Machine-learning-assisted artificial olfactory colorimetric sensor array for bladder cancer early detection from urinary volatile organic compounds.

Bladder cancer (BC) remains difficult to detect non-invasively because current diagnostic methods are either invasive or insufficiently sensitive for early-stage disease. In this study, we developed a machine-learning-assisted artificial olfactory colorimetric sensor array (CSA) for urinary volatile organic compound (VOC)-based BC screening. The CSA was fabricated by integrating eight cross-reactive dyes into electrospun poly(vinylidene fluoride) (PVDF) nanofibrous membranes. The porous and hydrophobic PVDF matrix enabled uniform dye distribution, reduced humidity interference, and improved colorimetric reproducibility. Urine samples from 134 subjects, including 81 BC patients and 53 controls, were tested using a static headspace exposure method. Colorimetric changes before and after exposure were quantified by Euclidean distance (ED), and the resulting eight-dimensional response fingerprints were analyzed using a multilayer perceptron (MLP) classifier. Based on five-fold cross-validation, the MLP model achieved an accuracy of 0.806 ± 0.078, recall of 0.913 ± 0.050, F1-score of 0.851 ± 0.059, and AUC of 0.861 ± 0.042, outperforming conventional machine-learning models in overall classification performance. SHAP analysis and occlusion-based feature ablation further identified Dye 6 (bromocresol green) as the most influential sensing unit. These results suggest that electrospun CSA-based urinary VOC fingerprinting may provide a low-cost and non-invasive strategy for BC screening.