Non-targeted metabolomics via high-resolution mass spectrometry methods, such as direct infusion Fourier transform-ion cyclotron resonance mass spectrometry (DI-FT-ICR MS), produces data sets with thousands of features. By contrast, the number of samples is in general substantially lower. This disparity presents challenges when analyzing non-targeted metabolomics data sets and often requires custom methods to uncover information not always accessible via classical statistical techniques. In this work, we present a pipeline that combines a convolutional neural network with traditional statistical approaches and an adaptation of a genetic algorithm. The developed method was applied to a lifestyle intervention cohort data set, where subjects at risk of type 2 diabetes underwent an oral glucose tolerance test. Feature selection is the final result of the pipeline, achieved through classification of the data set via a neural network, with a precision-recall score of over 0.9 on the test set. The features most relevant for the described classification were then chosen via a genetic algorithm. The output of the developed pipeline encompasses approximately 200 features with high predictive scores, providing a fingerprint of the metabolic changes in the prediabetic class on the data set. Our framework presents a new approach which allows to apply complex modeling based on convolutional neural networks for the analysis of high-resolution mass spectrometric data.