Abstract:
"This research addresses the challenge of detecting diabetes accurately, given its significant health and economic impacts. Leveraging advancements in machine learning, particularly in univariate models, the study aims to enhance the accuracy of diabetes diagnosis. The limitations of existing univariate models, such as their inability to effectively generalize novel data and extract useful features, underscore the need for improvement.
To overcome these limitations, the research proposes a novel approach utilizing both patient symptoms and retinal images for diagnosis. The technical solution involves implementing a random forest algorithm for symptom prediction and a pre-trained EfficientNetB0 architecture for retinal image analysis. The retinal model architecture incorporates global average pooling, dense layers, and sigmoid function for binary categorization, while the symptoms prediction model employs the random forest algorithm for classification.
Through thorough testing, including model-specific testing and multimodal testing, the proposed approach demonstrates outstanding results. The retinal model attains a 97% accuracy rate, showcasing high precision, recall, and F1 scores across both categories. Similarly, the symptoms prediction model, applying the random forest algorithm, achieves a commendable accuracy of 97%, with balanced precision, recall, and F1 scores. The multimodal approach, using late fusion techniques to combine the outcomes of the two models, achieves an even higher accuracy of 98%, further validating its effectiveness."
