Beta-Validation of a Non-Invasive Method for Simultaneous Detection of Early-Stage Female-Specific Cancers

ObjectivesIn this report, we assessed the accuracy of our previously developed method for simultaneous diagnosis of the four female cancers of the Breast, Endometrium, Cervix, and Ovary, in a clinical set-up with blinded protocol. Materials and MethodsOur test protocol combined global serum metabolome profiling wherein data was analyzed with machine learning algorithms to extract metabolite signatures that correlated with early-stage cancers. High-resolution mass spectrometry was employed to profile the serum metabolome and the resulting data were subjected to a pre-processing pipeline to obtain the data set. The data was then analyzed using artificial intelligence algorithms to identify early-stage cancer metabolic signatures. ResultsOverall, a total of 1000 blinded samples were analyzed by generating the serum metabolome profiles, followed by sequential algorithms for cancer detection and multiclass cancer type identification. Of these 1000 samples, 797 were identified as cancer positive, while, 203 samples were identified as cancer-negative. The multiclass algorithm was then applied to the 797 cancer-positive samples, to distinguish between samples that were from patients with either endometrial, breast, cervical, or ovarian cancer. After completion of the analysis, the sample code was broken to estimate the accuracy of the results. Concerning the identification of samples that were cancer-positive, the sensitivity obtained was 99.6% whereas the specificity was 100%. For the second stage of analysis which involves tissue of origin, all 107 breast cancer samples were correctly identified without any false calls. The accuracy for identification of cervical and ovarian cancers was between 95-96% for each, whereas 91% for endometrial cancer. ConclusionsOur present study validates the performance of our method for the early-stage detection of female-specific cancers in a clinical setting. Importantly, the algorithms for cancer detection and tissue of origin prediction, which were initially trained using samples from Caucasian patients, retained the accuracy on samples from Indian women patients. This suggests that the performance of these algorithms was minimally influenced by variables such as ethnicity and race. Present results, therefore, also underscore the potential clinical utility of our method for early-stage diagnosis of cancers that are specific to females.