Abstract
Arterial blood pressure (ABP) and photoplethysmography (PPG) waveforms contain vital physiological information for cardiovascular disease prevention and treatment. However, the lack of standardized tools for feature extraction limits their clinical utilization. This study proposes an automatic feature extraction tool that detects temporal landmarks—systolic phase onset, peak, dicrotic notch, and diastolic peak—using the iterative envelope mean (IEM) method. The tool extracts 852 features per cardiac cycle across time, statistical, and frequency domains. Evaluation on the MLORD dataset (17,327 patients) and real-time monitor data showed high performance, with F1-scores exceeding 97% for all landmarks. This tool provides a robust foundation for feature-based machine learning models in various clinical applications.
Introduction
Arterial blood pressure (ABP) and photoplethysmography (PPG) are ubiquitous in clinical settings, providing continuous monitoring of cardiovascular health. While these waveforms are rich in data—capable of predicting hypotension or estimating cardiac output—the absence of standardized, open-source extraction tools has hindered large-scale research and cross-study comparisons. A primary challenge lies in "landmark detection": identifying specific physiological events within a cardiac cycle, such as the dicrotic notch, which signifies the closure of the aortic valve.
Methodology
The proposed tool utilizes a two-stage pipeline to ensure robust feature extraction.
- Landmark Detection via IEM: Traditional derivative-based methods are often sensitive to noise and artifacts. To overcome this, the tool employs the Iterative Envelope Mean (IEM) method. This technique decomposes the signal into stationary and non-stationary components, making landmarks like the dicrotic notch significantly easier to identify, even when obscured in raw signals.
- Comprehensive Feature Set: Following landmark detection, the tool calculates 852 features per cycle. These include time-domain metrics (amplitudes, durations), statistical properties (skewness, kurtosis), frequency-domain features (via FFT), and derivative-based characteristics.
Landmark Identification
The tool focuses on four key temporal landmarks: the systolic phase onset, systolic phase peak, dicrotic notch, and diastolic phase peak. By standardizing the detection of these points, researchers can reliably calculate the durations and amplitudes that define a waveform's morphology.
Experimental Validation and Results
The tool's efficacy was tested against two major data sources:
- MLORD Dataset: A perioperative dataset from UCLA involving 17,327 patients. Over 68,000 cardiac cycles were analyzed.
- Real-time Data: A retrospective analysis of 3,000 cycles collected from a Philips IntelliVue MX800 patient monitor to simulate clinical viability.
The tool demonstrated exceptional accuracy compared to manual expert markings. In the MLORD dataset, F1-scores for systolic onset and peaks exceeded 99.5%, while the more challenging dicrotic notch and diastolic peaks maintained scores above 98.2%. In real-time simulations, performance remained high, with all landmark F1-scores exceeding 97.7% and error rates consistently below 4%.
Conclusion and Future Work
By providing a standardized, high-performance pipeline for ABP and PPG feature extraction, this tool enables the development of more nuanced machine learning models. Future work will focus on enhancing noise sensitivity and exploring the clinical significance of the high-dimensional feature set in predicting surgical complications and other diagnostic tasks.
Summary updated