Pulse Wave Analysis to Estimate Cardiac Output: Comment

We read with great interest the recent review in Anesthesiology, “Pulse Wave Analysis to Estimate Cardiac Output.”¹  As military anesthesiologists, we are very excited about the future potential to allow advanced waveform analysis to help better guide treatment when noninvasive monitors are not available or impractical for the clinical situation. Our group has investigated the use of pulse wave analysis for military applications, particularly for traumatic hemorrhage. There are several nuances we think are important to discuss that have thus far limited the clinical use of these devices.

Technologies for imputing hemodynamic parameters from pulse wave analysis primarily use multivariate linear and polynomial²  regression to identify associations between pulse wave features and parameters such as blood pressure, stroke volume, cardiac output (CO), systemic vascular resistance (SVR), and preload. For example, while there have been several updates to the Vigileo software (Edwards Lifesciences Corp., USA), the original algorithm was a simple linear regression based on the formula CO = κ (α × PWTT × β) HR, where κ, α, and β are constants unique to each patient and determined by pulse pressure, pulse wave transit time (PWTT), CO, and heart rate (HR).³  These systems have enhanced clinical care in a variety of settings and physiologic states, and have a high degree of accuracy.

However, pulse wave analysis models often rely on assumptions to relate the models to pertinent hemodynamic variables. An example of these assumptions is the parameters used in the Windkessel model used by VolumeView and ClearSight (Edwards Lifesciences Corp.). As the authors point out, there are ranges of invasiveness and calibration techniques employed to approximate model parameters, but the inaccuracy of several of these technologies in extreme physiologic states implies that this approach has critical limitations. Although polynomial regression allows modeling nonlinearities in data (unlike simple linear regression), it can be very sensitive to outliers and produce highly inaccurate predictions at extremes of the data range. In fact, this is discussed in the Edwards Lifesciences patent application for the technology used in VolumeView: “The accuracy of this method may be low in some extreme clinical situations where the basic empirical relationships of the model are not valid.” This is a particularly important caution for perioperative care when relevant physiology is often extreme and clinicians are in need of reliable hemodynamic information to make critical decisions.

Advances in computer technology and electronic medical records have promoted computer-based learning algorithms to analyze data.⁴  This involves several techniques that fall under the category of machine learning, including the use of neural networks, more commonly referred to as artificial intelligence. These methods enable computers to help identify trends in data and physiologic processes that may be too nuanced for traditional data analysis techniques such as traditional regression techniques to identify.

A recent study compared the accuracy of stroke volume prediction during liver transplant in 34 patients using the Edwards Lifesciences EV1000 (a system that uses pulse wave analysis) compared to a deep convolutional neural network.⁵  The deep learning model outperformed the pulse wave analysis approach in all phases except the anhepatic phase (where comparative performance was equivocal), but most notably had drastically improved performance during reperfusion, a time of extreme hemodynamic stress during which it is arguably most important to have reliable information to care for the patient. The authors discuss previous studies that demonstrated limitations in modeling rapid cardiovascular changes, and in particular, extremes of SVR. Since the deep learning model extracted features on its own, without clinician input, it may have found hidden features important for modeling rapid and extreme physiologic changes that were too complex for more traditional pulse wave analysis technology to capture.

In vivo, many hemodynamic indices are collinear. For example, in hypovolemia, compensation occurs with increases in heart rate, contractility, and SVR, all of which may impact the accuracy of pulse wave analysis models. Furthermore, they are autocorrelated since they are time series data. Consider walking into an operating room and trying to predict the next set of vitals on the patient monitor. The provider in the room would likely be much more accurate than you because you lack memory of previous states. As individual hemodynamic features are related, often in complex and nonlinear ways specific to an individual’s unique physiology and their previous states, it is apparent that the dynamics of a fluid wave through a blood vessel may become unpredictable in abnormal physiology, or at extremes of hemodynamics when this type of monitoring becomes increasingly important. It follows that standard statistical approaches have significant limitations.

The nonlinearities observed in pulse wave arrival time and other plethysmography features are likely a consequence of the dynamic interplay between stroke volume and SVR, which requires artificial intelligence to fully capture. Thus, major improvements in diagnostic ability may be achieved by the use of machine learning in pulse wave analysis. This is particularly true at extremes of CO, SVR, and preload, when traditional pulse wave analysis becomes unreliable. The machine learning approach, uniquely equipped to capture complex nonlinearities in hemodynamic variables, may significantly enhance our understanding of human physiologic responses and our ability to monitor noninvasively.

The authors declare no competing interests.