Takotsubo syndrome (TTS), initially considered relatively benign, has emerged as a potentially life-threatening condition with impaired short- and long-term prognosis, comparable to acute coronary syndrome in terms of in-hospital complications and long-term major adverse cardiac and cerebrovascular events (MACCE). Accurate prediction of TTS patients' clinical course is crucial for clinical decision-making and prognostic assessment. Existing risk scores, like the German and Italian Stress Cardiomyopathy (GEIST) score, demonstrate moderate accuracy (AUC around 0.70) in predicting in-hospital complications. Artificial intelligence (AI) and machine learning (ML) offer a promising alternative, often outperforming traditional risk scores in various cardiovascular applications. This study aimed to develop an ML-based model using common clinical variables to predict in-hospital death risk in the largest TTS registry. A secondary aim was to cluster TTS patients into distinct risk profiles based on relevant variables, facilitating personalized patient management.

The literature highlights the significant adverse event rate associated with TTS, necessitating improved risk stratification tools. While existing risk scores, such as the GEIST score, provide some predictive value, their accuracy is limited. The emergence of AI and ML in medical risk prediction has shown superior performance compared to traditional methods in various cardiovascular settings, including predicting mortality in suspected coronary artery disease and estimating ischaemic and bleeding risk in acute coronary syndrome. This forms the basis for the current study's hypothesis that an ML-based approach can improve risk prediction for in-hospital death in TTS patients.

This study utilized data from two datasets: the International Takotsubo (InterTAK) Registry (train and internal validation cohorts) and the Takotsubo Italian Network (external validation cohort). The InterTAK Registry, a prospective and retrospective observational registry, included patients from 2011 to 2021 across 58 cardiovascular centers in 17 countries. Patients were included based on InterTAK diagnostic criteria. Data included demographics, triggering factors, cardiovascular risk factors, haemodynamic and angiographic findings, electrocardiography and echocardiography parameters, laboratory values, medications, in-hospital complications, and management. Ethical approval was obtained from relevant committees. In-hospital all-cause death was the primary outcome, with ROC-AUC, sensitivity, and specificity as primary endpoints. Secondary endpoints included patient clustering into phenotypic groups based on risk of in-hospital death. The InterTAK Registry dataset (initially 3703 patients) underwent cleaning, removing discharge-related and proxy death features, and variables/records with >30% missing values. Missing values were imputed using median (continuous) and mode (categorical) during cross-validation. 31 clinically relevant variables were selected for the prediction model. A ridge penalized logistic regression (PLR)-based ML model and a K-medoids clustering algorithm were employed. The InterTAK cohort was split into train (75%) and internal validation (25%) sets, considering the imbalance towards survivors. An ensemble of 100 optimal PLRs was built through repeated five-fold cross-validation. Model performance was assessed on the internal and external validation sets and compared with existing scores (InterTAK 'traditional' model and GEIST score). A classical logistic regression model was also developed for comparison. For clustering, K-medoids was implemented using the 10 top predictive features from the PLR model and the TTS triggering factor. The optimal number of clusters was determined using the elbow method and silhouette coefficient. Sensitivity analyses were performed by sex, presence of coronary artery disease, and cardiogenic shock. Python was used for all analyses, and an online calculator for the InterTAK-ML model was developed.

Of 3482 patients, 183 (5.2%) died during hospitalization. Cardiovascular deaths accounted for 45.9%, non-cardiovascular for 46.4%, and unknown causes for 7.6%. In-hospital death was significantly associated with several factors (Table 1): lower prevalence of female sex, hypertension, and hyperlipidemia; more frequent physical triggers; higher prevalence of coexisting coronary artery disease, cardiogenic shock (CS), lower systolic blood pressure, higher heart rate, lower LVEF, increased need for catecholamine support, atrial fibrillation, ST-segment elevation, and higher white blood cell (WBC) count. The InterTAK-ML model showed excellent discriminative ability for predicting in-hospital death, with an AUC of 0.89 (0.85-0.92), sensitivity of 0.85 (0.78-0.95), and specificity of 0.76 (0.74-0.79) in the internal validation cohort, and an AUC of 0.82 (0.73-0.91) in the external validation cohort (Figure 3). A simplified model using the top 10 variables maintained good accuracy (AUC 0.88 internal, 0.83 external). The ML model significantly outperformed a standard logistic regression model, especially in sensitivity. K-medoids clustering identified six distinct phenotypic clusters (Table 2), each with varying risk of in-hospital death (Figure 4). Clusters with higher mortality rates were characterized by physical triggers, lower LVEF, lower diastolic blood pressure, higher heart rate, higher WBC count, cardiogenic shock, and acute neurological disorders. The clustering analysis remained consistent across subgroups (male, no coronary artery disease, no cardiogenic shock) and the external validation cohort.

This study demonstrates the significant in-hospital mortality rate (5.2%) associated with TTS, highlighting the need for improved risk stratification. The developed InterTAK-ML model, based on readily available clinical variables, provides superior predictive performance compared to existing scores. The identified top 10 predictive features, including LVEF, WBC count, cardiogenic shock, diastolic blood pressure, physical stress, acute neurological disorder, age, heart rate, atrial fibrillation, and asthma, are clinically relevant and have strong pathophysiological rationales. Notably, the model highlights the prognostic significance of sympathetic activation, reflected in increased heart rate independent of cardiogenic shock. The clustering analysis revealed six distinct TTS phenotypes with varying mortality risks, emphasizing the heterogeneity of the syndrome. The results support the use of ML for personalized risk assessment and management of TTS. The study underscores the value of AI-driven approaches for enhancing clinical decision-making and optimizing quality of care in managing TTS.

The InterTAK-ML model offers a robust and accurate tool for predicting in-hospital mortality in TTS patients, outperforming existing methods. The identification of six distinct patient clusters with varying risks further enhances risk stratification and facilitates personalized treatment strategies. The model's user-friendly online calculator promotes practical implementation. Future research should focus on validating the model in diverse populations and exploring the interaction between identified variables to deepen our understanding of TTS pathophysiology.

The observational and partly retrospective design of the InterTAK Registry introduces limitations. Limited ethnic diversity in the dataset may affect the generalizability of the model. The selection of the top 10 variables for clustering was somewhat arbitrary. Exclusion of patients on mechanical support and the lack of comprehensive ethnic data are further limitations. The timing of TTS onset relative to other events was incomplete. Further validation in an external, more diverse cohort is needed to confirm the model's generalizability.