Machine learning in diastolic dysfunction: Left atrial strain trace superior to single points for estimation of filling pressure†

This editorial refers to ‘Noninvasive assessment of left ventricular end-diastolic pressure using machine learning–derived phasic left atrial strain’, by M.M. Gruca et al., https://doi.org/10.1093/ehjci/jead231.

Assessment of left ventricular filling pressure (LVFP) is of high clinical interest to monitor the progression of disease or response to treatment or when evaluating signs or prognosis of heart failure. However, such pressure measurements are performed only in selected patients due to the invasiveness, costs, and availability of cardiac catheterization. Non-invasive estimation of LVFP by echocardiography is therefore an attractive alternative in clinical practice. LVFP impacts filling velocities and tissue velocities as well as chamber geometry and deformation. Such parameters can be assessed by echocardiography and used as non-invasive estimates of LVFP, but their indirect relationship to LVFP limits their accuracy. It is therefore a continuous search to find new parameters to improve the estimation of LVFP. The introduction of speckle-tracking echocardiography has facilitated the assessment of left atrial (LA) strain, which has emerged as a promising parameter for estimating LVFP but not in time to be included in the 2016 joint recommendations to evaluate LVFP from the American Society of Echocardiography (ASE) and the European Association of Cardiovascular Imaging (EACVI).¹ These recommendations propose a multi-parametric approach as no single parameter has sufficient accuracy alone. The 2016 ASE/EACVI algorithm for use in patients without specific cardiovascular disease combines peak early (E) and late (A) diastolic mitral inflow velocities, early diastolic mitral annular velocity (e′), maximum LA volume index (LAVI), and peak tricuspid regurgitation velocity.

In a recent multicentre study, we tested the accuracy of this algorithm and showed how LA strain could be incorporated to improve its performance.² The LA strain time trace quantifies the longitudinal lengthening and shortening of the left atrium over the cardiac cycle. By convention, the trace starts at ventricular end-diastole and has normally three main components: (i) The left atrium stretches to a maximum (LA reservoir strain) during LV ejection in response to LV long-axis shortening. (ii) There is a passive LA shortening during the early LV filling phase and (iii) an active LA shortening (LA booster or pump strain) during atrial contraction. A limitation of the 2016 ASE/EACVI algorithm is that a missing parameter may leave a patient unclassified. In our study, we showed how either LA reservoir or pump strain could be applied on the unclassified patients, increasing the feasibility of the algorithm without compromising accuracy. This approach has since been incorporated in the 2022 EACVI consensus document for evaluating heart failure patients with preserved ejection fraction.³

The fact that both LA reservoir and pump strain contained information about LVFP illustrates the point that strain waveforms may contain more information than just one value at a given time. This is recognized in ischaemia where early-systolic stretch and post-systolic shortening are common features of ischaemic segments, as well as in left bundle branch block where septal flash and rebound stretch are common hallmarks. Integrating the information of the entire strain trace is therefore appealing. This is effectively done in pressure-strain loop analysis as pressure-strain loop area or the work index, which may in part explain the strengths of this method.⁴

This is also a strength of machine learning that typically is used to identify the link between a large number of features for a classification purpose. Machine learning models have been trained to grade diastolic dysfunction based on conventional echocardiographic parameters⁵ and recently also include LA reservoir strain.⁶ Another alternative would be to provide the whole LA strain waveform to the machine learning model. The strain value at each time point could be considered a separate feature and the machine learning model would identify the features in the waveform that are linked to LVFP and extract those. In this issue of the journal, Gruca et al.⁷ tested such a method where a machine learning model found the useful information in the LA strain waveform and condensed that to one value that they named the LA strain index (LASi), linked to the probability of a patient having normal or elevated LV end-diastolic pressure (LVEDP). They concluded that LASi classified LVEDP as well or better than LA reservoir strain alone and the 2016 ASE/EACVI algorithm in patients who underwent left heart catheterization. Furthermore, as only patients with assessable LA strain were included, 100% were classified by LASi and LA reservoir strain while 19% were left unclassified using the 2016 ASE/EACVI algorithm. Their study contributes to the growing support for LA strain as a non-invasive marker of LVFP.

With respect to LVFP, different measures can be used such as LVEDP, pulmonary capillary wedge pressure and LV pre-A pressure, where the latter two are markers of mean LA pressure (LAP). While LVEDP is more relevant for LV function, mean LAP may be a more relevant pressure for pulmonary congestion. It should be noted that there is a difference of clinical conditions with an increase in LVEDP only vs. an increase in both mean LAP and LVEDP. An elevated LVEDP without an increase of mean LAP includes patients in the early stages of diastolic dysfunction.¹ Braunwald⁸ showed that LVEDP exceeded mean LAP averaging 9 mmHg in patients with LV disease, providing evidence that LA contraction could elevate LVEDP, while mean LAP remained at a lower level. Another experimental study showed that LA reservoir function was determined by LA contraction, LA stiffness, and LV long-axis shortening.⁹ These studies addressed the importance of LA pump function to maintain LA reservoir capacity, thus preventing an increase of mean LAP and pulmonary congestion. In the study by Gruca et al.,⁷ the machine learning model was trained using LVEDP ≥ 15 mmHg as cut-off, and the resulting LASi may be used to identify patients including pre-clinical heart failure, enabling earlier strategies for further heart failure development.

Their study was carried out in patients without any specific cardiovascular disease, while the performance of LASi in conditions such as, for example, atrial cardiomyopathy, recent atrial arrhythmias, or tachycardia may be more challenging and should be tested. Having LASi as a single marker of LV filling pressure may be of limited value in some of these conditions.

The use of machine learning is rapidly growing. Computer algorithms can go deeper in the level of abstraction to a point beyond where humans can define patterns and features. This means that features that may not be visible to the human brain are ‘visible’ to the computer. In the study of Gruca et al.,⁷ only the LA strain waveform was used as input to the machine learning model. However, all relevant measurements could in theory be given to a machine learning model including flow and tissue velocity traces and even the echocardiographic raw images themselves so the machine learning model can find previously undiscovered patterns linked to LVFP and improve the classification. The main obstacle for development of such machine learning models is the lack of data as vast amounts are needed. However, the interest in testing machine learning drives generation of more data, and the more data become available, the more machine learning will be used. Hence, this self-reinforcing effect will most likely accelerate the growth of machine learning further.

Funding

None declared.

Data availability

No new data were generated or analysed in support of this research.

References

Author notes