Many studies have been carried out related to the analysis of cognitive workload assessment using the N-back task. However, fixed analytic functions like time-frequency spectrum and wavelet-based approaches have been primarily used to analyze non-stationary EEG signals. Moreover, these approaches removedredundant information present in the input features by implementing the feature reduction approaches like correlation analysis and Principal Component Analysis, which are primarily based on the assumption of linearity in the input features. In the present work, we have analyzed multichannel EEG data for the N-back (0, 1, 2-back) task using a data-driven technique called multivariate empirical mode decomposition (MEMD). MEMD breaks down multichannel data into a fixed number of intrinsic mode functions (IMFs). Various features have been extracted from each IMF based on statistical parameters (variance, skewness, and kurtosis), spectral power (related to brain waves: delta, theta, alpha, beta, and gamma), and parameters based on time-series data (relativeMEMD energy and zero-crossing rate). The effective feature reduction is obtained by kernel principal component analysis (kPCA). These new reduced transformed features are taken as input for training and testing different machine learning (ML) models viz K-nearest neighbor (KNN), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and random forest. The best average classification accuracy of 97.34% could be achieved using KNN with kPCA transformed features (based on third-order polynomial kernel function). Theproposed approach performs better in classifying the N-back EEG data than earlier techniques.