Monitoring tear biomarkers is crucial for assessing both ocular and systemic health. Abnormal levels of vitamin C, pH, Ca2+, and proteins in tears are associated with various diseases, including corneal injury, dry eye, aniridia, and rosacea. Wearable biosensors offer a promising approach for non-invasive, real-time health monitoring, but current wearable tear sensors, mostly electrochemical, are often expensive, complex, and prone to unreliability. Microfluidic colorimetric sensors provide a more user-friendly alternative, visualizing biomarker concentrations through colorimetric reactions. However, challenges remain in data acquisition, interpretation, and correction for variations in pH and ambient light. This study addresses these limitations by developing an AI-assisted system that combines a flexible microfluidic colorimetric sensor with a sophisticated deep-learning algorithm for accurate and efficient biomarker detection.

Existing wearable tear sensors primarily rely on electrochemical methods, which involve integrating electrodes, power sources, and wireless transmission units, leading to high fabrication costs and reduced reliability. Microfluidic colorimetric sensors offer advantages in simplicity and user-friendliness, using smartphones for data acquisition and processing. Several studies have demonstrated the effectiveness of colorimetric sensors for sweat biomarker detection. However, existing colorimetric tear sensors often have complex data readout processes or pose infection risks (e.g., contact lens sensors). Furthermore, variations in tear pH and ambient light significantly affect the colorimetric reactions, compromising data accuracy. The integration of machine learning algorithms has shown promise in improving the performance of colorimetric sensors, but a comprehensive system addressing both data processing and environmental variations in tear analysis remains lacking.

The AI-WMCS system consists of a flexible polydimethylsiloxane (PDMS) microfluidic patch and a cloud server data analysis system (CSDAS) on a smartphone. The patch collects tears and facilitates colorimetric reactions with chromogenic reagents specific to vitamin C, pH, Ca2+, and proteins. A smartphone camera captures images of the color changes, and the CSDAS processes these images using a well-trained multichannel CNN-GRU neural network. The CNN-GRU model corrects for pH and color temperature variations, improving the accuracy of biomarker concentration predictions. The fabrication of the PDMS microfluidic patch involves a 'cut-and-paste' method using a copper mold and UV lithography. Chromogenic reagents are applied to filter paper chips and incorporated into the microfluidic reservoirs. Artificial tear samples with varying biomarker concentrations and pH levels were used for model training and testing. The performance of the AI-WMCS system was evaluated by comparing its predictions with results from standard methods.

The study achieved high accuracy in predicting biomarker concentrations using the AI-WMCS system. The 3D-CNN-GRU model demonstrated excellent performance in predicting Ca2+, vitamin C, and protein concentrations (R² = 0.994), while the 1D-CNN-GRU model accurately predicted pH (R² = 0.998). The deep-learning algorithm effectively corrected for errors caused by pH and color temperature variations, significantly improving the accuracy of the predicted concentration data. Spiked recovery tests using artificial tear samples showed that the AI-WMCS system produced results within 95.03–113.13% of the true values, highlighting the accuracy of the pH and color temperature correction. The system also demonstrated good performance in analyzing real tear samples from healthy volunteers, with all biomarker concentrations falling within their physiological ranges.

The AI-WMCS system successfully addresses the limitations of existing wearable tear sensors by combining a user-friendly microfluidic colorimetric device with a powerful deep-learning algorithm. The ability to correct for pH and color temperature variations is a significant advancement, ensuring accurate and reliable biomarker detection under varying conditions. The high accuracy and non-invasive nature of the system make it suitable for continuous health monitoring and early disease detection. The findings demonstrate the potential of this integrated approach to revolutionize point-of-care diagnostics and personalized healthcare.

This study presents a novel AI-WMCS system for the simultaneous and accurate detection of multiple tear biomarkers. The integration of a flexible microfluidic colorimetric sensor with a deep-learning-based data analysis system significantly improves accuracy and user-friendliness. This system holds great promise for non-invasive health monitoring, advancing the fields of telehealth and personalized precision medicine. Future work could focus on expanding the panel of detectable biomarkers, miniaturizing the device further, and integrating it with other wearable technologies.

The current study uses artificial tear samples and a limited number of real tear samples. Further validation with larger cohorts and diverse populations is needed to confirm the generalizability of the findings. Long-term stability and durability of the sensor patch under real-world conditions require further investigation. The reliance on a smartphone and cloud server might limit accessibility in regions with limited infrastructure.