Optical Fiber Sensor For Enhanced Biomarker Detection

Abstract

The early identification of Diabetic Retinopathy (DR) is crucial for the prevention of long-term visual loss. However, currently available techniques are both expensive and reliant on specialists which limits their use in developing countries that lack access to these resources. The tear fluid has been demonstrated to contain significantly higher levels of the principal molecular effector of DR - VEGF than healthy individuals. Therefore, it represents a potential non-invasive biomarker for the point-of-care diagnosis of DR. An interdisciplinary device integrating an array of biofunctionalised SDI optical fibre biosensors capable of detecting multiple levels of VEGF concentration simultaneously using a machine learning algorithm was developed to classify VEGF concentration ranges in spectral data. Artificial tear solution containing varying levels of VEGF over a clinically relevant range was used to validate the experimental response of the biofunctionalised SDI sensor. Responses obtained were dependent upon VEGF concentration. However, there existed considerable overlap within intermediate concentrations and thus limited the ability of classifiers to accurately classify samples. Classification accuracy was found to be heavily influenced by feature extraction and reference calibration. A deviation based on references to a blank sample (zero-concentration), substantially enhanced the accuracy of classifications, resulting in a classification accuracy rate of 97.02%, when classifying in a clinically relevant three-class scheme. Hyperparameter optimization demonstrated XGBoost with Moving Average filter to provide the best results.