Integration of a photoplethysmogram sensor to a single-board computer with cBP-Tnet edge AI for continuous blood pressure estimation

Deep learning (DL) methods for continuous blood pressure (BP) monitoring utilizing photoplethysmogram (PPG) signals have been confined only to research modeling and not applied. Recently, research on the weighing-scale-based system proved potential DL practical deployment (Edge AI) as existing systems rely on the Pulse Transit Time (PTT) method. The PTT method depends only on manual feature extraction, limiting its performance, and requires inconveniently one additional sensor per method (electrocardiogram, ballistocardiogram, and impedance plethysmography). This study deployed a DL model, cBP-Tnet Edge AI, into a single clip-on BP measuring system mounted on the sensor bar of a body-fat weighing scale, aiming for practical use case. This setup ensures familiar and secure skin contact minimizing motion artifacts. Specifically, we evaluated the performance of integrating a PPG sensor (MAX30102) to single-board computer (Raspberry Pi 5), deployed with cBP-Tnet edge AI for continuous BP estimation. To validate, we recruited 10 volunteer participants and compared cBP-Tnet Edge AI’s performance to established PTTbased and traditional cuff oscillometric BP measurement methods. In comparison to PTT-based, cBP-Tnet Edge AI attained Pearson correlations of $\mathrm{r}=0.788$ for systolic blood pressure (SBP) and $\mathbf{r} \boldsymbol{=} \mathbf{0. 8 4 1}$ for diastolic blood pressure (DBP). When compared to traditional cuff-based BP, measurements were preliminarily feasible (SBP: $\mathrm{p}=0.063$; DBP: $\mathrm{p}=0.791$). The mean absolute errors were 9.7 mmHg for SBP and 6.3 mmHg for DBP, with overall accuracies of $91.7 \%$ and $91.2 \%$, respectively. Thus, cBP-Tnet Edge AI is comparable in performance to both established PTT-based and traditional cuff-based methods, providing potential to include SBP and DBP measurements to commercial body-fat scales.