Comparing feature selection methods to improve the classification of Mild Cognitive Impairment patients based on Magnetoencephalography data

Background Machine learning (ML) algorithms are generating great interest for early diagnose of Alzheimer’s disease (AD). Despite several approaches have been developed, limited research has been conducted on the impact of feature selection in classification performance. In this study we compare two feature selection methods, LASSO and Elastic Nets (EN), applied to ML algorithms based on magnetoencephalography (MEG) data from mild cognitive impairment (MCI) patients. Method An eyes‐closed resting‐state MEG recording was performed by 262 participants (117 healthy controls and 145 MCI). Functional connectivity measures were estimated for all classical frequency bands using the Phase‐Locking Value. To develop the ML algorithms, specific regions of interest from the default mode network, based on the Automatic Anatomical Labelling atlas, were selected, resulting in a total of 325 features. After dividing the dataset into training (80%) and test (20%) sets, feature selection with 10000 iterations and five‐fold cross validation was performed independently for each band using LASSO and EN algorithms. Afterwards, several classification models were trained with a five‐fold cross validation and then validated on the test set, over 500 iterations. Result The classifier showing the best performance was the linear Support Vector Machine for the beta frequency band (12‐30 Hz). Given the similarity of the training and test performance values, the results reveal that overfitting was avoided when using the 45 features selected by LASSO instead of the 47 features selected by EN, reaching a validation accuracy of 77%. Additionally, LASSO consistently identified three functional connectivity links as relevant in all frequency bands but alpha: right superior frontal gyrus & right superior frontal gyrus orbital, left posterior cingulate cortex & left hippocampus, right posterior cingulate cortex & right precuneus. Conclusion ML algorithms applied to MEG data can successfully discriminate MCI patients from healthy controls. In addition, LASSO algorithm applied for feature selection leads to better classification performance. Moreover, the resulting relevant functional connectivity links involve areas affected by AD pathology even in early stages of its development (López‐Sanz et al., 2019). Further exploration of this methodology is needed, combining this data with structural neuroimaging or complementary biomarkers.