Machine learning applied to electrical impedance tomography for the improved management of nocturnal enuresis

Dunne, Eoghan
Nocturnal Enuresis (or bedwetting) is a condition that commonly affects children but can continue into adulthood. The condition can have serious implications on the quality of a child’s life. The current first-line medical treatment options are limited and none alert the user before voiding occurs. Electrical Impedance Tomography (EIT) is a low-cost, wearable and safe imaging modality that is investigated in this thesis as a potential proactive bedwetting alarm technology for nocturnal enuresis. In the literature, methods to estimate the bladder volume have been established in EIT. However, no research has been performed to determine when to alert the user in time to void their bladder. Consequently, this thesis investigates classification of bladder fullness in terms of not-full and full states using machine learning. After reviewing the literature, numerical forward models were critically investigated. Using anatomical information, a novel three-dimensional EIT pelvic numerical forward model was devised. Bladder state classification using machine learning was employed with both EIT measurement data and images. The classifiers were tested against varying noise sources previously established in the literature. A novel child pelvic numerical forward model (with realistic anatomical and electrical conductivity information) was then designed and implemented for bladder monitoring studies. Bladder state classification was then tested on child EIT data for the first time. Feature processing and classifier optimisation were also examined to improve classification performance and memory storage. Finally, the classifiers were investigated against the electrode misplacement noise source using EIT measurement data from the numerical child model. The overall findings indicated that high performance can be achieved with EIT bladder state classification using machine learning in the presence of established sources of error. Electrical impedance tomography measurement data was found to be the best input to the classifiers. Feature processing and classifier optimisation were found to improve the performance of an EIT-based bedwetting alarm solution. Electrode misplacement was found to greatly reduce classification performance. However, it has been shown that this error can be reduced by including data from varying electrode misplacements in the training dataset. Overall, this thesis has made significant strides towards the design and development of a non-invasive, proactive bladder monitor for children with nocturnal enuresis based on EIT.
NUI Galway
Publisher DOI
Attribution-NonCommercial-NoDerivs 3.0 Ireland