Poster R18 - Deep learning based continuous blood pressure prediction via forehead NIRS-PPG (FlexNIRS)

Introduction: Existing light based noninvasive and continuous blood pressure monitoring devices, such as the ones based on finger pulse-ox, provide insufficient accuracy. To address this challenge, we investigated the feasibility of forehead NIRS-PPG, allowing for a more robust measurement of pulsatile signal than in a finger, coupled with artificial intelligence to predict arterial blood pressure (ABP).

Materials and

Methods: We performed measurements in patients undergoing carotid endarterectomy (CEA) surgery, which entails large blood pressure changes and requires ABP monitoring with an arterial line. We employed a wearable NIRS device called FlexNIRS [1], sampled at 266 Hz, positioned on the forehead, contralateral to the surgery site. We implemented a deep learning model that combines convolutional neural networks (CNN) for morphological feature extraction and long short-term memory (LSTM) networks for temporal feature extraction, aiming to accurately predict ABP. This model was largely adapted and modified from previously published architectures [2-4]. Specifically, the model consists of one bi-layer LSTM and two LSTMs with 56 hidden units each, followed by two CNNs featuring 100-200 kernels ranging in size from 5 to 20, in combination with rectified linear unit activation layers, pooling layers, and fully connected networks for dimension reduction (Figure 1). The raw NIRS-PPG signal at 850 nm and 2.8 cm source-detector separation was resampled to 250 Hz and bandpass-filtered at 0.55 – 25 Hz. Four consecutive beats were ensemble averaged, and its derivative (d/dt(NIRS-PPG)) was paired, zero-padded, and filled into a bin of 300 samples (2 x 300 x #beats; except for one subject needed to use a bin of 350 samples to accommodate an unusually low heart rate) and matched with mean arterial blood pressure (MAP) values (1 x #beats) for training the model.

Results, Conclusions, and Discussions: In six CEA patients (one female), with an average of 2.7 hours of valid data, we trained the CNN-LSTM model using a set of periods no more than 10-20 minutes among when the MAP exhibited large fluctuations, which occurred around clamping and unclamping. MAP was predicted for the remaining untrained data, revealing a strong correlation between the predicted and actual MAP values (R=0.96, 0.80, 0.87, 0.95, 0.91, and 0.86, respectively; Figure 2). Our findings suggest that deep learning models, utilizing time-series NIRS-PPG data directly without prior feature extraction and dimension reduction, hold promising potential for robust, accurate, noninvasive, continuous ABP monitoring. In five additional CEA subjects, we are optimizing our model, including signal pre-processing, smoothing, parallel training seeds, and training data selection strategies, to improve prediction accuracy and robustness.

Acknowledgements (Optional): [1] Wu, K.C., Martin, A., Renna, M., Robinson, M., Ozana, N., Carp, S.A., Franceschini, M.A., 2023. Enhancing diffuse correlation spectroscopy pulsatile cerebral blood flow signal with near-infrared spectroscopy photoplethysmography. Neurophotonics 10, 1–15. https://doi.org/10.1117/1.nph.10.3.035008
[2] Azar, J., Makhoul, A., Couturier, R., Demerjian, J., 2021. Deep recurrent neural network-based autoencoder for photoplethysmogram artifacts filtering. Comput. Electr. Eng. 92, 107065. https://doi.org/10.1016/j.compeleceng.2021.107065
[3] Xia, K., Huang, J., Wang, H., 2020. LSTM-CNN Architecture for Human Activity Recognition. IEEE Access 8, 56855–56866. https://doi.org/10.1109/ACCESS.2020.2982225
[4] Jeong, D.U., Lim, K.M., 2021. Combined deep CNN–LSTM network-based multitasking learning architecture for noninvasive continuous blood pressure estimation using difference in ECG-PPG features. Sci. Rep. 11, 1–7. https://doi.org/10.1038/s41598-021-92997-0