Echocardiographic evaluation of left ventricular function using an automated analysis algorithm is feasible for beginners and experts: comparison with invasive and non-invasive methods

Evaluation of left ventricular function by transthoracic echocardiography (TTE) is at the corner of many diagnostic and therapeutic decisions in cardiovascular care and one of the most often performed examinations in clinical practice. Measurement of left ventricular ejection fraction (LVEF) is often done by visual estimation from experienced echocardiographers or manual biplane measurement according to the Simpson method [1, 2]. Assessment of left ventricular parameters with three-dimensional echocardiography circumvents the limitations associated with two-dimensional echocardiography such as apical foreshortening and geometric assumptions and has been shown to be reliable and reproducible [3‐5]. Use of three-dimensional echocardiography is therefore recommended by recent guidelines for the assessment of left ventricular volume parameters [2]. However, translation of this recommendation has been hampered by lack of experience of clinicians with three-dimensional echocardiography and the need for time-consuming three-dimensional reconstructions, limiting effective clinical workflow. Additionally, manually reconstructed 3D volumes might show substantial variability [6]. Novel, algorithm-based approaches are tempting as they possibly allow quick and accurate measurements for even unexperienced investigators. Yet, widespread use of such algorithms is limited due to uncertainty whether automated measurement of left ventricular function is reliable in daily practice. The aim of this work was therefore to compare left ventricular functional parameters obtained from three-dimensional transthoracic echocardiography using an automated software algorithm (HeartModel^A.I., Philips Healthcare) against qualitative, visual grading using left ventricular angiography and quantitative assessment of LV-function using cardiovascular magnetic resonance imaging (CMR), which is currently accepted as the clinical gold standard in this matter [7, 8].

In total, transthoracic studies from 198 patients were suitable for analysis.

Within this cohort, 39 (19.7%) patients received cardiac magnetic resonance imaging (CMR) and 159 (80.3%) left ventricular angiography within 24 h to the echocardiographic examination. Echocardiographic image quality was graded as “poor” in n = 0 patients (0%), “good” in n = 75 patients (37.9%) and “excellent” in n = 123 patients (62.1%). Left ventricular parameters such as LVEF, LVEDV and LVESV as derived from CMR and 3D-heart model were 53.1% (± 11.5), 160.3 ml (± 51.4), 78.3 ml (± 46.9) and 53.9% (± 14.9), 167.1 ml (± 65.5) and 82.9 ml (± 60.5) respectively, showing no significant differences between the two aforementioned methods (all p > 0.05, see Table 2). In the group of patients that underwent left ventricular angiography and echocardiography (n = 159), LVEF, LVEDV and LVESV derived from 3D-Heartmodel were 55.1 % (± 9.5), 164.8 ml (± 76.0) and 78.0 ml (± 55.2), respectively. Comparison of left ventricular angiography and 3D-Heartmodel showed no significant difference with respect to left ventricular ejection fraction (p > 0.05, see Table 2).

Correlation of left ventricular parameters was assessed using Pearson’s method. Correlation of LVEF for measurement by 3D-Heartmodel and cardiac magnetic resonance was overall good to excellent with r = 0.824, p < 0.0001, while correlation of LVEDV and LVESV was r = 0.921 and r = 0.944, respectively.

Correlation of 3D-Heartmodel versus left ventricular angiography was less pronounced, however still good with r = 0.745, p < 0.0001. Similar, Bland–Altman plots demonstrated good agreement for both modalities with a bias of − 0.74 for CMR and − 0.89 for LVA. (see Table 3 and Fig. 2). Intra- and interobserver variability of measuring LVEF using 3D-Heartmodel was low (ICC = 0.979 and 0.917, respectively). When comparing examinations based upon experience of the clinician, there was overall high significant correlation between 3D-HM and CMR in clinicians with low experience (beginner, EF r = 0.815, p = 0.002) and expert echocardiographers (EF r = 0.838, all p < 0.001). Comparing 3D-HM and LVA yielded similar results, however with lower correlation with EF r = 0.605 and r = 0.765, respectively (all p < 0.05, see Table 4)

Next, we aimed to understand whether 3D-HM could confirm the determination of heart failure class by CMR or LVA: all patients were categorized into (1) EF < 40% (= heart failure with reduced EF or HFrEF), (2) EF 41-49% (= heart failure with mildly reduced EF or HFmrEF) or (3) EF > 50% (= heart failure with preserved EF or HFpEF). High agreement was seen between CMR and 3D-HM as well as LVA and 3D-HM regarding confirmation of heart failure class (Fig. 3): this means that for the majority of patients, respective heart failure category (determined by CMR) was confirmed when using 3D-HM (agreement of 85.7% of cases for HFrEF, 33.3% for HFmrEF and 96.5% for HFpEF, Fig. 3a). Similarly, when comparing LVA with 3D-HM, agreement was seen for 69.2% in cases of HFrEF, 81.8% for HFmrEF and 91.9% for HFpEF (Fig. 3b). In total, complete agreement was detected in 89.7% of patients (CMR with 3D-HM) and 89.3% (LVA with 3D-HM).

Finally, we aimed to investigate the impact of echocardiographic image quality upon correlation of 3D-Heartmodel with CMR and LVA: in echocardiographic examinations with image quality graded as “good”, LVEF measurement by 3D-Heartmodel and cardiac magnetic resonance was r = 0.806, r = 0.874 for EDV and r = 0.945 for ESV. R for correlation of 3D-Heartmodel versus left ventricular angiography was 0.694. In patients in which image quality was scored as “excellent”, correlation was slightly higher for EF and EDV (r = 0.827 and r = 0.919) and slightly lower for ESV (r = 0.940, see Table 5). Additionally, patients with an obese phenotype (BMI >25) showed lower correlation between 3D-Heartmodel and cardiac magnetic resonance for EF (r = 0.720 in patients with BMI >25 versus r = 0.889 for patients with BMI < 25) and EDV (r = 0.891 versus r = 0.940) but similar correlation for ESV (0.959 in patient with BMI < 25 and 0.917 in BMI > 25, see Table 6).

