Swelling-Aware Face Recognition for Hemodialysis Patients Using Contrastive Deep Learning
- 4 October 2025
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 445-450
Abstract
Facial recognition technologies offer a promising solution for secure and contactless patient identification in clinical environments, particularly in hemodialysis centers where patients require frequent and repeated visits. However, their reliability is significantly reduced in the presence of facial edema - a common side effect of renal replacement therapy - which alters facial geometry and degrades the accuracy of conventional models trained on healthy faces. This study introduces a clinically relevant and scalable deep learning framework for facial edema classification, forming the foundation of the HEM-ID biometric module for national deployment in Kyrgyzstan. A synthetic dataset was constructed by applying controlled warping and augmentation techniques to a filtered set of 910 neutral facial images, simulating three swelling levels: neutral, mild (swollen1), and moderate (swollen2). Each class was augmented to 1,820 images using affine transformations, color augmentation, and horizontal flips, resulting in a fully balanced dataset of $\mathbf{5, 4 6 0}$ images. We evaluated five deep learning architectures - ResNet50, Efficient-Net-B0, a self-supervised contrastive model (SimCLR), a supervised contrastive model (SupCon), and a custom lightweight CNN named HemEdem. All models were trained using transfer learning, early stopping, and standardized augmentation strategies. Quantitative evaluation shows that SupCon achieved the highest macro F1-score of 0.87 and accuracy of 88%, followed by SimCLR (accuracy 0.82, macro F1 0.80). Baseline CNNs performed lower (ResNet50: accuracy 0.76, macro F1 0.74; EfficientNet-B0: accuracy 0.74, macro F1 0.71), while the lightweight HemEdemCNN reached accuracy 0.68 and macro F1 0.68, reflecting its efficiency-oriented design. The proposed approach demonstrates the feasibility of deploying robust, interpretable, and privacy-preserving facial identification systems in resource-limited clinical settings. HEM-ID offers a scalable solution to ensure secure and reliable patient verification across distributed hemodialysis clinics, even in the presence of dynamic physiological changes.
Keywords
This publication has 8 references indexed in Scilit:
- Deep learning analysis of blood flow sounds to detect arteriovenous fistula stenosisnpj Digital Medicine, 2023
- Deep Learning Approaches for Data Augmentation in Medical Imaging: A ReviewJournal of Imaging, 2023
- Edema Estimation From Facial Images Taken Before and After Dialysis via Contrastive Multi-Patient Pre-TrainingIEEE Journal of Biomedical and Health Informatics, 2022
- End-to-End Object Detection with Transformers2020
- MedGAN: Medical image translation using GANsComputerized Medical Imaging and Graphics, 2020
- A survey on Image Data Augmentation for Deep LearningJournal of Big Data, 2019
- MobileNets: Efficient Convolutional Neural Networks for Mobile Vision ApplicationsarXiv, 2017
- Deep Residual Learning for Image Recognition2016