AI-Driven Workflow for Automated Colorimetric Isothermal Amplification for Robust Pathogen Detection on Portable Devices
Ke Xu, Matthew L. Cavuto, Kenny Malpartida-Cardenas, Marcus Pond, Dimbintsoa Rakotomalala Robinson, Francois Kiemde, Bernard Hernandez, Shiranee Sriskandan, Umberto D’Alessandro, Annette Erhart, Halidou Tinto, Aubrey J. Cunnington, Alison Holmes, Pantelis Georgiou, Nicolas Moser, Jesus Rodriguez-ManzanoAbstract
Point-of-Care (POC) portable devices have shown significant potential for rapid pathogen identification in limited-resource settings. Colorimetric LAMP (cLAMP)-based POC nucleic acid testing is particularly promising due to its speed, sensitivity, and simple visual readout. However, there is a need for a universal, automated color detection method to improve ease of use and reduce readout subjectivity using portable devices. We proposed an AI-driven workflow for automatic detection of multiple pathogens using cLAMP in a multitube strip format. Following sample extraction, reagent rehydration, and reaction incubation, the cLAMP panel was photographed with a portable device. The algorithm performed: (1) image preprocessing to locate tubes; (2) segmentation and cleaning to extract the liquid-filled region using deep learning and Otsu's binarization, followed by pixel cleaning; (3) empty tube and invalid control checks; and (4) color detection using relative Maximum Mean Discrepancy (MMD; a nonparametric statistical distance between the hue distributions of two tube images) to identify samples. The AI workflow was evaluated on 1,636 photos from 527 tests for the differential diagnosis of three panels, for skin-tropic viruses, respiratory viral pathogens, and malaria parasites, under various lighting conditions in POC environments. The AI workflow was compared to expert-adjudicated human visual readouts as the reference standard. Rejection of invalid photos, defined as panels containing empty tubes or failed controls, achieved 98.94%, 98.29%, and 96.68% accuracy (95% CI 97.83−99.57%, 96.97−99.15%, 94.13−98.33%) for the three panels, respectively. Misclassified images were caused by human mistakes or extreme lighting conditions. Color identification on accepted images achieved 99.97%, 100%, and 99.69% accuracy (95% CI 99.82−100% for Data set 1 and 98.27−99.99% for Data set 3; one-sided 95% lower bound 99.90% for Data set 2). This study demonstrates a reliable automatic color readout workflow for cLAMP-based nucleic acid detection, showing potential for integration into POC testing apps for cloud-based use.