Integrated machine learning and signal understanding for movement disorder recognition
Embargo Date
Indefinite
OA Version
Citation
Abstract
In this thesis, we establish that the IPUS approach for the Integrated Processing and Understanding of Signals can be used to address the signal representation challenges in an application involving complex biosignals from a highly unscripted and unconstrained real-world environment. Specifically. we have designed an IPUS-based system to detect movement disorders on a per-second basis in Parkinson's disease (PD) patients wearing small numbers of surface electromyographic (sEMG) and triaxial accelerometric (ACC) sensors. It is also designed to operate without the need
for any patient-specific training. When evaluated on a database of 44 hours of sensor signals, our system is found to yield detection error rates significantly below 10% for each of the movement disorders. Such low error rates have previously been reported only for systems that severely constrain the physical activities performed by the PD patients.
In developing the signal processing for the IPUS system, we first established that various short-term autocorrelation-based features of the sElVIG and ACC signals are an appropriate basis for the classification of tremor, dyskinesia, and freezing-of-gait. Next, as a key innovation within the IPUS approach, we designed machine learning algorithms (dynamic neural networks, dynamic support vector machines, and hidden Markov model algorithms) that on the basis of those features produce numerical ratings between -1 and + 1 for every one second of the input data. We then designed rule-based signal understanding algorithms that convert the sequences of ratings into inferences about the presence or absence of the disorders and that detect various conditions that are then used as triggers to prompt changes in the signal processing. This triggering mechanism completes the IPUS loop for changing the signal processing on the fly, on a per-need basis.
In conclusion, the research results presented in this thesis illustrate the power of combining machine learning algorithms and !PUS-based adaptation of signal processing in order to successfully address a complex biosignal application. Our results also constitute a proof of principle and a potential blueprint for the development of solutions to other complex real-world biosignal problems.
Description
Thesis (Ph.D.)--Boston University
PLEASE NOTE: Boston University Libraries did not receive an Authorization To Manage form for this thesis or dissertation. It is therefore not openly accessible, though it may be available by request. If you are the author or principal advisor of this work and would like to request open access for it, please contact us at open-help@bu.edu. Thank you.
PLEASE NOTE: Boston University Libraries did not receive an Authorization To Manage form for this thesis or dissertation. It is therefore not openly accessible, though it may be available by request. If you are the author or principal advisor of this work and would like to request open access for it, please contact us at open-help@bu.edu. Thank you.