Implications of Spatiotemporal Deep Learning Neural Networks in Echocardiographic Diagnosis and Prognostication of Takotsubo Syndrome - Project Summary/Abstract Differentiating Takotsubo syndrome (TTS) from acute myocardial infarction (AMI) is often challenging in real time. Meanwhile, the poor phenotypic grouping of TTS patients prevents investigation of effective therapeutic strategies for long-term risk reduction. Built on our current study that already yielded an accurate and broadly applicable spatiotemporal deep convolution neural network (DCNN) for echocardiographic diagnosis of TTS, in this proposal, we plan to augment and optimize a DeSeg-TTSD algorithm using large-scale multi-institution and multi-vendor echocardiographic datasets, as well as clinical metadata to increase the robustness, generalizability, and interpretability of deep learning modeling in real-time imaging analysis. We also plan to develop and validate a DeSeg-TTSP model for TTS prognostication and phenomapping by extracting latent spatiotemporal features correlating with TTS pathophysiology and outcome, so as to develop personalized treatment. We hypothesize that, when trained on an echocardiographic video task, a spatiotemporal DCNN can unlock sub-visual predictive information with advanced learning and computational analysis, to discover distinctive myocardial motion patterns and assimilate latent spatiotemporal imaging features to improve the accuracy of diagnosis and prognostication for TTS patients. The proposed research project brings together multiple innovations: It is based on a solid scientific premise, builds on already achieved results and extends state of art of spatiotemporal deep learning modeling in real time imaging, to provide decision support for clinical diagnosis and prognostication using imaging information routinely available in daily practice. Other than our local research team assembling imaging engineering experts, cardiologists and statisticians, an inter-institutional team including 57 board-certified expert human readers will perform human classification, data visualization and evaluate the feasibility of the application and integration into a clinical setting of the established and validated DL model. We will fulfill the following specific aims 1: Develop and validate fully automated spatiotemporal DL diagnosis models from echocardiographic videos that enable discrimination of TTS from AMI in real time. 2: Integrate spatiotemporal DL prognostication into clinical prediction models to endorse long-term TTS prognostication. 3: Perform data visualization on spatiotemporal DL framework and investigate TTS pathophysiology to develop personalized treatment Establishing this spatial-temporal hybridized deep learning framework will become a foundation for the development of additional precision medicine decision support tools for patients with acute cardiovascular disorders to address urgently-needed diagnostic decisions, resolve time-sensitive therapeutic dilemmas and obtain advanced imaging markers to develop specific primary and secondary prevention strategies.
