Integrating Imaging Markers for Clinical Risk Stratification of Takotsubo Syndrome

BACKGROUND: The purpose of this study was to develop machine learning (ML) models that integrate clinical and echocardiographic markers to predict adverse cardiovascular events in patients with Takotsubo syndrome (TTS). METHODS: ML-based prediction models were developed on the basis of 2 retrospective data sets comprising 450 TTS patients (252 in an internal cohort and 198 in an external cohort). Baseline clinical characteristics included demographic characteristics, medical history, triggering factors, symptoms, and vital signs. All echocardiograms were performed within 24 hours of hospital admission. Feature selection and model development were performed using the random forest algorithm for prediction of 3-month major adverse cardiovascular events (MACE). Model performance was evaluated using area under the curve (AUC), calibration, sensitivity, and specificity. Model interpretability was evaluated using SHapley Additive exPlanations. RESULTS: The established ML models, which integrated key clinical and echocardiographic variables, demonstrated good discrimination for 3-month MACE prediction. The machine learning-based clinical-echocardiographic-4 model, which incorporated mitral valve inflow (E wave) deceleration time, systolic blood pressure, right ventricular fractional area change, and heart rate showed an AUC of 0.80 (95% confidence interval, 0.73-0.86) in the internal cohort and an AUC of 0.76 (95% confidence interval, 0.69-0.84) in the external cohort. Calibration analyses showed good agreement between predicted and observed risks internally and acceptable calibration externally. SHapley Additive exPlanations analysis highlighted the dominant contributions of key features to MACE prediction. Restricted cubic spline analyses identified clinically relevant thresholds for elevated MACE risk: deceleration time < 129.9 ms, systolic blood pressure < 125 mm Hg, right ventricular fractional area change < 34.7%, and heart rate > 93 beats per minute. CONCLUSIONS: ML-based clinical-echocardiographic prediction models showed prognostic value in early risk stratification in hospitalized TTS patients.