Automated Digit Recognition and Measurement-Type Classification from Blood Pressure Monitor Images Using Deep Learning

Blood pressure is a vital indicator of cardiovascular health and plays a crucial role in the early detection and management of heart-related diseases. However, current practices for recording blood pressure readings are still largely manual, leading to inefficiencies and data inconsistencies. To address this challenge, we propose a deep learning-based method for automated digit recognition and measurement-type classification (systolic, diastolic, and pulse) from images of blood pressure monitors. A total of 2147 images were collected and expanded to 3649 images using data augmentation techniques. We developed and trained three YOLOv8 variants (small, medium, and large). Post-training quantization (PTQ) was employed to optimize the models for edge deployment in a mobile health (mHealth) application. The quantized INT8 YOLOv8-small (YOLOv8s) model emerged as the optimal model based on the trade-off between accuracy, inference time, and model size. The proposed model outperformed existing approaches, including the RT-DETR (Real-Time DEtection TRansformer) model, achieving 99.28% accuracy, 96.48% F1-score, 641.40 ms inference time, and a compact model size of 11 MB. The model was successfully integrated into the mHealth application, enabling accurate, fast, and automated blood pressure tracking. This end-to-end solution provides a scalable and practical approach for enhancing blood pressure monitoring via an accessible digital platform.