Comprehensive data integration—Toward a more personalized assessment of diastolic function

The main challenge of assessing diastolic function is the balance between clinical utility, in the sense of usability and time‐efficiency, and overall applicability, in the sense of precision for the patient under investigation. In this review, we aim to explore the challenges of integrating data in the assessment of diastolic function and discuss the perspectives of a more comprehensive data integration approach.

proportion of patients, whereas more complex algorithms, incorporating increased decision points and parameters, have proved to have low clinical utility in the real-world practical setting. Achieving a universal approach to the assessment of diastolic function therefore seems to be an intricate task that can hardly be approached with traditional algorithms, either simplified or complex. The optimal assessment of diastolic function and filling pressures should ideally integrate all relevant clinical information with all available structural and functional echocardiographic data, not a preselected set of parameters. The described assessment envisions a personalized approach to patient care, a high-reaching future goal in medicine.
In this review, we aim to explore the challenges of integrating data in the assessment of diastolic function and discuss the perspectives of a more comprehensive data integration approach.

| A SS E SS ING D IA S TOLI C FUN C TI ON-THE QUE S T FOR A UNIVER SAL APPROACH
The majority of current ideas and pitfalls surrounding noninvasive assessment of diastolic function were recognized and defined in the seminal work from Appleton, Hatle, and Popp, 1 relating distinct transmitral flow velocity patterns to LV diastolic function. The observed flow patterns were more related to myocardial dysfunction and hemodynamic status than the type of underlying disease, setting ground for future classification of diastolic dysfunction into grades ( Figure 1). Although these grades are pathophysiologically interpretable, the patterns of mitral inflow represent a dynamic continuum, changing with regard to disease progression, medical therapy or alterations in hemodynamic status. Ongoing research showed that the correlation of mitral inflow parameters and pressure measurements is influenced by overall cardiac function, resulting in the fact that transmitral flow parameters do not correlate with LV filling pressures in patients with preserved ejection fraction, whereas they do in reduced LV function. 2 Interpreting any surrogate diastolic parameter is inherently complex, as most Doppler patterns demonstrate varying dependency on the inotropic state, volume loading, ventricular relaxation, chamber compliance, and left atrial pressure, as well as on additional factors such as age, heart rate, blood pressure, mitral valve pathology, among others. [3][4][5][6] Therefore, to correctly interpret findings and assess function, it is crucial to recognize a wider pattern including clinical history, diagnostic data, echocardiographic patterns, and their dynamic changes.
F I G U R E 1 Diastolic assessment using the 2016 guidelines. (Rows) Four hypertensive patients with varying degrees of diastolic dysfunction. (Columns from left to right) The patient history, signs and symptoms, recommended echocardiographic parameters, and diastolic grades assessed using the 2016 guidelines. 14 Diastolic dysfunction can be assessed in a straightforward way using the four echocardiographic parameters proposed by the guidelines. The grade of dysfunction concurs with the associated clinical picture. NT-proBNP levels were normal in the first two patients (<14.34 pg/mL), and slightly elevated in the third and fourth (303 and 731 pg/mL, respectively). BMI = body mass index; EF = ejection fraction; DM = diabetes mellitus; ARB = Angiotensin II receptor blocker; ACEi = Angiotensinconverting-enzyme inhibitors; FA = atrial fibrillation; PW TDI = pulsed wave tissue Doppler imaging; LAVI = left atrial volume indexed to body surface area; LV GLS = left ventricular global longitudinal strain; STE = speckle-tracking echocardiography To address these challenges and resolve the ambiguity of the pseudonormalization pattern, various additional tests and parameters were suggested over time-the alteration of loading conditions with a Valsalva test, 7 the addition of pulmonary venous velocity curves [8][9][10] or tissue Doppler imaging (TDI) 2,11,12 -ultimately resulting in more complex algorithms. As an example, with the addition of the ratio between early diastolic transmitral flow and TDI velocities of the mitral ring (ie, E/e′) the assessment of diastolic function in patients with preserved EF was somewhat simplified. However, this addition ultimately created a new grayzone in the intermediate range of the ratio, where further assessment and parameters were mandatory to assess underlying diastolic function (eg, pulmonary flow velocities or the Valsalva manoeuvre). 2 This need for a wide combination of parameters in noninvasive diastolic function assessment, together with alterations of algorithms in specific patient populations, was thus emphasized in the ASE/EACVI 2009 guidelines for diastolic assessment. 13 Besides parameters of diastolic function and associated measurements (ie, mitral inflow velocities, Valsalva manoeuvre, pulmonary venous flow, and TDI velocities), morphologic and functional ing the diagnosis of first grade dysfunction. 15 A major limitation of the guidelines was still the lack of consideration of age-where age influences the findings of diastolic parameters. 16 Recent efforts have been made in addressing the challenges of age-appropriate interpretation of diastolic patterns by applying age-specific multivariate ref- Several studies 19,20 demonstrated that the 2016 guidelines proved to have higher sensitivity in estimating the filling pressures in patients with reduced EF as compared to the 2009 guidelines, while the low sensitivity was still present in patients with normal EF and normal filling pressures. However, more data integration-combining demographic and clinical variables with noninvasive echocardiographic parameters-showed an incremental value when diagnosing elevated filling pressures. 21 On the other hand, stratification into diastolic grades has been strained by the lack of relationship to cardiovascular outcomes, complicating the clinical utility of undergoing complex algorithms to identify a diastolic class. While various diastolic parameters proved predictive of clinical outcomes in studies, [22][23][24][25][26] combining parameters in classifications to define grades showed no consistent relation to outcomes 27,28 -showing worse outcomes in moderate/severe compared with mild diastolic dysfunction, 29 or only in severe dysfunction. 30 A universal diastolic grading approach therefore evidently lacked clear clinical value.
Novel imaging techniques like speckle tracking echocardiography (STE) are also increasingly in focus, as they can offer a wealth of embedded information on the systolic and diastolic function, and provide insight into patterns of myocardial mechanics that correlate with diastolic parameters and cardiovascular outcomes. 14,31 The wealth of data that can be obtained using these techniques is still under research and therefore clinically underused. 32 Analysis of single-beat STE based LV and LA volume and strain peak velocity and timing measurements resulted in patient groups with increasing severity of diastolic dysfunction and LV filling pressures (validated by invasive measurements), proving that information derived from STE variables can indeed be useful for assessment of diastolic dysfunction. 38 Moreover, STE indices of diastolic function showed to be an important discriminator between heart failure phenogroups. 34 Deformation data also carry immense information in exercise testing, especially in the subset of patients with diastolic dysfunction that may have normal hemodynamic profile at rest but symptoms of heart failure or dyspnea in effort. Typically, the data from these exercise tests are complex to integrate and therefore conclusions are reduced to the comparison of only selected measurements at rest and exercise.

