs are passed through fully connected layers, and, then to the softmax layer for classification.
The effect of medication and stimulation “off/on” conditions on characteristics of LFP signals and the behavior classification performance is studied. The beta power of LFP signals under different stimulation and medication paradigms is investigated. Experimental results confirm that the beta power is suppressed significantly when the patients take medication or therapeutic stimulation. The results also show that the behavior classification performance is not impacted by different medication or stimulation conditions.
Identifying human behavioral activities from physiological signals is a stepping-stone toward adaptive closed-loop DBS systems. To design such systems, however, there are other open questions that need to be addressed, which are beyond the scope of this dissertation, such as developing event-related biomarkers, customizing the parameter of DBS system based on the patients’ current state, investigating the power consumption and computational complexity of the behavior recognition algorithms.
Copyright is held by the author. User is responsible for all copyright compliance.
Hosein Golshan Mojdehi
Received from ProQuest
Golshan Mojdehi, Hosein, "Developing Machine Learning Algorithms for Behavior Recognition from Deep Brain Signals" (2020). Electronic Theses and Dissertations. 1768.
Computer engineering, Electrical engineering