Imaging Heart Failure With Artificial Intelligence

he technological origami is unfolding before our eyes; each leaf that unfurls bestows us with new challenges and opportunities. During the last 2 decades, several modalities in cardiovascular imaging have evolved with the sheer number of parameters utilized in the assessment of cardiac function. Particularly, the assessment of cardiac motion and deformation from echocardiography and magnetic resonance imaging creates a large volume of spatially and temporally diverse data. In the absence of suitable techniques, the majority of the data analysis currently is time constrained by graph-based assessments of spatially averaged functional data over a cardiac cycle. Computational modeling and the emerging algorithmic objectivity in machine learning can provide opportunities to resolve such dilemmas and minimize impediments that arise from high-dimensional imaging datasets. To this end, Sanchez-Martinez et al, 1 in this issue of Circulation: Cardiovascular Imaging, proposed the adoption of algorithmic objectivity to differentiate patients with heart failure with preserved ejection fraction from those who suffered dyspnea from other causes. This work extends the potential approaches suggested in previous investigations toward automation and acceleration of complex pattern recognition using cardiac imaging data for disease phenotyping. tilizing the data from MEDIA study (Metabolic Road to Diastolic Heart Failure), the authors examined 156 patients aged 60 years subjected to stress echocardiography. The group contained healthy, heart failure with preserved ejection fraction, dyspnea without heart failure with preserved ejection fraction, and asymptomatic hypertensive patients. A total of 22 descriptors with 300 features were collected pertaining to the tissue Doppler-derived longitudinal myocardial velocity traces at rest and during exercise. The large dimensions in the data set increase complexity in the model and can fit to noise. Therefore, to reduce the dimensionality by selecting principal features and preserving the elemental characteristics of the data, the authors applied an unsupervised machine-learning algorithm, which finds naturally occurring patterns and relationships within the data without labeled responses. his resulted in the large data getting condensed meaningfully and represented in low-dimensional data that is rich in features for differentiating healthy from diseased patients. Specifically, the technique of the unsupervised hierarchical clustering algorithm used in the study represents an effective way to visualize structure in the data to aid the assessment of data quality, bias detection, and discovery of informative groups. The resultant analysis revealed 2 clusters with important clinical differences in age, body mass index, left ventricular mass index, and left ventricular filling pressures assessed using conventional echocardiography and blood biomarkers. More importantly, the 2 groups were different symptomatically with 6-minute walk test distance being 31% shorter in one group than the other.