This work reports on the implementation of a neural network (NN) feature selection technique to identify the design features that impact the rise of spurious modes in high performance X-cut Lithium Niobate MEMS resonators. A metric based on spurious modes location, number, and peak-to-peak excursion is introduced to serve as a target function for the NN training. Measured data from 814 fabricated resonators operating between 50 MHz and 1 GHz is used to train and validate the NN and identify the most significant design features that introduce spurious responses. Number of electrode pairs, electrode length, and anchor width are found to be the driving feature of this process. To verify the NN outcome, finite element simulations based on optimized and non-optimized feature sets are finally reported, showing significant spurious modes suppression.