Blood Pressure Estimation for Hemodynamic Monitoring Using Reduced Photoplethysmography Features in Intensive Care Units: A Prospective Machine Learning Study
ABSTRACT
Background:
Continuous, non-invasive blood pressure (BP) monitoring is clinically important for managing hemodynamic instability in the ICU. However, current ICU practice relies primarily on invasive arterial catheterization. While this is the clinical gold standard for continuous measurement, it carries risks such as infection and thrombosis and is typically limited to select critical care settings.
Objective:
To reduce reliance on invasive arterial lines, we developed an interpretable estimation framework for systolic (SBP), mean arterial (MAP), and diastolic blood pressure (DBP) from single-sensor photoplethysmography (PPG).
Methods:
Methods:
Synchronized PPG and arterial line 10-minute recordings from 100 ICU patients were analyzed. Using a hybrid feature selection strategy (RReliefF and intrinsic model-specific rankings), we reduced 201 initial features to an optimized subset of 30 core predictors. Model development and evaluation were designed to minimize information leakage and overfitting, incorporating nested cross-validation and independent patient-wise testing to assess generalizability across heterogeneous ICU populations. We developed and validated predictive models using machine learning techniques, specifically XGBoost, Random Forest, and Support Vector Regression (SVR), with XGBoost serving as our primary model.
Results:
On the test cohort, the optimized model achieved a mean absolute error (MAE) of 3.50 mmHg for MAP and 2.62 mmHg for DBP, consistently meeting AAMI/BHS performance standards (MAP: mean differences (MD) ± SD = -0.09 ± 4.54 mmHg; DBP: MD ± SD = -1.07 ± 3.51 mmHg), confirming clinical-grade accuracy. While SBP estimation slightly exceeded the SD threshold due to high-frequency fluctuations and intra-subject variability, the mean bias for all targets remained within the 5-mmHg limit. Under BHS standards, the model achieved Grade A for both MAP and DBP. Although SBP performance was more sensitive to feature reduction, the 30-feature subset provided the optimal balance for ICU monitoring, prioritizing the precision of MAP and DBP essential for detecting hypotension and maintaining organ perfusion. This 85% reduction in dimensionality maintained predictive accuracy while enabling a per-patient inference latency of 20–23 ms.
Conclusions:
Interpretability analysis identified key predictors: pulse reflection, systolic timing, vasomotor oscillations, and autonomic regulation. By explicitly linking predictive features to underlying cardiovascular physiology, the proposed framework addresses a key limitation of existing black-box cuffless BP models. These results demonstrate that accurate, lightweight, and interpretable cuffless BP estimation is feasible using single-sensor PPG, potentially supporting real-time bedside deployment, continuous BP trending, and early detection of hypotension in critical care settings.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.