Echocardiography machine learning based to improve detection of transthyretin cardiac amyloidosis: The R3M Algorithm

Abstract

Introduction

Transthyretin cardiac amyloidosis (ATTR-CA) is an emerging cause of heart failure. The screening of ATTR-CA remains difficult since its echocardiographic features are analogous to those observed in patients with age- and hypertension-related cardiac remodeling.

Method

We retrospectively included 264 patients (76 ± 13 years old, 59% male) referred for suspected ATTR-CA. A supervised machine learning diagnosis algorithm differentiating patients with (n = 112) and without (n = 152) ATTR-CA was constructed based on echocardiographic data, and subsequently validated in an external multicenter cohort of 455 patients (76 ± 13 years old, 61% male).

Results

Patients with ATTR-CA had a lower systolic function (LVEF 47.4 ± 11 vs. 54.3 ± 12%, P < 0.001), left ventricular (LV) global longitudinal strain (GLS) (11.0 ± 3.7 vs. 14.2 ± 4.5%, P < 0.001) and more significant relative apical longitudinal sparing (RALS) (1.5 ± 1.2 vs. 0.9 ± 0.4, P < 0.001) compared to controls. Machine learning identified right ventricular free wall thickness (RVFWT), RALS, relative wall thickness (RWT), and LV mass index as key variables for identifying ATTR-CA (AUC 0.88 [0.84–0.92]; P < 0.001). The diagnostic value of this R3M (RVFWT, RALS, RWT and LV Mass index) algorithm was good in the validation multicenter cohort with an AUC of 0.79 [0.75–0.83] P < 0.001. The R3M algorithm further improved diagnostic accuracy over the IWT (Increased Wall Thickness) guidelines score (increase in C-index of 0.15 [0.10–0.21], P < 0.001).

Conclusion

The simple R3M algorithm based on echocardiographic data exploring RVFWT, apical sparing, and concentric hypertrophy displays good diagnostic accuracy for ATTR-CA and could represent an efficient screening tool (Fig. 1).