Implementing a Personalized Model in Edge via FPGA for Non-Invasive Blood Flow Volume Measurement Based on PPG for Security

This study proposes a personalized blood flow volume (BFV) prediction method using PPG signals for improved accuracy and model robustness in dialysis treatment. Experimental results validate the effectiveness of the proposed system, demonstrating enhanced raw PPG signal processing through band-pass filtering, power spectrum density (PSD) calculation, and DC drift analysis for PPG quality assessment. Extracted features from pre-processed PPG signals are utilized as inputs for personalized models trained using transfer learning. The dedicated hardware implementation accelerates both training and BFV calculation, significantly reducing latency compared to software. The hardware-based implementation exhibits notably faster latency, reducing training latency from 469.45ms to 115.88ms and testing latency from 1.2ms to 0.0041ms. Despite a slight decrease in R-square (R2) from 0.985 to 0.984, the hardware implementation successfully achieves improved latency, highlighting its practical efficiency. This approach offers a reliable and efficient method for personalized BFV estimation using PPG signals, with potential benefits in dialysis treatment. The FPGA-based hardware implementation utilizing the CDC method enables efficient training of personalized models, achieving comparable calculation speeds while reducing latency. Overall, this approach demonstrates the potential for enhanced patient care in dialysis treatment through reliable and efficient personalized BFV estimation using PPG signals, featuring an improved R2 of 0.985, reduced latency, and efficient hardware implementation.