Machine learning-based prediction of in-hospital death for patients with takotsubo syndrome: The InterTAK-ML model

The study addresses the need for accurate short-term prognostic assessment in takotsubo syndrome (TTS), a condition once considered benign but now recognized to carry significant in-hospital complications and long-term adverse events similar to acute coronary syndromes. Existing risk tools, such as the GEIST score derived via stepwise logistic regression, demonstrate only moderate accuracy. The primary aim was to develop a machine learning (ML)-based model integrating routinely available clinical variables to predict in-hospital all-cause death in TTS patients from the International Takotsubo (InterTAK) Registry. A secondary aim was to perform data-driven clustering to identify phenotypic groups with differing in-hospital mortality risk to support personalized management.

Prior work includes the GEIST score for in-hospital complications with AUC around 0.70 and the InterTAK prognostic score derived from classical statistical methodologies. Other expert-opinion-based tools have been proposed but lacked rigorous derivation and validation. Artificial intelligence and ML approaches have outperformed traditional scores in various cardiovascular contexts, including mortality prediction in suspected coronary artery disease and post-acute coronary syndrome adverse event prediction. Previous registry analyses identified factors associated with poor outcomes in TTS (e.g., cardiogenic shock, physical triggers, neurological disorders), but comprehensive, externally validated ML models for in-hospital death in TTS had not been reported.

Design and data sources: The study utilized the International Takotsubo (InterTAK) Registry (observational, prospective/retrospective, 58 centers in 17 countries; patients enrolled 2011–2021) for model derivation and internal validation, and the Takotsubo Italian Network (TIN; 2007–2018; centers overlapping with InterTAK excluded) for external validation. Inclusion followed InterTAK diagnostic criteria. Ethics approvals were obtained locally. Outcome: In-hospital all-cause death. Data preprocessing and feature selection: Starting from 3703 InterTAK patients, discharge-related and proxy-for-death variables were removed. Variables and records with >30% missingness were excluded. Missing values were imputed within cross-validation (median for continuous, mode for categorical). A set of 31 clinically relevant, readily obtainable variables were selected considering collinearity. LVEF was taken from angiography when available, otherwise echocardiography. After cleaning, 3482 records remained. Modeling: A ridge penalized logistic regression (PLR) ensemble was developed. The dataset was split into training (75%) and internal validation (25%) sets with attention to class imbalance. Five-fold cross-validation on the training set was repeated 100 times with different randomizations, yielding 100 optimal PLR models. The final InterTAK-ML model is a consensus/ensemble of these 100 models. Performance metrics (ROC-AUC, sensitivity, specificity) were estimated with 1000 bootstrap resamples. Comparators: The InterTAK traditional model and, in the external cohort, the GEIST score. A standard logistic regression using the same variables was also trained for comparison in the internal validation set. Simplified model: A reduced InterTAK-ML using only the top 10 predictive features was evaluated. Clustering: To phenotype patients, K-medoids clustering (PAM algorithm) using Gower distance was applied to the top 10 predictive features plus triggering factor (physical, emotional, both, none). Candidate cluster numbers (2–10) were assessed using the elbow method (explained variance) and silhouette coefficient; six clusters were selected. Dimensionality reduction via t-SNE was used for visualization. Subgroup and sensitivity analyses: Model performance and cluster distributions were examined by sex, presence of coronary artery disease, and presence of cardiogenic shock on admission. Implementation: An online calculator is available (https://compbiomed.hpc4ai.unito.it/intertako/).

