Abstract
Early detection of breast cancer remains challenging due to limitations of current screening methods, including reduced sensitivity in dense tissue, false positives that lead to additional imaging and invasive biopsies. Untargeted metabolomics using noninvasive matrices such as urine has emerged as a promising complementary approach. In this study, a voltammetric electronic tongue consisting of 12 sensors, bare and modified with three isomeric conjugated polymers, was developed to transduce urinary metabolomic differences into electrochemical fingerprints. Performance was first evaluated on artificial urine and then tested on a larger set of clinical specimens. Differential pulse voltammetry signals were preprocessed to reduce dimensionality, analyzed by PCA and PLS-DA for pattern recognition and outlier detection, and classified into cancer and control groups using a range of linear, nonlinear, and ensemble-based supervised learning. On artificial urine, PCA showed clear separation, and gradient boosting achieved the highest test accuracy (96%). In clinical urine, separation by PCA was less pronounced, whereas PLS-DA and supervised models improved discrimination, with gradient boosting yielding 97% accuracy. Overall, the results show that the proposed electronic tongue captures clinically relevant urinary signatures and that supervised methods are advantageous when moving from artificial to real-world samples.