# Enhancing tertiary cardiology triage with vectorcardiographic features: a machine learning approach using real-world data

## TL;DR

A machine learning model using vectorcardiographic features improves triage decisions for patients needing specialized cardiology care, with high sensitivity but low specificity.

## Contribution

The study introduces a novel machine learning model integrating vectorcardiographic-derived GEH features for tertiary cardiology triage using real-world data.

## Key findings

- The model achieved 94% sensitivity and 68% AUC in predicting tertiary care needs.
- GEH parameters like QRST angle and SVG magnitude were statistically significant predictors.
- SHAP analysis confirmed the model's alignment with clinical risk factors and GEH features.

## Abstract

•ML model uses VCG-derived GEH to predict tertiary care (AUC 68 %, Sens 94 %) for triage.•Interpretable XGBoost prioritizes GEH and prior PCI, matching clinical risk factors.•High sensitivity and low specificity highlight its use as a screening support tool.•SHAP confirms robustness of SVG, QTc, and age across several sensitivity scenarios.•VCG biomarker extraction enables low-cost, scalable support for clinical triage use.

ML model uses VCG-derived GEH to predict tertiary care (AUC 68 %, Sens 94 %) for triage.

Interpretable XGBoost prioritizes GEH and prior PCI, matching clinical risk factors.

High sensitivity and low specificity highlight its use as a screening support tool.

SHAP confirms robustness of SVG, QTc, and age across several sensitivity scenarios.

VCG biomarker extraction enables low-cost, scalable support for clinical triage use.

To assess whether electrocardiographic markers of Global Electrical Heterogeneity (GEH) improve the identification of patients requiring tertiary care, either alone or combined with an explainable machine learning model, compared with standard ECG features and clinical risk factors in a real-world tertiary cardiology population.

Patients were forwarded to a specific evaluation in a cardiology-specialized hospital performed an ECG and data collection. A series of follow-up attendances occurred in periods of 6-months, 12-months and 15-months to check for cardiovascular-related events (mortality or new nonfatal cardiovascular events (Stroke, MI, PCI, CS), as identified during 1-year phone follow-ups. The first attendance ECG was measured by a specialist and processed in order to obtain the Global Electric Heterogeneity (GEH) using the Kors Matriz. The ECG measurements, GEH parameters, and risk factors were combined for training multiple instances of XGBoost decision tree models. Each instance was optimized for the AUCPR, and the instance with the highest AUC was chosen as representative of the model. The importance of each parameter for the winner tree model was compared to better understand the improvement from using GEH parameters.

GEH parameters were statistically significant in this population (p < 0.001), particularly the QRST angle and SVG magnitude. The combined model integrating GEH, standard ECG features, and clinical risk factors achieved the best performance, with a sensitivity of 94.1 %, specificity of 30.8 %, AUC of 67.6 %, and F2 score of 0.62. SVG feature importance and SHAP analyses were consistent with the statistical findings, indicating that the model's decision patterns align with clinically relevant information and reinforce the role of GEH features. The modeling approach was carefully designed to prevent overfitting, ensure generalizability, and facilitate implementation through its decision tree architecture.

VCG-derived features may improve the identification of patients requiring tertiary care, either alone or integrated into an explainable and robust machine learning model trained on real-world data. Its clinical value will ultimately depend on prospective validation and seamless integration within existing care pathways.