Lateral flow assays (LFAs) are extensively utilized in point-of-care diagnostics due to their affordability, simplicity, and rapid time-to-results. However, their low sensitivity remains a significant limitation, particularly for detecting foodborne pathogens at concentrations below regulatory thresholds. This study evaluated two distinct sensing modalities—photothermal speckle imaging and colorimetric line intensity analysis—for their potential to enhance the sensitivity of commercially available LFAs. Photothermal imaging quantified refractive index shifts induced by plasmonic heating of gold nanoparticles, while colorimetric analysis used smartphone-acquired images processed with machine learning. The photothermal method achieved a limit of detection (LOD) of 2.13 × 10⁵ CFU/mL, while the colorimetric approach, using a logistic regression model with LASSO regularization, achieved an LOD of 10⁵ CFU/mL. While both approaches demonstrated detection thresholds comparable to traditional visual interpretation, the colorimetric method provided an added advantage by enabling quantitative prediction of bacterial concentration through regression modeling. With further optimization of each sensing method, these findings demonstrate the feasibility of improving unmodified commercial LFAs through optical and computational enhancements, offering a promising pathway toward the development of portable biosensing systems for real-time food safety monitoring.

Foodborne pathogens remain a major global concern, demanding rapid, accessible, and determination technologies. Conventional methods, such as culture assays and polymerase chain reaction, offer high accuracy but are time-consuming for on-site testing. This study presents a portable, smartphone-assisted dual-mode biosensor that combines colorimetric and photothermal speckle imaging for improved sensitivity in lateral flow assays (LFAs). The prototype device, built using low-cost components ($500), uses a Raspberry Pi for illumination control, image acquisition, and machine learning-based signal analysis. Colorimetric features were derived from normalized RGB intensities, while photothermal responses were obtained from speckle fluctuation metrics during periodic plasmonic heating. Multivariate linear regression, with and without LASSO regularization, was used to predict Salmonella concentrations. The comparison revealed that regularization did not significantly improve predictive accuracy indicating that the unregularized linear model is sufficient and that the extracted features are robust without complex penalization. The fused model achieved the best performance (R² = 0.91) and consistently predicted concentrations down to a limit of detection (LOD) of 10⁴ CFU/mL, which is one order of magnitude improvement of visual and benchtop measurements from previous work. Blind testing confirmed robustness but also revealed difficulty distinguishing between negative and 10³ CFU/mL samples. This work demonstrates a low-cost, field-deployable biosensing platform capable of quantitative pathogen detection, establishing a foundation for the future deployment of smartphone-assisted, machine learning-enabled diagnostic tools for broader monitoring applications.