Harnessing Mechanistic Simulators for Rapid Diagnostic Test Capture and Deep Learning Classification

Rapid Diagnostic Tests (RDTs) are affordable tools for disease diagnosis and surveillance. There is growing evidence of the benefits of Machine Learning models to aid accurate decision making through classification of test results, improving sensitivity and specificity and reducing the risk of false positives/negatives. Yet to date these models have relied on large, clinically sensitive, costly, and time-intensive real-world image training libraries. Herein, we present SynSight – a classification pipeline for RDTs trained using mechanistically generated synthetic data. SynSight combines test region of interest segmentation using real-time object detection and result classification. We validate SynSight with HIV RDTs achieving 98.2% sensitivity and 99.2% specificity, and further with COVID-19 RDTs (two antigen, achieving 99% and 97% accuracy; one antibody, achieving 95% accuracy). This agile approach opens training opportunities for machine learning classifiers for RDTs with no available real-world image libraries, keeping pace with new RDT development and innovation.