- Cohort and outcome: Among 3482 InterTAK patients, in-hospital death occurred in 183 (5.2%); 3299 (94.8%) were discharged alive. Causes of death (n=169 known): 45.9% cardiovascular, 46.4% non-cardiovascular, 7.6% unknown. Patients who died had worse hemodynamics on admission, including higher prevalence of cardiogenic shock (45.8% vs 6.3%), lower systolic BP (120.2 ± 33.3 vs 131.1 ± 29.2 mmHg), higher heart rate (99 ± 26 vs 87 ± 21 bpm), lower LVEF (33.1 ± 11.1% vs 41.0 ± 11.9%), more frequent atrial fibrillation (18.2% vs 5.8%), and higher WBC counts (14.1 ± 7.4 vs 10.6 ± 4.9 ×10^3/μL). Physical triggers were more common in fatal cases (82.0% vs 39.3%). - Feature importance: The 10 most predictive variables were (direction as per PLR coefficients): LVEF (protective), WBC count (risk), cardiogenic shock (risk), diastolic BP (protective), physical trigger (risk), acute neurological disorder (risk), age (risk), heart rate (risk), atrial fibrillation at presentation (risk), and asthma (direction reported among top 10). - Model performance: InterTAK-ML (full 31-variable model) achieved internal validation ROC-AUC 0.89 (95% CI 0.85–0.92), sensitivity 0.85 (0.78–0.95), specificity 0.76 (0.74–0.79); external validation ROC-AUC 0.82 (0.73–0.91) with sensitivity 0.74 (0.61–0.87). The simplified 10-variable InterTAK-ML reached internal AUC 0.88 (0.85–0.91), sensitivity 0.83 (0.74–0.91), specificity 0.75 (0.73–0.78); external AUC 0.83 (0.74–0.91), sensitivity 0.70 (0.52–0.87), specificity 0.80 (0.78–0.83). A standard logistic regression showed AUC 0.84 (0.79–0.90) and specificity 0.99 (0.98–1.00) but very low sensitivity 0.09 (0.02–0.15). - External cohort: In TIN, in-hospital mortality was 2.2% (n=23), with similar associations between poor outcomes and physical triggers, acute neurological disorders, and worse admission hemodynamics. - Clustering: Six phenotypic clusters were identified with distinct observed in-hospital death rates: cluster 1 (0.5%), cluster 2 (0.8%), cluster 3 (1.7%), cluster 4 (5.4%), cluster 5 (15.5%), cluster 6 (28.8%). Physical triggers concentrated in clusters with intermediate to high mortality (clusters 4–6). Cluster 6 had universal cardiogenic shock; clusters 1–3 had low shock prevalence (0–3%) and low mortality, often associated with emotional or no identifiable triggers.

The InterTAK-ML model demonstrates that ML applied to routine clinical variables can provide high discrimination for in-hospital mortality in TTS, outperforming traditional scores and standard logistic regression particularly in sensitivity—critical for identifying high-risk patients in a relatively low-event setting. The model’s top predictors have strong pathophysiologic plausibility: low LVEF reflects more extensive myocardial stunning; cardiogenic shock, lower diastolic BP, higher heart rate, and atrial fibrillation reflect hemodynamic compromise and sympathetic activation; elevated WBC suggests an inflammatory component; physical triggers and acute neurological disorders mark more severe systemic stress; age modulates risk; asthma may contribute via respiratory compromise and potential medication effects. The six-cluster phenotyping captures non-linear interactions among these variables, offering a layered, patient-centered risk context beyond a single probability estimate. Clusters dominated by physical triggers and markers of hemodynamic compromise and inflammation carry substantially higher mortality, while emotional or absent triggers with stable hemodynamics correspond to low risk. These insights can guide triage, monitoring intensity, and communication of risk, and they suggest mechanistic avenues—sympathetic activation and inflammation—as targets for future research and therapeutic strategies.

An ML-based tool (InterTAK-ML) was developed and externally validated to predict in-hospital mortality in TTS, surpassing existing traditional scores. Using readily available clinical variables, it achieves high discrimination and identifies six phenotypic clusters with distinct risk profiles, enabling both individualized risk prediction and contextualized phenotyping. All identified predictors are pathophysiologically grounded, including heart rate as a novel prognostic marker independent of cardiogenic shock. The freely available online calculator facilitates clinical implementation. Further external validation is warranted and may support broader adoption and refinement, including exploration of long-term outcomes and integration of additional data modalities.

- Registry design: InterTAK is observational and partly retrospective, with inherent biases. - Ethnicity data were limited; the cohort is predominantly European with an Asian minority, potentially limiting generalizability to other populations. - Clustering used only the top 10 predictive variables plus trigger for practicality, potentially omitting other relevant features (e.g., ST-segment elevation, troponin). - Exclusion of mechanical ventilation and mechanical circulatory support aimed to reduce confounding and collinearity with cardiogenic shock but may limit applicability to the sickest patients. - Timing of TTS diagnosis relative to triggers was generally unavailable (except in acute neurological disorders), precluding assessment of temporal effects. - Methodological choices: K-medoids (PAM) was chosen for robustness and compatibility with Gower distance; alternative clustering approaches might yield different partitions. - Focus on short-term (in-hospital) mortality by design; long-term prognostication was not addressed.