Table 3.

Phenotyping studies using echocardiographic-derived parameters

Authors (year)	Patients (n)	Training data	Outcome of interest	Inclusion criteria	Algorithm used	Findings	Validation
Wojnarski et al (2017)39	656 patients.	Unsupervised training: the cross-sectional diameter of the aorta at each level taken from CT. Supervised training: 56 patient-level preoperative and echocardiographic variables trained to the outcome of the cluster label.	Association of bicuspid valve with patterns of aortopathy.	Patients with bicuspid aortic valve and aortic aneurysms for surgical repair.	Unsupervised training: partitioning around medoids. Supervised training: polytomous random forest analysis to an outcome.	Three aneurysm subtypes identified: those with root (13%), ascending (55%) and arch (32%) predominant. Severe valve regurgitation was most associated with the root phenotype (57%). AS was most commonly associated with arch phenotype (62%). Patient age increased as the extent of the aneurysm is more distal. Root phenotype had the highest male predominance (94%).	The clustering and resulting phenotypes could be defined by algorithmic rules which were then used for manual phenotyping, giving 94% accuracy. Then during supervised training, variable importance for classification was studied, with the five most important measurements being peak AV gradient, mean AV gradient, LV inner diameter, LV relative wall thickness, and bicuspid or unicuspid valve.
Kwak et al (2020)42	Training data: 398 patients. Validation data: 262 patients.	From 32 variables from demographics, physical examination, laboratory data, LV geometry, LV systolic function, LV diastolic function and aortic valve, 11 were used for clustering. 11 identified via dimensionality reduction, using Pearson coefficient and Bayesian information criteria that penalise model complexity. The 11 parameters in order of importance for clustering were haemoglobin, tricuspid regurgitant jet velocity, creatinine, left atrial volume, E-wave velocity, LVEF, BMI, heart rate, A-wave velocity, platelets and white cell count.	Primary: all-cause mortality. Secondary: cardiac mortality, non-cardiac mortality and death after AVR.	Newly diagnosed patients with moderate or severe AS.	Training model: model-based clustering. Validation model: agglomerative hierarchical clustering (Ward’s method).	Three clusters were identified: cluster 1 contained patients with depressed LVEF, more LV hypertrophy, more severe diastolic dysfunction and atrial fibrillation and had higher rates of cardiac death; cluster 2 contained elderly patients with comorbidities, specifically end-stage renal disease and had more non-cardiac death; cluster 3 was the lowest risk group with lowest event rates. These new labels showed improved discrimination (integrated discrimination improvement 0.029) and net reclassification improvement (0.294) for the outcome of 3-year all-cause mortality.	On an independent sample of 262 patients, model-based clustering was repeated and similar trends were identified. The modelling was repeated using agglomerative hierarchical clustering and three similar clusters with similar clinical characteristics were identified.
Sengupta et al (2021)41	1052 patients with CT, MRI and echos from three centres, prospective cohort.	Aortic valve area index, LV ejection fraction, aortic valve mean gradient, stroke volume indexed and aortic valve peak velocity.	Progression to AVR or progression to death.	Asymptomatic AS or discordant AS.	Patient similarity analysis: topological data analysis. Then, using new labels, created a supervised ML classifier.	High severity (57% of patients) and low severity (43% of patients) phenotypes. High severity group had higher aortic valve calcium score, more late gadolinium enhancement, higher BNP and high sensitivity troponin, and five times the risk of AVR. High severity patients who received AVR were also twice as likely to progress to death than all low severity patients. Echo-supervised classifier was very accurate and applied to the training data.	High and low severity labels were used to train a supervised machine learning classifier that had AUC of 0.988. This algorithm had better discrimination (integrated discrimination improvement of 0.07) and reclassification (net reclassification improvement of 0.17) for the outcome of AVR at 5 years compared with traditional valve severity grading.
Casaclang-Verzosa et al (2019)40	Training data: 346 patients with mild to severe AS. Validation data:155 mice at 3, 6, 9 and 12 months of age.	79 clinical and echocardiographic data (AVA, LVEF, LV mass index, relative wall thickness).	Understanding the progression of disease from mild to severe AS.	Patients with mild to severe AS.	Topological data analysis.	Topological data analysis created a Reeb graph that formed a loop with mild and severe AS on either spectrum and moderate AS forming two separate paths. Suggests that the path from mild to severe AS follows two prototypical paths via moderate AS. The severe AS area of the map was associated with higher mean gradient, LV mass index, E/e’, concentric hypertrophy and smaller AVA. This area of the Reeb graph associated with severe AS had 3.88 times the risk of valve intervention. When examining the two arms of moderate AS, although there was no difference in AVA, patients in the upper arm had lower EF, more frequently were men and had higher incidence of coronary artery disease. Patients in the upper arm had lower peak velocity, lower mean gradient, higher LV mass index and higher left atrial volume. In follow-up post-AVR, the patients’ loci in the Reeb graph regressed from the severe to the mild position.	A murine model for which echos were performed at 3, 6, 9 and 12 months of age. Topological data analysis of mice echo measurements also showed a similar trend to human patterns of AS. In addition, longitudinal imaging of the murine model provided insight into the natural progression of AS, with mice moving along the spectrum from mild to severe AS over time.

AS, aortic stenosis; AUC, area under the receiver operator curve; AV, aortic valve; AVA, aortic valve area; AVR, aortic valve replacement; BMI, body mass index; BNP, brain natriuretic protein; EF, ejection fraction; LV, left ventricular; LVEF, left ventricular ejection fraction; ML, machine learning.