Blood Pressure Estimation Using Explainable Deep-Learning Models Based on Photoplethysmography.

Area of Science:

• Biomedical Engineering
• Medical Devices
• Artificial Intelligence in Medicine

Background:

• Arterial lines offer superior hemodynamic monitoring but are invasive and not for routine use.
• Continuous, noninvasive arterial pressure monitoring is needed to overcome limitations of intermittent cuff measurements.
• Photoplethysmography (PPG) signals offer a potential noninvasive source for continuous arterial pressure estimation.

Purpose of the Study:

• To develop and validate a deep-learning model for reconstructing continuous mean arterial pressure (MAP) from PPG signals.
• To compare the performance of the deep-learning model against the arterial line gold standard.
• To assess the potential of noninvasive MAP estimation as a decision-support tool in clinical settings.

Main Methods:

• A deep-learning model utilizing convolutional and recurrent layers was trained on high-frequency PPG signals from 117 patients.
• Input features included PPG signals and derived metrics like dicrotic notch amplitude and heart rate.
• Model performance was evaluated using Mean Absolute Error (MAE), with explainability assessed via Grad-CAM saliency maps.

Main Results:

• The deep-learning model achieved a Mean Absolute Error (MAE) of 3.5 (± 4.4) mm Hg, a 42.6% improvement over a baseline cuff-based model (MAE 6.1 ± 14.5 mm Hg).
• The model met international standards for accuracy (grade A) and demonstrated narrow confidence intervals.
• Grad-CAM analysis indicated the model primarily utilizes information from the dicrotic notch region of the PPG waveform.

Conclusions:

• The developed deep-learning model accurately estimates arterial pressure noninvasively using PPG signals.
• This technology holds potential as a valuable decision-support tool in operating rooms, especially when invasive monitoring is not feasible.
• Further integration into clinical workflows could enhance patient monitoring and safety.