What question did this study set out to answer?

This study aims to evaluate a deep learning algorithm's performance in estimating blood pressure from photoplethysmographic signals using invasive arterial measurements as a reference.

June 3, 2026

Performance Evaluation of a Deep Learning Algorithm for Blood Pressure Estimation From Photoplethysmographic Signals

Key Points

This study aims to evaluate a deep learning algorithm's performance in estimating blood pressure from photoplethysmographic signals using invasive arterial measurements as a reference.
Utilized a two-stage cascaded neural network architecture trained on synchronized PPG and ABP waveforms.
Data sourced from the pre-processed MIMIC-III database of ICU patients.
Performance evaluated through correlation analysis and absolute error distributions.
SBP demonstrated a strong correlation with ABP (r = 0.936, P < 0.001) and absolute errors within ±5 mmHg for 68% of samples.
DBP also showed strong association (r = 0.893, P < 0.001), with 83% of samples having errors ≤5 mmHg.
Bland-Altman analysis revealed mean differences of 3.0 ± 10.5 mmHg for SBP and -1.9 ± 6.8 mmHg for DBP.

Abstract

Objective: Continuous blood pressure (BP) monitoring is important in many clinical settings. Invasive arterial blood pressure (ABP) measurement remains the reference standard but is associated with procedural risks and limited applicability. Non-invasive cuff-based techniques provide intermittent measurements, whereas photoplethysmography (PPG) offers a non-invasive approach that may enable continuous BP estimation. This study evaluated the performance of a deep learning–based algorithm for estimating systolic (SBP) and diastolic blood pressure (DBP) from PPG waveforms using invasive ABP as reference. Design and method: A two-stage cascaded neural network architecture developed by Redwave Medical GmbH (Jena, Germany) was trained using synchronized PPG and ABP waveform segments extracted from the pre-processed MIMIC-III database comprising intensive care unit (ICU) patients. Data were split into training and independent test sets. Algorithm performance was evaluated on a held-out test cohort comprising 26,981 samples in total. Association and agreement with invasive ABP measurements were assessed using Pearson correlation analysis, Bland–Altman analysis, and absolute error distributions according to British Hypertension Society (BHS) criteria. Results: Predicted SBP and DBP values showed strong statistical association with invasive reference ABP (SBP: r = 0.936; DBP: r = 0.893; both P < 0.001). For SBP, absolute errors were <=5 mmHg in 68%, <=10 mmHg in 85%, and <=15 mmHg in 91% of samples. Corresponding values for DBP were 83%, 92%, and 96%, respectively. Bland–Altman analysis demonstrated mean differences of 3.0 ± 10.5 mmHg and -1.9 ± 6.8 mmHg for SBP and DBP, respectively, with no relevant proportional bias observed. Conclusions: In this retrospective analysis, a deep learning–based approach enabled estimation of SBP and DBP from PPG waveforms, demonstrating agreement relative to invasive ABP measurements that warrants further evaluation in prospective validation studies. These results support continued investigation of PPG-based BP estimation as a non-invasive adjunct to established continuous monitoring methods. Prospective studies in diverse patient populations, beyond ICU settings, are required to assess clinical applicability and robustness.

Mark Helpful

Bookmark

Relay