0026
A Machine Learning approach to Predict Age-Related Motor Performance using MRI-MRS data1University of Leuven, leuven, Belgium, 2Philips UK, Belfast, United Kingdom, 3icometrix, leuven, Belgium
Synopsis
Aging is associated with gradual alterations in structural and neurochemical characteristics of the brain, which can be assessed in vivo by MRI and MRS modalities. The process of brain aging occurs in accord with a general decline in cognitive-motor performance and increases the risk of neurodegeneration. We used MRI-MRS data from 86 individuals as inputs for machine learning models5 to predict motor performance in healthy individuals. Our analysis shows that application of machine learning algorithms on combination of age, gender and MR data can accurately predict motor performance and has potential to be used as a biomarker for neuro related diseases.
Introduction
Aging is associated with gradual alterations in structural (gray and white matter) and neurochemical characteristics of the brain, which can be assessed in vivo by magnetic resonance imaging (MRI) [1] and proton magnetic resonance spectroscopy (1H-MRS) [2-3]. The process of brain aging occurs in accord with a general decline in cognitive and motor performance [4]. Although the changes associated with brain aging are not explicitly pathological, the risk of neurodegenerative disease and dementia increases with age [5]. However, the wide range of onset ages for age-associated brain diseases indicates that the effects of aging on the brain vary greatly between individuals. Therefore, advancing our understanding of brain aging and identifying biomarkers of the process are vital to help improve detection of early-stage neurodegeneration and predict age-related cognitive and motor performance. One approach to identifying individual differences in brain aging is to use MRI and MRS data as inputs for machine learning models to predict motor performance in healthy individuals [6]. By correlating structural and neurometabolic MR data with task performance, machine-learning algorithms such as Random Forest [7] can formulate regression models fitting MR datasets as independent variables to predict motor performance as a dependent variable.Materials and Methods
Whole brain tissue fractions and brain metabolite concentrations (N-acetylaspartate (NAA), Glutamate-Glutamine (Glx), Creatine (Cr), Choline (Cho) , Taurine (Tau) and Myo-inositol (mI)) were acquired from 86 apparently healthy subjects. Two motor tasks, Purdue pegboard task (PPT) [8] and Bimanual coordination task (BCT) [9], were used in this study. A detailed description of data used in this study can be found in Table 1. Data was randomly split: 70% into training (60 patients) and 30% into testing (26 patients) data sets. The root mean squared error (RMSE) was computed for the testing set. The regression experiment was run 1000 times for BCT and PPT, in order to robustly estimate if different feature groups will lead to statistically significant differences within RMSE values. Two non-parametric Kruskal-Wallis rank tests were applied for PPT and BCT separately, followed by Dunn-Šidák's post-hoc tests to determine which feature group had the lowest rank (lowest RMSE). Finally, feature importance for BCT and PPT prediction was estimated from the entire dataset using Random Forest with 100000 trees. PPT values and BCT values were separately predicted based on four groups of features: group 1 (age and gender), group 2 (MR spectroscopy features), group 3 (MR whole brain volume tissue fractions), and group 4 (all features). Prediction was done using Random Forest regressor with 1000 trees [10] from Python’s scikit-learn module.Results
A detailed description of RMSE% (100*RMSE divided by average BCT or PPT) mean values, standard deviation, 5th percentile, and 95th percentile can be found in Table 2 for predicting BCT and PPT using different features. Figures 1 and 2 show BCT and PPT RMSE rank estimates and intervals for different feature groups. Intervals are shown as horizontal lines, while rank mean estimates are in the middle of the intervals as either filled diamonds or empty circles. Two groups are significantly different if their intervals are disjoint. Feature importance values are shown in Table 3.Discussion
Regression data evaluated by RMSE% indicates that the BCT prediction accuracy is relatively higher than PPT for all features (Table 2). The RMSE rank estimates show that combining all features leads to better predictions for both BCT and PPT, while the next best feature group is age and gender. A closer analysis into feature importance shows that age accounts for over 70% of the total variation when predicting BCT, and only 30% when predicting PPT. CSF and Creatine are the second and third most important features for predicting PPT. Gender does not have any influence in predicting BCT, but it is as significant as NAA and grey matter fraction in predicting PPT. The differences in feature importance may be due to the task-dependency where BCT is designed specifically to test multiple coordination constraints (i.e., temporal and spatial) as well as cognitive speed and attentional control whereas the PPT tests focuses mainly on upper limb dexterity and gross and fine motor skills.Conclusions
Our analysis shows that training Random Forest on a combination of age, gender, MRI and MRS data can accurately predict motor performance. This type of predictive techniques may allow physicians to identify people at risk of developing neurodegenerative disorders.Acknowledgements
No acknowledgement found.References
1. Farokhian F, Yang C, Beheshti I, Matsuda H, Wu S. Age-Related Gray and White Matter Changes in Normal Adult Brains. Aging Dis. 2017;8(6):899-909.
2.Ding, X.Q., Maudsley, A.A., Sabati, M., Sheriff, S., Schmitz, B., Schütze, M., Bronzlik, P., Kahl, K.G., Lanfermann, H., 2016. Physiological neuronal decline in healthy aging human brain - An in vivo study with MRI and short echo-time whole-brain1H MR spectroscopic imaging. NeuroImage 137, 45–51.
3. Duarte, J.M.N., Do, K.Q., Gruetter, R., 2014. Longitudinal neurochemical modifications in the aging mouse brain measured in vivo by 1H magnetic resonance spectroscopy. Neurobiol. Aging 35,1660 – 1668.
4.Bernard JA, Seidler RD (2014). Moving forward: age effects on the cerebellum underlie cognitive and motor declines. Neurosci Biobehav Rev. 42, 193–207.
5. Abbott, A., 2011. Dementia: A problem for our age. Nature 475, S2-S4.
6.Cole, J.H., Poudel, R.P., Tsagkrasoulis, D., Caan, M.W., Steves, C., Spector, T.D., Montana, G., 2016. Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker. arXiv preprint arXiv:1612.02572.
7. Denisko D, Hoffman MM (February 2018). "Classification and interaction in random forests". Proceedings of the National Academy of Sciences of the United States of America. 115 (8): 1690–1692.
8. Serbruyns L., Gooijers J., Caeyenberghs K., Meesen R. L., Cuypers K., Sisti H. M., et al. (2015). Bimanual motor deficits in older adults predicted by diffusion tensor imaging metrics of corpus callosum subregions. Brain Struct. Funct. 220 273–290.
9. Sisti H. M., Geurts M., Clerckx R., Gooijers J., Coxon J. P., Heitger M. H., et al. (2011). Testing multiple coordination constraints with a novel bimanual visuomotor task. PLoS ONE 6:e2361910.1371/journal.pone.0023619.
10. Breiman, “Random Forests”, Machine Learning, 45(1), 5-32, 2001 from Python’s scikit-learn module.
Figures
