Smartphone-based video deep learning enables rapid and accurate lateral flow diagnostics

Lateral flow assays (LFAs) are widely used in point-of-care diagnostics due to their simplicity, portability, and rapid results. Despite their advantages, LFAs face challenges in terms of sensitivity and quantitative accuracy, particularly when relying on static end-point measurements, which capture a single time point rather than dynamic changes throughout the assay. In this study, we introduce a new approach to enhance LFA performance by integrating temporal video data and deep learning algorithms. We compare the diagnostic accuracy of static image-based analysis and video-based analysis using a smartphone-based system for human chorionic gonadotropin (hCG) detection. Temporal video data, captured in three distinct time segments (1–4, 4–7, and 7–10 min after the reaction starts), were processed using a YOLOv8-based deep learning model, which improved diagnostic performance compared to static images. The video-based method achieved a sensitivity of 92.7%, specificity of 99.1%, and overall accuracy of 98.4% within the 1–4 min time segment, surpassing the static image method (87.3%, 98.4%, and 97.2%, respectively). Moreover, the video approach facilitated faster results, with accurate detection in just 4 min, compared to 15 min used in traditional methods. These findings demonstrate that incorporating temporal video sequences into deep learning models can enhance the sensitivity, specificity, and speed of LFA-based diagnostics, offering a promising strategy for improving point-of-care testing. This approach holds potential for broader applications in clinical diagnostics, enabling faster and more accurate biomarker detection.