Data-Driven Machine-Learning Quantifies Differences in the Voiding Initiation Network in Neurogenic Voiding Dysfunction in Women With Multiple Sclerosis
- Author:
Christof KARMONIK
1
;
Timothy BOONE
;
Rose KHAVARI
Author Information
- Publication Type:Original Article
- Keywords: Neurogenic lower urinary tract dysfunction; Multiple sclerosis; Functional magnetic resonance imaging; Machine learning
- MeSH: Area Under Curve; Brain; Female; Gyrus Cinguli; Humans; Machine Learning; Magnetic Resonance Imaging; Multiple Sclerosis; Urinary Tract; Urodynamics
- From:International Neurourology Journal 2019;23(3):195-204
- CountryRepublic of Korea
- Language:English
- Abstract: PURPOSE: To quantify the relative importance of brain regions responsible for reduced functional connectivity (FC) in their Voiding Initiation Network in female multiple sclerosis (MS) patients with neurogenic lower urinary tract dysfunction (NLUTD) and voiding dysfunction (VD). A data-driven machine-learning approach is utilized for quantification. METHODS: Twenty-seven ambulatory female patients with MS and NLUTD (group 1: voiders, n=15 and group 2: VD, n=12) participated in a functional magnetic resonance imaging (fMRI) voiding study. Brain activity was recorded by fMRI with simultaneous urodynamic testing. The Voiding Initiation Network was identified from averaged fMRI activation maps. Four machine-learning algorithms were employed to optimize the area under curve (AUC) of the receiver-operating characteristic curve. The optimal model was used to identify the relative importance of relevant brain regions. RESULTS: The Voiding Initiation Network exhibited stronger FC for voiders in frontal regions and stronger disassociation in cerebellar regions. Highest AUC values were obtained with ‘random forests’ (0.86) and ‘partial least squares’ algorithms (0.89). While brain regions with highest relative importance (>75%) included superior, middle, inferior frontal and cingulate regions, relative importance was larger than 60% for 186 of the 227 brain regions of the Voiding Initiation Network, indicating a global effect. CONCLUSIONS: Voiders and VD patients showed distinctly different FC in their Voiding Initiation Network. Machine-learning is able to identify brain centers contributing to these observed differences. Knowledge of these centers and their connectivity may allow phenotyping patients to centrally focused treatments such as cortical modulation.
