Prognostic value of texture analysis from cardiac magnetic resonance imaging in patients with Takotsubo syndrome: a machine learning based proof-of-principle approach

• Main Authors: Manoj Mannil, Ken Kato, Robert Manka, Jochen von Spiczak, Benjamin Peters, Victoria L. Cammann, Christoph Kaiser, Stefan Osswald, Thanh Ha Nguyen, John D. Horowitz, Hugo A. Katus, Frank Ruschitzka, Jelena R. Ghadri, Hatem Alkadhi, Christian Templin
• Format: Article
• Language: English
• Published: Nature Publishing Group 2020-11-01
• Series: Scientific Reports
• Online Access: https://doi.org/10.1038/s41598-020-76432-4

Abstract Cardiac magnetic resonance (CMR) imaging has become an important technique for non-invasive diagnosis of takotsubo syndrome (TTS). The long-term prognostic value of CMR imaging in TTS has not been fully elucidated yet. This study sought to evaluate the prognostic value of texture analysis (TA) based on CMR images in patients with TTS using machine learning. In this multicenter study (InterTAK Registry), we investigated CMR imaging data of 58 patients (56 women, mean age 68 ± 12 years) with TTS. CMR imaging was performed in the acute to subacute phase (median time after symptom onset 4 days) of TTS. TA of the left ventricle was performed using free-hand regions-of-interest in short axis late gadolinium-enhanced and on T2-weighted (T2w) images. A total of 608 TA features adding the parameters age, gender, and body mass index were included. Dimension reduction was performed removing TA features with poor intra-class correlation coefficients (ICC ≤ 0.6) and those being redundant (correlation matrix with Pearson correlation coefficient r > 0.8). Five common machine-learning classifiers (artificial neural network Multilayer Perceptron, decision tree J48, NaïveBayes, RandomForest, and Sequential Minimal Optimization) with tenfold cross-validation were applied to assess 5-year outcome including major adverse cardiac and cerebrovascular events (MACCE). Dimension reduction yielded 10 TA features carrying prognostic information, which were all based on T2w images. The NaïveBayes machine learning classifier showed overall best performance with a sensitivity of 82.9% (confidence interval (CI) 80–86.2), specificity of 83.7% (CI 75.7–92), and an area-under-the receiver operating characteristics curve of 0.88 (CI 0.83–0.92). This proof-of-principle study is the first to identify unique T2w-derived TA features that predict long-term outcome in patients with TTS. These features might serve as imaging prognostic biomarkers in TTS patients.