Endometrial cancer (EC) is the most common cancer of the female reproductive tract in developed countries. At the moment, no effective screening system is available. Here, we evaluate the diagnostic performance of a serum metabolomic signature. Two enrollments were carried out, one consisting of 168 subjects: 88 with EC and 80 healthy women, was used for building the classification models. The second (used to establish the performance of the classification algorithm) was consisted of 120 subjects: 30 with EC, 30 with ovarian cancer, 10 with benign endometrial disease, and 50 healthy controls. Two ensemble models were built, one with all EC versus controls (Model I) and one in which EC patients were aggregated according to their histotype (Model II). Serum metabolomic analysis was conducted via gas chromatography-mass spectrometry, while classification was done by an ensemble learning machine. Accuracy ranged from 62% to 99% for the Model I and from 67% to 100% for the Model II. Ensemble model showed an accuracy of 100% both for Model I and II. The most important metabolites in class separation were lactic acid, progesterone, homocysteine, 3-hydroxybutyrate, linoleic acid, stearic acid, myristic acid, threonine, and valine. The serum metabolomics signature of endometrial cancer patients is peculiar because it differs from that of healthy controls and from that of benign endometrial disease and from other gynecological cancers (such as ovarian cancer).
Keywords: classification models; endometrial cancer; ensemble machine learning models; metabolomics; partial least square discriminant analysis.