Machine-learning differentiation of precordial T-wave inversion in pulmonary embolism and acute coronary syndrome

Differentiating acute coronary syndrome (ACS) and acute pulmonary embolism (PE) in emergency settings is a notorious diagnostic challenge due to their overlapping symptoms and shared electrocardiographic abnormalities, such as precordial T-wave inversion (TWI). This study evaluated the ability of machine learning (ML) algorithms to distinguish acute PE from ACS by analyzing the morphology of inverted T waves in precordial leads.

This retrospective, case-control study included 200 patients (100 ACS confirmed by coronary angiography; 100 PE confirmed by computed tomography pulmonary angiography) admitted to Shahid Madani Hospital, Tabriz, Iran (2023-2024), who exhibited precordial TWI. Image processing with a region-growing technique traced negative T waves in leads V1, V2, and V3. Signals were filtered using Savitzky-Golay and Empirical Wavelet Transform (EWT) methods to extract time-frequency statistical parameters. Multiple supervised ML models were trained (80% training, 20% test with 5-fold cross-validation) over 50 random iterations.

The Ensemble (AdaBoost) model achieved the highest diagnostic performance, yielding an accuracy of 84.44%, sensitivity of 83.96%, specificity of 84.87%, and F1-score of 83.56%. Neural Network models performed comparably, achieving 84.00% accuracy and 85.14% sensitivity. Other evaluated algorithms (SVM, KNN, Naive Bayes) also generated acceptable discriminative accuracy (ranging from 78% to 82%).

Machine-learning algorithms can extract potentially disease-specific features from precordial T-wave morphology to differentiate PE from ACS. These findings support further investigation of AI-enhanced ECG interpretation as a rapid, low-cost, non-invasive adjunct to clinical decision-making in acute care. External validation in larger, clinically representative populations is required before clinical application.