A Semi-Decentralized and Parallel Bidirectional Deep Learning (SDP-BiDL) Framework for Binary Classification of Diabetic Retinopathy

Krishna Veer Singh

Keywords: Diabetic Retinopathy, Deep Learning, Semi-Decentralized Framework, Bidirectional Neural Networks, Multimodal Healthcare Analytics, Class Imbalance Handling, PCA Feature Extraction, SMOTE Oversampling

TL;DR: We propose a semi-decentralized bidirectional deep learning framework (SDP-BiDL) that integrates multimodal patient data to achieve state-of-the-art accuracy and real-time inference for diabetic retinopathy prediction.

Abstract: Diabetic Retinopathy (DR) remains a critical healthcare challenge, where timely and accurate assessment is essential to prevent vision loss and associated cardiovascular complications. Existing machine learning models often suffer from limited robustness, interpretability, and scalability in clinical environments. To address these issues, we introduce a semi-decentralized and parallel Bidirectional Deep Learning (SDP-BiDL) framework for joint DR classification and cardiovascular disease risk prediction. Our approach leverages a dataset of 1,151 patients with clinically validated DR, incorporating nineteen covariates derived from retinal images, including lesion features, anatomical descriptors, and global image statistics. To enhance predictive reliability, we employ statistical feature selection and address class imbalance using oversampling techniques. The proposed SDP-BiDL architecture integrates parallel bidirectional modules—based on BiGRU / BiLSTM / BiRNN—trained on complementary feature subsets, enabling independent temporal modeling and effective fusion of representations for final prediction. In extensive experiments, the proposed SDP-BiLSTM, SDP-BiGRU, and SDP-BiRNN variant achieves 98.94\%, 98.72\%, and 97.34\%, respectively accuracy with an area under the curve of 0.99, 0.98, 0.97, respectively, surpassing both traditional machine learning and conventional deep neural architectures. Moreover, the framework supports real-time inference, producing predictions in under one second, which is crucial for clinical applicability. These results demonstrate that SDP-BiDL offers a scalable and interpretable solution for multimodal healthcare analytics, effectively combining imaging phenotypes, biomarkers, and medication records, and holds strong promise for deployment in real-world clinical decision support systems.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 3800