Automated machine learning framework for interpreting paper-based biosensing of pathogens

Paper-based biosensors have emerged as a promising tool for the rapid detection of microbial contaminants. However, the interpretation of paper-based biosensor images commonly relies on manual intensity extraction and assay-specific decision criteria, while explicit thresholds for positive/negative classification are often not reported. Herein, an ML framework was developed for automated interpretation of paper-based colorimetric biosensor images across distinct paper-based biosensor studies targeting different pathogens. Using conventional normalized intensity, region-of-interest (ROI)-based descriptors, and image-only convolutional neural network approaches, within-study validation showed strong positive/negative classification performance. However, the optimized decision thresholds differed between studies, and this study-dependent shift resulted in degraded cross-study classification performance. We therefore introduced a color-space normalization approach that quantified each target signal according to its chromatic relationship with the paired positive and negative controls, rather than relying on single-channel fluorescence intensity alone. This color-space normalization improved the separation of positive and negative readouts across studies and yielded comparable decision-threshold ranges in both within-study and cross-study validation. Accordingly, the color-space normalization approach provided more stable cross-study classification than conventional single-channel normalization, ROI-based descriptors, and image-only deep learning approaches. The trained model and frozen decision threshold were applied to a separate dataset as a small proof-of-concept application. These results support ML-assisted biosensor interpretation as a means to automate image interpretation and establish more consistent decision boundaries for paper-based biosensor studies.