The aim of this study was to compare parameters of left ventricular function (i.e., LVEF, LVEDV and LVESV) derived by a novel, algorithm-based analysis (3D-Heartmodel^A.I.) with clinically established quantification (i.e., same-day measurement done by either cardiac magnetic resonance or left ventricular angiography). The main findings of our work are given as follows:

Evaluation of left ventricular function is extremely important in daily clinical care. Accurate and reliable evaluation of left ventricular ejection fraction in the clinical setting is essential, since it has a major impact on further diagnostic testing and clinical workflow and is most frequently performed by two-dimensional transthoracic echocardiography with Simpsons’ biplane measurements. This is especially true for heart failure patients where adequate monitoring of disease progression and pharmacological treatment is essential [9, 10]. However, limitations associated with two-dimensional echocardiography, such as foreshortening or poor endocardial demarcation especially in patients with unfavorable characteristics, make the use of biplane measurements of left ventricular parameters error-prone and observer-dependent [11, 12]. We speculate that the use of an automated algorithm may help to reduce time-consuming collection of left ventricular data, which is beneficial in high-volume echocardiography laboratories. As such, we were able to demonstrate that the proposed method yielded satisfying results even in patients with obese phenotype with BMI > 25 (r = 0.720).

Semi-quantitative assessment of left ventricular dimensions using cardiac magnetic resonance is considered the gold standard for quantification of left ventricular dimensions [8] but local availability of CMR is often limited, requires expert radiologists and is contraindicated in a significant subset of patients (e.g., after implantation of a pacemaker or ICD). Likewise, left ventricular angiography allows detailed visualization of the left ventricle [13]. However, this technique has several limitations since its monoplane technique is often inaccurate, its application is limited due to its invasiveness and the need for contrast medium and radiation as well as possible side effects associated with catheterization of the left ventricle. Echocardiography on the other side is widely available, non-invasive, inexpensive and free of radiation (Fig. 4). Three-dimensional echocardiography could circumvent the limitations associated with two-dimensional echocardiography [10]. As compared with cardiac magnetic resonance, three-dimensional echocardiography has been shown to be more accurate and precise than conventional two-dimensional measurement but underestimating left ventricular volumes [6], being cumbersome, time-consuming and requiring special expertise. Therefore, clinical application has still been limited. Recently developed novel software-based tools with automated and quick 3D-rendering (such as the 3D-Heartmodel) use a distinct algorithm trained upon approximately thousand different echocardiographic imaging data sets enabling automated and quick contour detection of the left ventricle in end-diastole and end-systole [14] and was shown to be accurate and more reproducible in estimating left ventricular volumes compared with manual 3D-echocardiography [15, 16]. This allows user-friendly and more investigator-independent measurement of left ventricular parameters. Several studies have shown that inter- and intraobserver variability of 3D-Heartmodel is fairly low, allowing high reproducibility of this method in clinical routine [14, 17, 18]. In this study, we decided to retrospectively investigate the real-world performance of 3D-Heartmodel, comparing it with two distinct approaches often used in daily practice: on the one hand, we used same-day magnetic resonance imaging with multiplane quantification of LVEDV, LVESV and derived LVEF by an expert radiologist. On the other hand, we used same-day left ventricular angiography obtained during left heart catheterization, in which LVEF was visually graded by an expert clinician.

Recently, a comparison between 3D-Heartmodel and manual 3D echocardiography showed good correlation (r = 0.79) [19]. Additionally, a study of 65 patients undergoing 3D-echocardiography in combination with the first generation of the AI-based heart model algorithm and magnetic resonance imaging showed high agreement between these two modalities with respect to left ventricular parameters (r = 0.84–0.95) [14]. Similar, Levy at al. found an equally high and significant correlation of LVEF between 3D-Heartmodel and CMR with r = 0.93 [17] which was later confirmed in a study of 20 patients by Narang et al. [20]. Finally, comparison of 3D-Heartmodel showed good agreement with conventional biplane method, manual 3D-echocardiography and CMR in 200 patients [21]. However, such studies mainly using experienced sonographers for echocardiographic imaging in a preset setting may not reflect real-life conditions where often unexperienced sonographers are involved. Therefore, our retrospective analysis is informative as the echocardiographic examinations done in our study were performed by clinicians with varying expertise in echocardiography, demonstrating high correlation even in unexperienced clinicians.

Additionally, we demonstrate that the 3D-Heartmodel is able to stratify the overwhelming majority of patients into correct heart failure categories (approximately 90% for both CMR and LVA). Hence, even though the EF might be slightly under- or overestimated, 3D-HM could allow proper risk stratification of patients with heart failure, which is important for therapeutic decision such as intensified pharmacotherapy.

To the best of our knowledge, no study has so far investigated the comparison of 3D-Heartmodel with visual estimation of left ventricular function by left ventricular angiography. While left ventricular angiography is no longer regarded as the clinical gold standard for quantification of left ventricular function and visual estimation of left ventricular function can be criticized for its possible inaccuracy, it is the authors’ experience that it is still widely used in daily clinical practice.

In conclusion, an echocardiographic-based automated software algorithm to obtain left ventricular functional parameter showed excellent agreement with clinically established methods such as CMR or invasive angiography and can easily be used in the clinical setting. Our findings support the use and application in daily clinical practice. Larger, prospective studies are needed to validate our findings and to assess its impact of subsequent clinical work-up.