Machine-learning approach on echocardiography to improve the detection of transthyretin amyloid cardiomyopathy: GRAAL algorithm.

Comprehensive Summary

This retrospective study analyzed the accuracy of a decision tree-based machine learning algorithm to detect transthyretin amyloid cardiomyopathy (ATTR-CM). The initial cohort (n = 260) consisted of patients referred for suspected ATTR-CM and diagnosed based on transthoracic echocardiography features and Perugini’s grade 2 or 3 cardiac uptake on scintigraphy. The decision tree identified four significant identifiers of ATTR-CM of which the GRRAL algorithm was named after: Global longitudinal strain (< 12%), RV free wall thickness (RVFWT ≥ 6.3 mm), Apical longitudinal spAring (RALS ≥ 0.94), and Left ventricular mass index (≥172 g/m²). The GRRAL algorithm had a significantly higher AUC (AUC = 0.90) compared to traditional detection scoring measures, such as the increased wall thickness (IWT) score (AUC = 0.73) and a clinical model derived from the RAISE score (AUC = 0.70). Adding the GRAAL algorithm on top of the clinical model, the clinical model and IWT score, and natriuretic peptides all resulted in an increased C-index (increase of 0.19, 0.08, and 0.4, respectively). Additionally, the combined C-index of the GRRAL algorithm on top of the clinical model and IWT score was 0.971.

Outcomes and Implications

ATTR-CM screening is largely limited by detection of echocardiographic features with low discriminative value, such as left ventricular hypertrophy or diastolic dysfunction, in populations with comorbidities. Thus, it is important to develop a high performance detection method to identify ATTR-CM in suspected patients, to ensure appropriate clinical intervention is performed as soon as possible. The GRAAL algorithm improved diagnostic accuracy on top of other scoring methods. GRAAL’s robust performance under a multi-center external validation (AUC = 0.83) indicates its potential as a highly accurate screening tool.