Lung cancer is the most frequent and mortal of all types of cancer for both genders. Approximately 80% of the newly diagnosed lung cancers are non-small cell lung cancers. Early diagnosis improves the chances of survival. Machine learning allows us to process a considerable number of variables involved in this disease. Using metabolites as attributes for the analysis, we can discern lung cancer patients from healthy patients. In addition, machine learning algorithms reveal us which metabolites has a determining contribution in the classification. The objective of this study is to demonstrate the accuracy, sensitivity and specificity of a supervised learning algorithm to classify and predict non-small cell lung cancer, using concentration values found in the serum and plasma metabolome of afflicted and healthy humans. We obtained the dataset from the Metabolomics Work-bench repository, which contains 335 samples and 139 known metabolites detected. Of all the models applied, Random Forest Classifier obtained the highest accuracy. It can classify participants according to diagnosis with >75% accuracy in serum samples. Important serum metabolites for the classification included aspartic acid, fructose, and tocopherol alpha. Cystine, pyruvic acid and tocopherol alpha for plasma. The specified metabolites are strongly associated with this condition, and are potential biomarkers for the disease. By giving clues for an earlier diagnosis, this study remarkably contributes in the field of personalized medicine, and the appreciation of the biological processes of lung cancer.