Author Of 1 Presentation
P0002 - Applying machine learning to multimodal neuroimaging data to predict visual episodic memory performance in multiple sclerosis (ID 1530)
Abstract
Background
Episodic memory (EM) impairment is common in MS. The thalamus, hippocampus, fornix and cingulum are important hubs in the Papez circuit involved in EM processing. The extent to which MS-related early structural and functional changes in these regions relate to visual EM is not known.
Objectives
To build a brain model which represents multimodal MRI measures mostly associated with visual EM processing in early MS using a machine learning approach (Elastic Net-EN).
Methods
A computerized nonverbal test of visual episodic memory (CANTAB Paired Associate Learning-PAL) was administered to 180 MS patients (RADIEMS cohort: 162 relapsing remitting, 18 clinically isolated syndrome, 123 female, mean age = 34.4 ± 7.5, mean years since diagnosis = 2.1 ± 1.4, mean premorbid intelligence = 108.3 ± 8.8). Raw scores of first trial memory were selected for statistical assessment. Neuroimaging data were acquired via a 3T MRI scanner. Lesion load over 12 regions across the brain, volume of bilateral 12 hippocampal subfields, 25 thalamic nuclei, functional and structural connectivity (assessed by fractional anisotropy and mean diffusivity) between bilateral subiculum of hippocampus and anterior thalamic nuclei and structural connectivity of fornix and cingulum were used to build a model predicting EM. Age, sex, IQ and EDSS scores were also included in the model. Five-fold cross-validation was used to train and test the model with 50 repetitions. Pearson’s correlation (r) was used to assess the univariate relationships between the continuous variables and model’s prediction accuracy.
Results
PAL test score was positively correlated with IQ (r = 0.24, p < 0.05), and negatively correlated with age and EDSS score (r = -0.25 and r = -0.23, p < 0.05). The correlation between the model’s predicted memory and actual memory scores was 0.26 ± 0.14, indicating moderate performance similar to pevious literature predicting cognitive scores from imaging variables. The most important MRI predictors of better memory were right hippocampus molecular layer and right thalamus parataenial nucleus volume, functional connectivity between left thalamus anteroventral nuclei and left hippocampus subiculum, left thalamus medial pulvinar and right hippocampus CA4 volume and lesion load in medulla and limbic lobe.
Conclusions
Atrophy and lesions in select regions of Papez Circuit are important imaging predictors of memory in early MS. A next essential step to understanding disease-specificity of these findings is comparison to matched health controls.