| A SS E SS ING FUN C TI ON IN CHALLENG ING PATIENTS -THE LIMITATI ON S OF A UNIVER SAL APPROACH
The described overview of noninvasive diastolic function assessment shows, consistently and somewhat paradoxically, that a universal approach is feasible only by sacrificing precise assessment in special patient populations where noninvasive parameters and the corresponding patterns are influenced by related comorbidities. For example, mitral valve disease or regional deformation impairment due to ischemic disease or genetic-sarcomere mutations can alter the mitral inflow pattern, TDI velocity profile and the related ratios,

| MOVING TOWARD MORE COMPREHEN S IVE DATA INTEG R ATI ON OF THE WHOLE C ARD IAC C YCLE IN THE A SS E SS MENT OF D IA S TOLI C FUN C TI ON
The addition of whole cardiac cycle data extracted from echocardiographic images (eg, volume, blood-pool and myocardial velocity, strain or strain-rate curves) to the assessment of diastolic function serves as a step toward a more sophisticated data integration strategy. Heterogeneity of diastolic dysfunction is an appropriate challenge for machine learning (ML), especially unsupervised approaches, 31 which aim to extract hidden patterns in available data and naturally cluster patients regardless of a priori knowledge or predefined clinical labels. Such algorithms have recently been used to approach diastolic dysfunction classification. Using recommended parameters for diastolic assessment, an unsupervised clustering approach identified unique patterns of diastolic dysfunction that showed a relationship to clinical outcomes, as opposed to current grading schemes. 32 Importantly, patients classified as indeterminate by guidelines were reclassified into an appropriate risk group.
In other studies, a combination of variables (ie, demographic, clinical, laboratory, ECG, and echo) have been used to explore heart failure phenotypes that differ in outcomes and therapy response 35,36 ; and also, to investigate HF phenogroups with data on invasive hemodynamics, altogether showing that the severity of diastolic dysfunction seems to be one of the main separating factors between these phenogroups. 36,37 Precise phenotyping of diastolic function inevitably influences patient care, for example, optimal patient management requires differentiation between abnormal relaxation, when heart F I G U R E 3 An example of data integration in the assessments of a complex patient. A female with long-standing arterial hypertension and clinically diagnosed obstructive hypertrophic cardiomyopathy. The posterior part of the mitral annulus is calcified, moderate mitral regurgitation is present, and the basal septum is hypertrophied, measuring 17 mm. All of the latter influence traditional interpretation of diastolic parameters. Additional investigation is needed. The patient had elevated blood pressure at assessment, which can influence findings. The obstruction is highest in the midventricular region, with the gradient reaching 51 mm Hg during the Valsalva manoeuvre. During Valsalva, the inversal of the pseudo-normal mitral inflow can be noted. The E/E′ ratio indicates elevated filling pressure, supported by the difference in the timings of the pulmonary vein and mitral inflow A-wave duration, LA is enlargement and tricuspid regurgitation velocity. (Abbreviations same as in Figure 1, COPD = chronic obstructive pulmonary disease) rate reduction is beneficial, and decreased compliance, when the latter is not the case. 38 The distinction can be found through comprehensive data assessment incorporating a wide set of parameters, stepping out of the scope of simplified algorithms of classification.
ML approaches can aid in standardizing echocardiographic evaluation using unlabelled variables without a priori algorithms, isolating prognostic phenotypes not visualized by guideline algorithms.
In disease exploration, both the traditional consensus-based and the described ML approaches are constrained to a limited number of key disease markers and clinical variables, such as selected peak value or timing measurements. These might not capture the full complexity and subtle changes of the underlying diseases. Specifically, spatiotemporal patterns of myocardial velocity curves, defined by peak and timing values throughout the whole cardiac cycle, are reflective of regional and global dysfunction in systole and diastole 39 and reveal intricate changes in myocardial mechanics in specific cardiac pathologies. 