Study design, materials and methods
Twenty-eight ambulatory female patients with stable MS and NLUTD were recruited for a study approved by the institutional review board (IRB). One patient had to be excluded due to motion artifacts. Remaining patients were separated into two groups as reported previously [1]. Group 1; voiders (V, n=15) and group 2; voiding dysfunction (VD, n=12). Voiding dysfunction was defined as having post-void residual of ≥ 40% of the maximum cystometric capacity or performing self-catheterization.
A double lumen 7Fr MRI-compatible bladder and rectal catheters were inserted to conduct urodynamic testing concurrent with the fMRI examination [1].
During the continuous acquisition of fMRI echo planar (EPI) images covering the entire volume of the brain every 3 seconds, several cycles of bladder filling/voiding were completed as tolerated by the subject. Each cycle consisted of the bladder filling (50 ml/min) until subjects reported a strong desire to void. Voiding was initiated after 30 seconds of holding. After voiding or attempt of voiding was completed, the cycle was repeated. If unable to void, bladder was emptied passively. Care was taken not too exceed 45 min for all cycles.
Significant activated brain regions (p<0.05) were identified at the time of voiding initiation using the generalized linear model (GLM) for each subject. Second level GLM analysis yielded significantly activated brain regions for each group and an average BOLD activation map averaged over all subjects was created.
A Voiding Initiation Network in MS was defined by including only the highest activated brain regions from the average BOLD activation map.
For each subject, FC of these brain regions was quantified.
Four machine learning algorithms (random forests, neural networks, generalized linear model and partial least squares), each based on a different principle algorithm, were used with the individual FC as predictor variables to classify a subject either as voider or not.
The entire dataset was split into a training set (50 %) and a test set (50 %). Ten-fold repeated cross validation with five repeats was used to train the machine-learning algorithms.
The area under the curve (AUC) of the receiver-operating characteristic curve (ROC) was used as an indicator to determine the best-performing algorithm.
Results
The following brain regions were identified as part of the Voiding Initiation Network: dorsal vagal motor nuclei, pontine storage and micturition center, periaqueductal gray (PAG), ventral tegmental area, substantia nigra, red nuclei, thalamus, cingulate, insula was well as cortical regions in the frontal, parietal and mesial temporal lobes (figure 1). Frontal brain exhibited strong connectivity while cerebellar regions showed reduced connectivity. Voiders showed higher polarization of connectivity, i.e. stronger in the frontal lobes and more disconnect in the cerebellar areas than patients with voiding dysfunction (The two best performing machine-learning algorithms yielded good values for the training data set (AUC=0.89, partial least squares and AUC=0.86, random forests). Worst performance of the generalized linear model (AUC=0.71) indicates the complexity of the data and the need for non-linear algorithms. Accuracies found from the test data set were acceptable (0.69 for partial least squares and 0.62 for random forest).
The 10 most important brain regions used in the classification were all located in the left frontal brain (middle, medial and inferior frontal gyrus) and in the left cingulate.
Interpretation of results
Female MS patients with voiding dysfunction exhibit different FC patterns than those who void spontaneously [2]. To quantify these differences, machine learning algorithms are of advantage as they allow access to the complex non-linearity of individual brain region FC for classification.
Interruption of white matter integrity in MS in female patients with voiding dysfunction inhibits formation of FC connectivity pattern observed in MS patients who void spontaneously. Inhibition is most pronounced I the left frontal brain and left cingulate.