Motion-Informed Neural Network Towards Diastolic Mechanics in Takotsubo Syndrome
Circulation: Cardiovascular Imaging · 2026-11-11
DOI: 10.1161/circimaging.126.020258 ↗Abstract
BACKGROUND: The pathophysiology of Takotsubo syndrome (TTS) remains perplexing. We developed a motion-informed neural network (MiNN-Echo) using echocardiograms to elucidate the diastolic mechanics of TTS. METHODS: We trained and validated MiNN-Echo, a deep convolutional neural network (2D+t) echocardiographic prediction model using multi-vendor echocardiograms from 1142 patients (596 TTS and 546 anterior myocardial infarction) in a multi-institutional cohort. A total of 68 expert echocardiographic readers performed human interpretation. The network’s dual-branch architecture independently learns spatial and temporal features, using 2-dimensional image frames for spatial representation and segmentation, and optical flow with sinusoidal phase encoding for temporal dynamics. The MiNN-Echo model integrates static cardiac anatomic changes with motion details from 2-dimensional echocardiographic videos to produce interpretable disease predictions. RESULTS: The MiNN-Echo model achieved an area under the curve of 0.77, compared with 0.69 among human readers in differentiating TTS from acute myocardial infarction. Myocardial motion features during diastole contributed more (≈71%) to the differential diagnosis than those during systole. Compared with patients with acute myocardial infarction, TTS patients showed increased isovolumic relaxation time (112.5 [interquartile range (IQR), 68.75–165.94] ms versus 87.5 [IQR, 54.17–140] ms, P
EUVIMED