Lateral ventricle volume trajectories predict response inhibition in older age—A longitudinal brain imaging and machine learning approach

Abstract

Objective In a three-wave 6 yrs longitudinal study we investigated if the expansion of lateral ventricle (LV) volumes (regarded as a proxy for brain tissue loss) predicts third wave performance on a test of response inhibition (RI). Participants and methods Trajectories of left and right lateral ventricle volumes across the three waves were quantified using the longitudinal stream in Freesurfer. All participants ( N = 74;48 females;mean age 66.0 yrs at the third wave) performed the Color-Word Interference Test (CWIT). Response time on the third condition of CWIT, divided into , and , was used as outcome measure in a machine learning framework. Initially, we performed a linear mixed-effect (LME) analysis to describe subject-specific trajectories of the left and right LV volumes (LVV). These features were input to a multinomial logistic regression classification procedure, predicting individual belongings to one of the three RI classes. To obtain results that might generalize, we evaluated the significance of a k -fold cross-validated f1-score with a permutation test, providing a p -value that approximates the probability that the score would be obtained by chance. We also calculated a corresponding confusion matrix. Results The LME-model showed an annual ∼ 3.0% LVV increase. Evaluation of a cross-validated score using 500 permutations gave an f1-score of 0.462 that was above chance level ( p = 0.014). 56% of the fast performers were successfully classified. All these were females, and typically older than 65 yrs at inclusion. For the true slow performers, those being correctly classified had higher LVVs than those being misclassified, and their ages at inclusion were also higher. Conclusion Major contributions were: (i) a longitudinal design, (ii) advanced brain imaging and segmentation procedures with longitudinal data analysis, and (iii) a data driven machine learning approach including cross-validation and permutation testing to predict behaviour, solely from the individual’s brain “signatures” (LVV trajectories).

Cognitive and MRI trajectories for prediction of Alzheimer’s disease

“…To construct subject specific trajectories for each measure we used linear mixed effects models 23 , 24 , a class of models able to produce regression models from dependent variables 25 . Our models are based on the one presented in 24 and similar to the ones employed in our previous works 19 , 26 . As the ventricles show quadratic cohort behaviour (Fig.…”

“…In a first set of analyses we defined features characterising longitudinal changes in memory function (Rey Auditory Learning Test (RAVLT)) 11 and in a more global measure of cognitive function (ADAS-Cog-13 (ADAS13)) 9 , 17 . Expecting more precise predictions by including information from MRI examinations 15 , 16 , we investigated the add-on effect of including morphometric brain measures associated with memory function (entorhinal cortex and hippocampus 14 ) and a global measure of cognitive function (the volume of the ventricles as a proxy for a global tissue loss 18 ). More specifically, we used a pipeline proposed by Mofrad et al 19 based on a combination of mixed effects and machine learning models for analysis of longitudinal data.…”

Brain aging mechanisms with mechanical manifestations

“…Similarly, Scahill et al observed an average ventricular volume expansion rate of 0.65 cm 3 /year, which they link to brain atrophy, changes in CSF dynamics, and the distribution of CSF in the ventricular and sulcal spaces ( Scahill et al, 2003 ). While most studies report total ventricular volume changes, Lundervold et al observed a left-right ventricle expansion asymmetry and reported an annual 2.9% increase of the left and 3.1% of the right ventricular volume ( Lundervold et al, 2019 ).…”

Predicting conversion to Alzheimer’s Disease in individuals with Mild Cognitive Impairment using clinically transferable features

“…There are several factors contributing to this 20 , with one crucial obstacle being that most of these algorithms are constructed using data that are expensive and/or invasive to obtain. Although the inclusion of more invasive biomarkers 21,22 and/or longitudinal data [23][24][25] would increase the predictive power of the algorithms, this information is rarely obtained in an initial clinical examination of a MCI subject. To the best of our knowledge, few studies have aimed at creating classification models based on clinically relevant features, with a study from Grassi and colleagues 26 being an exception.…”