Machine learning ECG model discriminated takotsubo syndrome from myocardial infarction with ROC AUC 0.85-0.87, sensitivity 0.60-0.90, and NPV near 0.99, outperforming cardiologists.
Does a machine learning-based ECG model improve the discrimination of takotsubo syndrome from myocardial infarction compared to cardiologists using conventional criteria in patients with suspected MI?
A machine learning model applied to admission 12-lead ECGs can differentiate takotsubo syndrome from myocardial infarction with high negative predictive value, outperforming conventional ECG criteria used by cardiologists.
Absolute Event Rate: 0% vs 0%
Abstract Background Numerous electrocardiographic (ECG) criteria have been evaluated, with relatively high specificity but low sensitivity, in diagnosing takotsubo syndrome (TS), mainly differentiating it from myocardial infarction (MI). Machine learning (ML) algorithms applied to the ECG have been successful in several cardiac diagnoses, however rarely been used for the diagnostics of TS. Purpose Our aim was to develop a ML-based ECG-model to discriminate TS among patients with a suspected MI based on a 12-lead ECG at admission. Methods A cross-sectional study in a city, using a neural network with UNet architecture. The network was trained and validated on 507 TS cases and 14,978 controls with suspected MI, identified from the Swedish coronary angiography and angioplasty register. Cross-validation in 10 splits was performed. The model was compared with cardiologists using previously proposed ECG criteria. Results The receiver operating curve (ROC) are under the curve (AUC) for discriminating TS against any patients with suspected MI was 0.85 (cross validation: 0.81-0.91) with sensitivity (0.60-0.90) and specificity (0.59-0.89) with low positive predictive value (PPV) (0.06-0.16) and high negative predictive values (NPV) (0.99). The ROC AUC for discriminating TS from verified MI was 0.87 (0.82-0.91) with sensitivity (0.60-0.90) and specificity (0.59-0.89) with low PPV (0.16-0.33) and high NPV (0.97-0.99). Results for ST-elevation and non-ST-elevation MI were even better than for verified MI. The model was compared against a committee of two cardiologists. The best combination of ECG criteria was prolonged QTc combined with ST-elevations in both anterior leads and -aVR. The best performing model achieved an ROC AUC of 0.71. Conclusion In this study, ML could discriminate TS against suspected and verified MI with high sensitivity and a NPV of almost one, outperforming cardiologists using conventional criteria. The model requires further refinement to increase PPV, precision-recall and external validation, but it holds promise for TS screening in the acute setting aiding the clinician in ruling out TS. The results point towards the need for a more complex model, possibly combined with other diagnostic tools such as biomarkers and echocardiography, to increase the specificity and PPV.
Hakansson et al. (Sat,) reported a other. Machine learning ECG model discriminated takotsubo syndrome from myocardial infarction with ROC AUC 0.85-0.87, sensitivity 0.60-0.90, and NPV near 0.99, outperforming cardiologists.