Development of colorimetric and machine learning based accurate glucose detection platform for point of care applications.

Improving overall health and preventing complications is crucial for timely and effective treatment of diabetes patients. In this direction, accurate measurement and detection of glucose concentration in blood is essential. The aim of this study is to develop an affordable, reliable and accurate Point-of-Care (POC) diagnostic platform for glucose concentration detection using microfluidic and colorimetric principles. A microfluidic chip is fabricated which provides the base for efficient colorimetric reaction. The proposed system requires only ~ 20 µL of sample per microwell, with a colorimetric reaction time of 3-4 min. The chip is positioned inside a compact, USB-powered, 3D printed image capture device affixed with a high-resolution camera so that variables such as camera position, focal distance and lighting conditions can be controlled. The analysis workflow is adaptable for integration with embedded systems or laptops, making it suitable for real-time deployment in Point-of-Care settings without the need for smartphones or calibration tools. The images captured inside the device (1280 in total, corresponding to 16 glucose concentration levels ranging from 50 to 200 mg/dL) were labeled based on known concentration levels and subjected to standardized preprocessing. Key preprocessing steps included Region of Interest (ROI) extraction using a fixed box detection algorithm, normalization of pixel values, image resizing to 128 × 128 pixels, and standardization using ImageNet parameters. These ensured uniformity and robustness prior to feature extraction and machine learning classification. The classifiers used are Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Neural Networks (NN), and K-Nearest Neighbours (KNN). The RF machine learning classifier achieved the highest cross-validation accuracy of 87.47% and precision of 89.14%, demonstrating its ability to effectively distinguish between different glucose concentration levels. The confusion matrix and ROC curve analysis further validated the model's robustness, with minimal misclassifications and a high mean AUC value of 0.9509. The results thus highlight the potential of image-based glucose concentration estimation as a cost-effective and scalable solution for real-time monitoring in biomedical applications.