# Establishment and validation of an electrocardiogram vector-based machine learning model for the conversion of prone position electrocardiograms into standard electrocardiograms

## TL;DR

This paper introduces a machine learning model that converts prone-position electrocardiograms into standard ECGs to improve the detection of heart conditions like STEMI.

## Contribution

A novel machine learning-based approach is developed to convert prone ECGs into standard ECGs, achieving high diagnostic accuracy.

## Key findings

- A hybrid model combining ML and regression approaches achieved strong morphology accuracy and amplitude similarity in ECG conversion.
- The model demonstrated robust diagnostic performance with AUCs ranging from 0.825 to 0.910 for detecting various cardiovascular conditions.
- Converted ECGs were reviewed by cardiologists and showed promising results for identifying STEMI and other CVDs.

## Abstract

The prone electrocardiogram (ECG) presents challenges in detecting anterior ST-segment elevated myocardial infarction (STEMI). This study aims to develop a method to convert prone ECGs to standard ECGs to facilitate physician diagnosis of STEMI and other cardiovascular diseases (CVD).

The standard ECGs, vectorcardiograms (VCGs), and prone ECGs were prospectively examined for model development. Three conversion approaches were developed: direct lead matching by linear regression (Approach 1), conversion from prone ECGs to VCGs via regression and then to standard ECGs (Approach 2), and machine-learning (ML)-based models (Approach 3). External validation was done with a separate cohort, and a hybrid model was created by integrating the best-performing morphology and amplitude models. The diagnostic performance of the converted ECGs was reviewed by nine cardiologists and benchmarked against the original ECG interpretations. Five hundred and ninety prone ECG cardiac cycles from seventy participants [median age 64 years, interquartile range (IQR) 27.0–70.0] were analysed for model development. The external validation cohort had 94 patients (median age 56.5 years, IQR 39.3–67.0). Approach 3 had the best morphology accuracy, and Approach 2 had the best amplitude similarity. These two models were combined into a hybrid model. In the external validation dataset, the AUCs (95% confidence intervals) for detecting normal ECGs, anterior ST-segment elevation/depression, old anterior myocardial infarction, and bundle branch blocks were 0.835 (0.734–0.908), 0.825 (0.693–0.923), 0.898 (0.799–0.957), 0.867 (0.622–0.956), and 0.910 (0.714–0.953), respectively.

The successful development of models for converting prone ECGs to standard ECGs demonstrated good and robust diagnostic performance for CVD.

Graphical AbstractConverting prone electrocardiograms (ECGs) to standard ECGs using three approaches. (A) Morphological differences between prone and standard ECG. Prone ECGs show significant morphological differences from standard ECGs, particularly in the precordial leads, making them less reliable for identifying anterior ST-segment elevated myocardial infarction. (B) Electrocardiogram leads placement during prone positioning. Leads V1 and V2 are situated on the sides of the seventh thoracic vertebra. Lead V3 is positioned equidistantly between leads V2 and V4. Lead V6 is placed in the fifth intercostal space at the left midaxillary line, and V4 and V5 are aligned with V6 at the same level. (C) Three approaches for prone ECG conversion into standard ECG: Approach 1: Utilizing multiple linear regression, the standard ECG leads V1, V2, and V4 were constructed by transforming the prone ECG leads V4, V2, and V1, respectively, into leads that exhibit inverted vectors in the horizontal plane. Approach 2: Utilizing multiple linear regression, a conversion model for prone ECGs and standard vectorcardiograms (VCGs) was established, followed by the application of mature transformation methods to convert them into standard ECGs. Approach 3: Machine learning methods were employed to establish models for converting prone ECG leads into standard ECG. (D) Comparisons of the converted standard ECGs and the original ECGs. Approach 3 (random forest model) demonstrated best performance in morphology. The morphology similarity between the converted and original ECGs for P waves, QRS complexes, and T waves in leads V1–V5 was 90.4% (10.1%), 87.4% (3.9%), and 94.9% (4.0%), respectively. (E) The diagnostic performance of the converted ECGs was assessed by cardiologists. The receiver operating characteristic (ROC) areas under the curve with 95% confidence intervals (CI) for normal ECGs, ST-segment elevation or depression in anterior leads, old anterior myocardial infarction, and bundle branch blocks were 0.835 (95% CI: 0.734–0.908), 0.825 (95% CI: 0.693–0.923), 0.898 (95% CI: 0.799–0.957), 0.867 (95% CI: 0.622–0.956) and 0.910 (95% CI: 0.714–0.953), respectively.

Converting prone electrocardiograms (ECGs) to standard ECGs using three approaches. (A) Morphological differences between prone and standard ECG. Prone ECGs show significant morphological differences from standard ECGs, particularly in the precordial leads, making them less reliable for identifying anterior ST-segment elevated myocardial infarction. (B) Electrocardiogram leads placement during prone positioning. Leads V1 and V2 are situated on the sides of the seventh thoracic vertebra. Lead V3 is positioned equidistantly between leads V2 and V4. Lead V6 is placed in the fifth intercostal space at the left midaxillary line, and V4 and V5 are aligned with V6 at the same level. (C) Three approaches for prone ECG conversion into standard ECG: Approach 1: Utilizing multiple linear regression, the standard ECG leads V1, V2, and V4 were constructed by transforming the prone ECG leads V4, V2, and V1, respectively, into leads that exhibit inverted vectors in the horizontal plane. Approach 2: Utilizing multiple linear regression, a conversion model for prone ECGs and standard vectorcardiograms (VCGs) was established, followed by the application of mature transformation methods to convert them into standard ECGs. Approach 3: Machine learning methods were employed to establish models for converting prone ECG leads into standard ECG. (D) Comparisons of the converted standard ECGs and the original ECGs. Approach 3 (random forest model) demonstrated best performance in morphology. The morphology similarity between the converted and original ECGs for P waves, QRS complexes, and T waves in leads V1–V5 was 90.4% (10.1%), 87.4% (3.9%), and 94.9% (4.0%), respectively. (E) The diagnostic performance of the converted ECGs was assessed by cardiologists. The receiver operating characteristic (ROC) areas under the curve with 95% confidence intervals (CI) for normal ECGs, ST-segment elevation or depression in anterior leads, old anterior myocardial infarction, and bundle branch blocks were 0.835 (95% CI: 0.734–0.908), 0.825 (95% CI: 0.693–0.923), 0.898 (95% CI: 0.799–0.957), 0.867 (95% CI: 0.622–0.956) and 0.910 (95% CI: 0.714–0.953), respectively.