40 Similarly to when a clinician integrates these data based on previous experience and knowledge, novel machine learning techniques offer the possibility to incorporate information embedded in the velocity data of the whole cardiac cycle, with the aim to extract the maximum amount of information reflective of cardiac function and disease from cardiac images. This approach could also be used to analyze the complex changes occurring between rest and exercise echocardiography. Moreover, as atrial fibrillation still serves as an exclusion criterion in many ML algorithms looking at cardiac echo measurements, 41,42 a whole cardiac cycle data approach incorporating multiple consecutive beats could present a potential way forward in addressing this challenge. 43 Moreover, pathology-related information is contained not only in the amplitude and profile of a velocity curve, but likewise in the timings and durations of different cardiac phases (eg, isovolumic contraction or early diastole). 44 Temporal differences, due to inter-patient variability in heart rate or intra-patient variability between rest and exercise, result in a challenging interpretation of the relationship between cardiac phases (eg, when assessing a shift in the onset of systole/diastole due to dysfunction, see Figure 4). Since the timings of cardiac phases can easily be defined with echocardiographic (valve flows) and ECG (onset of atrial contraction) data, time alignment of echo data is feasible as part of the ML approach. 39,[43][44][45] Velocity data can be time aligned to a common temporal reference within a patient cohort and quantitatively compared between patients. Data on the corrected differences in timings can be preserved and used as an additional parameter in later analysis.

F I G U R E 4
A scheme showing the utility of temporally aligning velocity traces. A, Temporal noncorrespondence of the velocity traces can be due to intersubject differences in heart rate and in the timing of cardiac phases. B, Temporal alignment can be used to express velocity traces within a common temporal reference. C, Temporally aligned velocity traces can be directly compared between patients enabling the assessment of the onset and duration of cardiac phases. A later onset of systolic LV ejection, and a prolonged LV ejection and isovolumic relaxation time can be seen in the patient on the right. This concurs with the delayed and reduced peak aortic velocity and the fusions of the early and late diastolic filling An important matter to assess is if the theoretical advantage of whole cardiac cycle data integration adds any real advantages in disease exploration. To address this question, a ML approach integrating spatiotemporal information from rest and exercise echocardiographic data (including velocity, strain, and strain rate curves, respectably) was used to create spatiotemporal-rest-exercise representations of the LV function. 39 This comprehensive whole cardiac cycle data proved more successful than traditional measurements

| CON CLUS ION
The balance between clinical utility, in the sense of usability and time-efficiency, and overall applicability, in the sense of precision for the patient under investigation, represents the main challenge in the assessment of diastolic (dys)function. The high-reaching aim of personalized medicine that could resolve these tensions may be feasible through a more comprehensive integration of all relevant data-from clinical to whole cycle echocardiographic data. Complete data integration seems to be a long-lasting goal, the way forward in diastology, and machine learning seems to be one of the tools suited for the challenge. Each successful integration of heterogeneous data to patient assessment offers incremental value to the goal of better understanding complex topics such as diastolic dysfunction or HFpEF. With more comprehensive approaches, we can see improved shaping of disease phenotypes and better relation of these phenotypes to underlying pathophysiological processes-which may help affirm diagnostic reasoning, guide treatment options and reduce limitations related to previously unaddressed confounders. The aim has slowly shifted from strict categorical classifications of disease/ health toward the exploration of disease as a continuous spectrum, ranging from health to dysfunction, with the novel goal being personalized positioning of patients into a certain part of this spectrum.
Finally, the main clinical value can be harvested from relating newfound distinct phenotypes to long-term patient trajectories, a goal consistently highlighted in contemporary publications. With perpetual proof that traditional approaches to complex problems are not the optimal solution, there is room for a steady and cautious, and inherently very exciting paradigm shift toward novel diagnostic tools and workflows to reach a more personalized, comprehensive and integrated assessment of cardiac function.