Estimating LVEDP from a PPG Signal: Feasibility, Signal Processing, and the Path to a Regulated Device
Left ventricular end-diastolic pressure is one of the most clinically meaningful hemodynamic parameters — and one of the hardest to measure without a catheter. Here is what the engineering looks like when you try to extract it from a photoplethysmography waveform.
Left ventricular end-diastolic pressure (LVEDP) is a direct measure of how hard the left ventricle is working to fill against its own stiffness and the pressures in the pulmonary circulation. Elevated LVEDP — typically above 12 mmHg — is one of the earliest and most sensitive indicators of diastolic dysfunction and heart failure with preserved ejection fraction (HFpEF). The problem is that measuring it accurately has traditionally required a left heart catheterization: an invasive, expensive procedure performed in a cath lab.
The promise of non-invasive LVEDP estimation from a photoplethysmography (PPG) signal is compelling. PPG sensors are cheap, wearable, and already embedded in hundreds of millions of consumer devices. If the signal contains recoverable information about cardiac filling pressures — and there is good reason to believe it does — the clinical implications are significant. So is the engineering challenge.
Why PPG Might Contain LVEDP Information
A PPG waveform reflects pulsatile changes in blood volume at the measurement site, typically the finger or wrist. The shape of that waveform is determined by the interaction between the stroke volume ejected by the left ventricle, the compliance of the arterial tree, and the reflection of pressure waves from peripheral vascular beds.
LVEDP influences this chain in several ways. Elevated filling pressures are associated with increased left ventricular stiffness, which alters the pressure-volume relationship during ejection and changes the contour of the aortic pressure waveform. That change propagates through the arterial tree and leaves a measurable imprint on the peripheral PPG waveform — in the timing and amplitude of the dicrotic notch, in the augmentation index, in the ratio of the late systolic to early systolic peak, and in the shape of the diastolic decay.
None of these features map to LVEDP through a simple, direct physiological equation. The relationship is confounded by arterial stiffness (which changes with age, blood pressure, and medications), heart rate, stroke volume, and the measurement site. This is what makes the problem interesting from an engineering standpoint and difficult from a clinical validation standpoint.
Signal Acquisition and Front-End Design
Before any signal processing can happen, you need a clean PPG signal. This sounds straightforward but is one of the most underestimated problems in wearable biosensor development. Motion artifact is the dominant noise source in ambulatory applications, and its frequency content overlaps substantially with the physiologically meaningful components of the PPG waveform.
A well-designed PPG front end starts with the optical configuration. Green wavelengths (around 530 nm) are commonly used at the wrist because they are absorbed more strongly by hemoglobin and produce a higher signal-to-noise ratio in superficial tissue. Infrared wavelengths penetrate more deeply and are preferred for finger probes. The LED drive current, photodetector transimpedance gain, and ambient light rejection circuit all need to be co-designed with the signal processing in mind.
For LVEDP estimation specifically, you need to preserve the morphological features of the PPG waveform with high fidelity. This means sufficient ADC resolution (16-bit is not excessive), a sampling rate of at least 250 Hz to capture the dicrotic notch accurately, and an analog front end with enough bandwidth to avoid distorting the waveform shape. Many off-the-shelf PPG AFE chips optimized for heart rate monitoring apply heavy filtering that destroys the morphological information you need.
Feature Extraction
The literature on PPG-derived hemodynamic parameters has converged on a set of waveform features that carry the most predictive information. In the time domain, the most useful include the time from the systolic peak to the dicrotic notch (related to the pre-ejection period and vascular tone), the augmentation index (the ratio of the late systolic augmentation to pulse pressure, reflecting wave reflection), and the inflection point area ratio.
Second-derivative analysis of the PPG waveform — the accelerated photoplethysmogram (APG) — amplifies subtle morphological features that are difficult to see in the raw waveform. The relative amplitudes of the APG waves (labeled a through e) have been correlated with arterial stiffness and, by extension, with filling pressures.
Pulse transit time (PTT) — the time delay between the R wave on an ECG and the arrival of the pulse at the PPG sensor — is one of the more physiologically grounded features. PTT is inversely related to pulse wave velocity, which is a direct measure of arterial stiffness. Since arterial stiffness and LVEDP are both elevated in diastolic dysfunction, PTT carries useful information — though it requires a simultaneous ECG channel, which adds hardware complexity.
Model Development and Validation
The relationship between PPG-derived features and LVEDP is multivariate and nonlinear. Classical regression approaches — even with careful feature engineering — tend to plateau in performance because they cannot capture the complex interactions between waveform shape, patient physiology, and hemodynamic state.
Machine learning models, particularly gradient boosting and recurrent neural networks applied to the raw waveform, have shown more promise. But they come with a challenge that is specific to medical device development: they are difficult to validate, explain to regulators, and debug when they fail. A model that achieves good mean absolute error in a retrospective dataset may perform very differently in a prospective deployment, particularly in patient populations not well-represented in the training data.
The ground truth problem is also non-trivial. Invasive LVEDP measurements from cardiac catheterization are the gold standard, but catheterization datasets are small, expensive to collect, and subject to significant selection bias — the patients who undergo cath are not representative of the broader heart failure population you ultimately want to monitor. Building a training dataset that supports a generalizable model is a research program in itself.
The Regulatory Path
A device that estimates LVEDP and presents that estimate to a clinician is a Class II medical device at minimum, and the FDA will scrutinize the clinical validation evidence carefully. The predicate device landscape for non-invasive hemodynamic monitoring is thin, which means a de novo submission may be required rather than a straightforward 510(k).
The key questions FDA will ask are about the limits of the claim, the accuracy in the intended use population, and what happens when the algorithm is wrong. Defining the intended use narrowly — for example, trending rather than absolute measurement, or screening rather than diagnosis — can substantially reduce the evidentiary burden and the risk profile of the device.
IEC 62304 software lifecycle requirements apply to the algorithm itself, which creates documentation obligations around training data, model versioning, performance monitoring, and post-market surveillance that are not typical of conventional firmware development. If the algorithm is updated based on post-market data, that update may trigger a regulatory submission depending on how the original clearance was scoped.
What a Realistic Development Program Looks Like
The path from a research result showing correlation between PPG features and LVEDP to a cleared, manufacturable device is long — typically three to five years for a first-in-class monitoring device. The feasibility phase involves hardware bring-up, signal quality optimization, and a small clinical study to validate that your feature extraction pipeline produces results consistent with the published literature. The pivotal clinical study, designed with FDA input through a pre-submission meeting, is the dominant cost driver and schedule risk.
Getting the hardware and software architecture right early is critical. Choices made in the feasibility phase — ADC resolution, sampling rate, front-end topology, algorithm framework — are expensive to reverse once clinical data collection has begun. Investing in a rigorous design process before the first patient visit pays for itself many times over.
At SiGenix, we have worked on PPG-based wearable devices across several applications and understand both the signal processing challenges and the regulatory requirements for this class of device. If you are developing a non-invasive hemodynamic monitoring product and want to talk through the technical or regulatory strategy, reach out to our team.