Development of a rapid metal oxide semiconductor-based sensory system for noninvasive neonatal sepsis detection.

A portable electronic nose system, termed cNose, is presented which combines an optimized metal oxide semiconductor gas sensor array with machine learning algorithms to determine fecal volatile organic compounds (VOCs) for noninvasive bedside screening of neonatal sepsis. The system was evaluated at a public hospital in Sleman, Yogyakarta, Indonesia, using 347 fecal samples (197 positive, 150 negative for sepsis confirmed by blood culture). This balanced dataset was used for exploratory machine learning analysis and proof-of-concept evaluation rather than for population-level inference. Features from averaged sensor responses served as inputs for four classifiers: linear discriminant analysis (LDA), decision tree, random forest, and extreme gradient boosting. Mutual information-based feature selection was employed to identify the most informative sensors to reduce redundancy in the array. Cross-validation on the training set indicated that mutual information combined with LDA achieved accuracy, sensitivity, and specificity of 91.42% (95% CI: 85.53–97.30%), 90.39% (95% CI: 84.19–96.58%), and 3.04% (95% CI: 87.70–98.39%), respectively, using only six sensors. On the independent testing dataset, the model achieved 89.41% (95% CI: 82.87–95.95%) accuracy, 89.36% (95% CI: 82.81–95.92%) sensitivity, and 89.47% (95% CI: 82.95–96.00%) specificity. The findings of this study suggest that cNose could serve as a low-cost, rapid, and noninvasive screening tool, potentially reducing sampling and analysis time compared to conventional blood tests.