Abstract
The Erythrocyte Sedimentation Rate (ESR) is a critical clinical indicator of inflammation, traditionally requiring a one-hour observation period via the Westergren method. This study presents a deep learning approach using Gated Recurrent Units (GRUs) to predict the 60-minute ESR value using only the first 30 minutes of sedimentation data. By analyzing sequential images of blood samples, the GRU model captures the complex temporal dynamics of red blood cell settling. Experimental results on 130 clinical samples demonstrate that the GRU model achieves a Mean Absolute Error (MAE) of 0.0143, significantly outperforming Random Forest and standard feed-forward architectures. This method offers a path toward doubling diagnostic throughput using standard optical equipment.
Introduction
Erythrocyte Sedimentation Rate (ESR) testing is a fundamental diagnostic tool used to detect non-specific inflammation associated with infections, autoimmune diseases, and cancers. When inflammatory proteins are present, red blood cells aggregate into stacks called rouleaux, which settle faster than individual cells. The standard Westergren method, while cost-effective, is time-consuming, requiring 60 minutes for a single reading. Rapid alternatives often necessitate expensive, specialized hardware. This research explores the use of computer vision and recurrent neural networks to halve the testing time by forecasting the final sedimentation height from early-stage data.
Methods and Materials
The experimental setup utilized 1 mL blood samples placed in syringes, with a Nikon D5600 camera capturing images at one-minute intervals. To extract the sedimentation height, the researchers employed Otsu’s method for automated image thresholding, converting photographs into binary images to track the plasma-erythrocyte interface.
The study employed a Gated Recurrent Unit (GRU) architecture, a type of recurrent neural network optimized for sequential data. Unlike static models, GRUs utilize memory gates to process the history of the settling process, which is essential for modeling the non-linear acceleration of erythrocytes. The model was trained and validated on 130 blood samples from Pusan National University Hospital, standardized to a 35% hematocrit level.
Results and Discussion
Analysis of the prediction window revealed a critical trade-off between speed and accuracy. While predictions were attempted at earlier intervals (e.g., 20 minutes), the 30-minute mark was identified as the optimal threshold for maintaining clinical stability and low error rates.
The GRU model demonstrated superior performance compared to Random Forest, Feed-Forward Neural Networks (FFNN), and Wide & Deep Neural Networks (WDNN). Specifically, the GRU achieved a Mean Absolute Error (MAE) of 0.0143 in normalized height. The success of the GRU over the FFNN underscores the importance of temporal context; the model effectively learns the "momentum" of the sedimentation process rather than relying on a single height measurement.
Conclusion
By integrating computer vision with recurrent neural networks, this study successfully demonstrates a method to predict 60-minute ESR values in just 30 minutes. The approach maintains high accuracy while utilizing standard camera hardware, offering a scalable solution for clinical laboratories. Future research will focus on increasing dataset diversity and exploring the implementation of this model on mobile devices for point-of-care diagnostics.
Summary updated