Portable multiclass lateral flow assay for dry eye disease based on machine learning and smartphone imaging

Dry eye disease (DED) is a widespread ocular surface disorder that requires early and accurate diagnostics for effective management. Among various biomarkers, matrix metalloproteinase-9 (MMP-9) has been validated as a reliable indicator of ocular inflammation. However, current lateral flow assays (LFAs), such as InflammaDry, provide only binary outcomes, which limits their ability to stratify disease severity and guide personalized treatment. More advanced and accessible diagnostic strategies are needed to address these gaps. This paper presents a portable smartphone-based platform that integrates machine learning (ML) with LFAs for multiclass detection of tear MMP-9. Customized LFA strips were developed to detect six different concentrations of MMP-9, each repeated six times. Images were captured using six smartphones under four distinct backgrounds, generating a dataset of 864 images. The segment anything model (SAM) was applied for strip segmentation, achieving 100% accuracy across all samples. ResNet34 and MobileNet_V2 ML model, pre-trained on ImageNet, were used for classification and achieved accuracies of 99% and over 97%, respectively. The results demonstrate that ML-assisted smartphone imaging enables reliable multiclass detection, overcoming the limitations of conventional binary LFA assays. This approach addresses variability across devices and environments, eliminates the need for external readers, and provides a practical, low-cost solution for DED diagnosis. More broadly, the framework can be extended to other LFA-based assays, supporting scalable point-of-care testing (POCT) in resource-limited settings. • The smartphone-assisted paper-based chemiluminescence detection platform offers high integration. • A fully automated “sample in, answer out” operation fashion. • The chemiluminescence detection process has been optimized for enhanced performance. • Rabbit IgG testing achieved a limit of detection of 62.4 pg/mL. • Detection of tau protein reached a limit of detection of 26.1 pg/mL, which is lower than clinical cut-off value.