Continuous non-invasive pulmonary artery pressure assessment via seismocardiography and twin neural network
E Korsgaard, A Agam, P Sogaard, B Kjaergaard, J J Struijk, N Giordano, S SchmidtAbstract
Background
Small and sustained increase in pulmonary artery pressure (PAP) often precede acute heart failure (HF) and are only later reflected in conventional vital signs and symptoms. Elevated PAP is currently considered as a pivotable biomarker to monitor HF decompensation as continuous invasive PAP monitoring has been shown to reduce hospitalizations and mortality. Current continuous PAP monitoring relies on invasive and costly catheter-based systems limiting the use in general HF. Seismocardiography (SCG), a non-invasive technique capturing cardiac vibrations, has shown potential for tracking cardiac hemodynamics and may allow continuous, low-cost HF monitoring to guide timely intervention and improve outcomes.
Purpose
To develop a non-invasive method using SCG and twin neural networks (TNN) to track changes in PAP over time using only the SCG signal.
Methods
Ten anesthetized pigs were subjected to alternating hypercapnia and hypoxemia. The pig returned to baseline values before a new hypercapnia or hypoxemia intervention was initiated. PAP was measured via right heart catheterization, while SCG was recorded from a sternal accelerometer. A TNN was trained to estimate relative changes in PAP by comparing SCG recordings at baseline with those recorded during the intervention. Model performance was evaluated based on correlation between the true and predicted difference between baseline and intervention PAPs using leave-one-out-strategy.
Results
The TNN predicted PAP changes with a mean correlation of 0.688 during hypercapnia and 0.432 during hypoxemia across the 10 pigs. The correlation was statistically significant different from 0 (p=0.0039) for hypercapnia and close to significant for hypoxemia (p=0.0580).
Conclusion(s)
SCG was able to non-invasively detect changes in PAP in pigs, with moderate to strong correlation to invasive measurements, particularly during hypercapnia. These findings suggest that using SCG alone can track dynamic changes in PAP. This method could potentially be utilized as a continuous, low-cost alternative to catheter-based PAP monitoring to monitor hemodynamic changes in HF patients. However, further evaluation in humans is required.Examples of model predictions over timeCorrelation plots for all models