Predicting takotsubo syndrome subtypes: An interpretable machine learning model for differentiating emotional versus physical aetiologies

Abstract

Background

Takotsubo syndrome (TTS) is an acute coronary syndrome characterized by a reversible, mostly apical dysfunction of the left ventricle. Based on the triggering event, TTS has been classified as primary due to emotional causes and secondary due to physical stress. Using a comprehensive machine-learning approach, we aimed to distinguish between these two types of TTS, an essential task for optimizing patient care.

Methods

Based on a dataset of 320 TTS patients from a research group in Rome, a logistic regression model was trained to develop an interpretable predictive model capable of accurately classifying the aetiology of TTS in individual patients using admission-based clinical markers.

Results

The developed model achieved 74 % accuracy, 75 % precision and recall, 72 % specificity, and an area under the curve (AUC) of 0.78. Based on the studies conducted, chest pain, dyspnoea, atrial fibrillation, sex, chronic obstructive pulmonary disease, heart rate, and cancer were identified as key clinical features for differentiating between the two TTS types. An external validation cohort of 121 TTS patients has been employed further to assess the performance of the trained classification model, obtaining 74 % accuracy, 77 % precision, 91 % recall, 27 % specificity, and an AUC of 0.62.

Conclusions

An interpretable machine learning model has been developed, demonstrating the ability to distinguish between emotional versus physical aetiologies in TTS, highlighting the most impactful clinical factors. As built considering clinical variables recorded at admission, the model may serve as an immediate tool that can guide clinicians in their practice.