Machine Learning in Cardiac Imaging: Exploring the Art of Cluster Analysis

Machine learning cluster analysis identifies increased 12-month mortality risk in transcatheter aortic valve replacement recipients

Background

Long-term mortality risk is seldom re-assessed in contemporary clinical practice following successful transcatheter aortic valve implantation (TAVR). Unsupervised machine learning permits pattern discovery within complex multidimensional patient data and may facilitate recognition of groups requiring closer post-TAVR surveillance.

Methods

We analysed and differentiated routinely collected demographic, biochemical, and cardiac imaging data into distinct clusters using unsupervised machine learning. k-means clustering was performed on data from 200 patients who underwent TAVR for severe aortic stenosis (AS). Input features were ranked according to their influence on cluster assignment. Survival analyses were performed with Kaplan-Meier and Cox proportional hazards models. Nested cox models were used to identify any incremental prognostic benefit cluster assignment achieved beyond conventional risk scores.

Results

Analysis identified two distinct clusters. Compared to Cluster 1, Cluster 2 demonstrated significantly worse all-cause mortality at 12 months (HR 6.3, p < 0.01), and was characterised by more advanced cardiac remodelling with worse indices of multi-chamber cardiac function, as quantified by strain imaging. Cluster assignment demonstrated greater predictive power for 12-month mortality as compared with conventional risk and frailty calculators.

Conclusion

k-means clustering identified two prognostically distinct phenogroups of patients who had undergone TAVR with better discriminatory power than conventional risk and frailty calculators. Our results highlight the utility of machine learning applications for clinical risk prediction and scope to improve patient surveillance.

Machine learning prediction of progressive subclinical myocardial dysfunction in moderate aortic stenosis

Background

Moderate severity aortic stenosis (AS) is poorly understood, is associated with subclinical myocardial dysfunction, and can lead to adverse outcome rates that are comparable to severe AS. Factors associated with progressive myocardial dysfunction in moderate AS are not well described. Artificial neural networks (ANNs) can identify patterns, inform clinical risk, and identify features of importance in clinical datasets.

Methods

We conducted ANN analyses on longitudinal echocardiographic data collected from 66 individuals with moderate AS who underwent serial echocardiography at our institution. Image phenotyping involved left ventricular global longitudinal strain (GLS) and valve stenosis severity (including energetics) analysis. ANNs were constructed using two multilayer perceptron models. The first model was developed to predict change in GLS from baseline echocardiography alone and the second to predict change in GLS using data from baseline and serial echocardiography. ANNs used a single hidden layer architecture and a 70%:30% training/testing split.

Results

Over a median follow-up interval of 1.3 years, change in GLS (≤ or >median change) could be predicted with accuracy rates of 95% in training and 93% in testing using ANN with inputs from baseline echocardiogram data alone (AUC: 0.997). The four most important predictive baseline features (reported as normalized % importance relative to most important feature) were peak gradient (100%), energy loss (93%), GLS (80%), and DI < 0.25 (50%). When a further model was run including inputs from both baseline and serial echocardiography (AUC 0.844), the top four features of importance were change in dimensionless index between index and follow-up studies (100%), baseline peak gradient (79%), baseline energy loss (72%), and baseline GLS (63%).

Conclusions

Artificial neural networks can predict progressive subclinical myocardial dysfunction with high accuracy in moderate AS and identify features of importance. Key features associated with classifying progression in subclinical myocardial dysfunction included peak gradient, dimensionless index, GLS, and hydraulic load (energy loss), suggesting that these features should be closely evaluated and monitored in AS.

Unsupervised Machine Learning for Assessment of Left Ventricular Diastolic Function and Risk Stratification

BACKGROUND

The 2016 American Society of Echocardiography (ASE) guidelines have been widely used to assess left ventricular diastolic function. However, limitations are present in the current classification system. We aimed to develop a data-driven, unsupervised machine learning approach for diastolic function classification and risk stratification using the left ventricular diastolic function parameters recommended by the 2016 ASE guidelines; the guideline grading was used as the reference standard.

METHODS

Baseline demographics, heart failure hospitalization and all-cause mortality data were obtained for all adult patients who underwent transthoracic echocardiography at Mayo Clinic Rochester in 2015. Patients with prior mitral valve intervention, congenital heart disease, cardiac transplant, or cardiac assist device were excluded. Nine left ventricular diastolic function variables (mitral E and A wave peak velocities, E/A, deceleration time, medial and lateral annulus e' and E/e', and tricuspid regurgitation peak velocity) were used for an unsupervised machine learning algorithm to identify different phenotype clusters. The cohort average of each variable was used for imputation. Patients were grouped according to the algorithm-determined clusters for Kaplan-Meier survival analysis.

RESULTS

Among 24,414 patients, age 63.6 ±16.2 years, all-cause mortality occurred in 4,612 (18.9%) patients during median follow-up 3.1 years. The algorithm determined 3 clusters with echocardiographic measurement characteristics corresponding to normal diastolic function (n= 8,312), impaired relaxation (n=11,779) and increased filling pressure (n =4,323), with 3-year cumulative mortality of 11.8%, 19.9% and 33.4%, respectively (p<0.0001). All 10,694 (43.8%) patients classified as indeterminate were reclassified into the 3 clusters (3,324, 5,353, and 2,017, respectively) with 3-year mortality of 16.6%, 22.9% and 34.4%, respectively. The clusters also outperformed guideline-based grade for prognostication (c-index: 0.607 vs. 0.582, p=0.013).

CONCLUSIONS

Unsupervised machine learning identified physiologically and prognostically distinct clusters based on 9 diastolic function Doppler variables. The clusters can be potentially applied in echocardiography laboratory practice and future clinical trials for simple, replicable diastolic function related risk stratification.

Automated Pattern Recognition in Whole-Cardiac Cycle Echocardiographic Data: Capturing Functional Phenotypes with Machine Learning

BACKGROUND

Echocardiography provides complex data on cardiac function that can be integrated into patterns of dysfunction related to the severity of cardiac disease. The aim of this study was to demonstrate the feasibility of applying machine learning (ML) to automate the integration of echocardiographic data from the whole cardiac cycle and to automatically recognize patterns in velocity profiles and deformation curves, allowing the identification of functional phenotypes.

METHODS

Echocardiography was performed in 189 clinically managed patients with hypertension and 97 healthy individuals without hypertension. Speckle-tracking analysis of the left ventricle and atrium was performed, and deformation curves were extracted. Aortic and mitral blood pool pulsed-wave Doppler and mitral annular tissue pulsed-wave Doppler velocity profiles were obtained. These whole-cardiac cycle deformation and velocity curves were used as ML input. Unsupervised ML was used to create a representation of patients with hypertension in a virtual space in which patients are positioned on the basis of the similarity of their integrated whole-cardiac cycle echocardiography data. Regression methods were used to explore patterns of echocardiographic traces within this virtual ML-derived space, while clustering was used to define phenogroups.

RESULTS

The algorithm captured different patterns in tissue and blood-pool velocity and deformation profiles and integrated the findings, yielding phenotypes related to normal cardiac function and others to advanced remodeling associated with pressure overload in hypertension. The addition of individuals without hypertension into the ML-derived space confirmed the interpretation of normal and remodeled phenotypes.

CONCLUSIONS

ML-based pattern recognition is feasible from echocardiographic data obtained during the whole cardiac cycle. Automated algorithms can consistently capture patterns in velocity and deformation data and, on the basis of these patterns, group patients into interpretable, clinically comprehensive phenogroups that describe structural and functional remodeling. Automated pattern recognition may potentially aid interpretation of imaging data and diagnostic accuracy.