Naturally occurring longitudinal data such as electronic health records are in great utility to derive evidence-based data-driven decision in this precision medicine era. Mixture-based multilevel analyses of curve approximation are mostly used for identifying subgroups of homogeneous trajectories in these datasets. The methods are implemented only in M-plus and R among standard software packages. Limited by computational complexities that increases with sample size, degree of unbalancedness, curve complicacies, and number of random effects; methods often yield convergence problem and inconsistent results. Loess, an alternative flexible non-parametric curve approximation technique, usually captures the general patterns of the curves when approximating values at fixed common time points. Application of mixture model on the correlated variables of approximated values can identify subgroups of distinct trajectories using eigenvalue decomposition of the covariance matrix. This study appraised the classification performance of this method as well as compared with that of existing methods using real and simulated data of varying levels of complexities.