Abstract
Blood clot formation and capillary fiber blockage in dialyzers remain critical challenges for patients with end-stage kidney disease (ESKD) undergoing hemodialysis. This study aimed to develop a machine learning model that effectively quantifies residual blood clots in dialyzer images captured using bedside smartphone cameras.
Dialyzer images were collected using mobile phones, and preprocessing techniques-such as background noise removal and image segmentation-were applied to focus on relevant regions. Data augmentation was used to increase model robustness. Composite images were created by combining views from both ends of the dialyzer, enhancing the model's ability to detect residual clots. We developed a binary classification model to distinguish between <10% and ~30% blood clot levels using a pre-trained ConvNeXt architecture. Explainable AI (LIME) was incorporated to ensure the model focused on clinically relevant areas in its predictions.
The dataset was split into training (60%), validation (20%), and testing (20%) sets, with 10 random trials for robustness. The ConvNeXt model achieved an accuracy of 0.6971 without pre-training or data augmentation, which increased to 0.7572 with pre-trained weights. Our combined framework yielded the highest accuracy (0.7672) and reduced standard deviation, indicating greater robustness. For comparison, two nephrology nurses achieved accuracies of 0.6271 and 0.6005 when manually classifying clot levels based solely on end images.
Our approach effectively detects residual blood clots in dialyzers using ConvNeXt by leveraging image data from both ends. The use of explainable AI tools confirmed the model's ability to accurately identify blood clots by focusing on relevant regions. Our study emphasizes the need to balance model complexity with computational efficiency. The ConvNeXt base model successfully avoided overfitting while maintaining practical performance, which could lead to improved clinical decision-making by minimizing circuit downtime and optimizing anemia management.