Advancing Continuous Blood Pressure Estimation with Transformer on Photoplethysmography in Operation Room

Continuous invasive arterial blood pressure (ABP) monitoring can immediately monitor changes in blood pressure in the operating room and intensive care unit and detect hemodynamic disorders such as hypotension for faster response. We propose a transformer model for real-time blood pressure estimation using only noninvasive photoplethysmography (PPG) without specific feature engineering. We collected vital signs during the operation, including invasive arterial blood pressure (ABP) and PPG at the Soonchunhyang University Bucheon Hospital. The transformer model represents vital signs as high-dimensional vectors and includes context vectors using self-attention to understand them in the context of neighbouring values. We develop a unified model that estimates continuous blood pressure, offering insights for detecting abnormal signals and enabling early intervention, particularly in surgical scenarios under anaesthesia where hypotension symptoms may appear. Our model outperforms previous approaches for the intensive care unit data and the operation room data. The mean absolute error (MAE) ± standard deviation (SD) was 1.20 ± 1.04 mmHg for arterial blood pressure (ABP), 1.04 ± 0.91 mmHg for systolic blood pressure (SBP), and 0.89 ± 0.68 mmHg for diastolic blood pressure (DBP) on the operation room data. The performance of the proposed model based on MIMIC II data is 1.54 ± 1.31 mmHg for arterial blood pressure (ABP), 1.24 ± 1.14 mmHg for systolic blood pressure (SBP), and 1.59 ± 0.82 mmHg for diastolic blood pressure (DBP). In this study, we succeeded in estimating real-time ABP using a transformer-based model, emphasizing its ability to handle vital signs data from the operating room, its use of token embedding and self-attention, and the unified modelling of continuous blood pressure for improved detection and intervention in medical scenarios.