Optimizing embedding spaces for recurrence-based analysis of multivariate cardiac signals: Myocardial infarction detection from 12-lead ECG data

Myocardial infarction (MI) remains one of the leading global causes of mortality, requiring accurate diagnostic methods. Electrocardiogram (ECG) signals in the embedded phase space (PS) have shown great potential for detecting cardiac abnormalities, but selecting embedding parameters (delay and dimension) remains debated. This study evaluates the impact of these parameters on classifying ECG signals from healthy individuals and MI patients. A k-nearest neighbors (KNN) model was trained on recurrence quantification analysis (RQA) attributes from two PS reconstruction methods: one using optimized embedding parameters per signal and another with fixed parameters. The study assessed MI detection across affected cardiac regions and tested different normalization techniques. Results showed that fixed embedding outperformed variable parameters for binary MI detection (ROC AUC: 0.815 ± 0.1; F1-score: 0.726 ± 0.1), though significance was marginal. In multiclass classification, fixed embedding with Robust Scaler excelled in one-vs-rest classification (ROC AUC: 0.746 vs. 0.578, p < 0 . 01 ) and distinguishing anterior vs. inferolateral infarctions (median ROC AUC: 0.909). Min-Max normalization proved the most effective, balancing feature scaling and outlier resilience. The use of fixed embedding parameters, after adjustment based on most frequent values, improved the computational efficiency by eliminating per-signal optimization. However, differentiating adjacent infarction regions (e.g., anterior vs. anterolateral, ROC AUC: 0.426 ± 0.17) remains challenging, requiring advanced feature engineering and multimodal integration. This study underscores the viability of fixed embedding parameters for PS reconstruction combined with RQA for MI detection, supporting the development of deployable ECG-based diagnostic tools.