This study explores the feasibility of predicting echocardiographic (EchoCG) parameters from electrocardiogram (ECG) data using machine learning techniques. Two modeling approaches are investigated: regression for continuous parameter prediction and multi-class classification for clinically significant parameter ranges. A dataset of 37 patients with matched ECG and EchoCG data is used. Strong correlations between selected parameter pairs are identified. Results demonstrate that ensemble models such as Random Forest outperform linear models in most prediction tasks. Limitations due to data imbalance and potential improvements using balancing techniques are also discussed.
Electrocardiography (ECG) and transthoracic echocardiography (EchoCG) are two fundamental diagnostic tools in cardiology. While ECG provides information on the electrical activity of the heart, EchoCG ofers insights into its mechanical and structural function. These modalities are often used together in clinical settings to diagnose and monitor cardiovascular diseases.
In recent years, machine learning (ML) has demonstrated substantial potential in processing and
interpreting ECG data, enabling automatic detection of arrhythmias, structural abnormalities, and even
prediction of patient outcomes such as mortality [
Several recent eforts suggest that non-invasive ECG data may contain enough information to infer
certain echocardiographic abnormalities, especially when leveraged through advanced ML techniques
[
At the same time, researchers acknowledge significant barriers to ML application in clinical
cardiology [
In this paper, we explore the feasibility of predicting a wide set of echocardiographic parameters from ECG-derived features using machine learning techniques. We utilize a dataset consisting of 37 military persons, free of heart disease, with matched ECG and EchoCG measurements, comprising 172 ECG parameters and 134 EchoCG parameters. Our approach includes: • Performing Pearson correlation analysis to identify strongly associated ECG–EchoCG parameter pairs. • Training regression models—linear regression and Random Forest—to predict quantitative EchoCG values from ECG data. • Formulating a multi-class classification problem based on clinically meaningful ranges of selected
EchoCG parameters (e.g., LV dimensions, LVEF). • Evaluating classification performance and analyzing limitations due to data imbalance, with suggestions for addressing them through oversampling (e.g., SMOTE), cost-sensitive learning, and ensemble models.
This research aims to evaluate the potential of ECG-based prediction models as a non-invasive tool for estimating echocardiographic measurements. The results provide insight into the correlation between electrical and mechanical cardiac markers and lay the groundwork for intelligent clinical decision support systems.
The dataset used in this study comprises paired records of electrocardiographic (ECG) and transthoracic echocardiographic (EchoCG) parameters collected from a cohort of 43 patients. After preprocessing, only 37 patients had both valid ECG and EchoCG records and were included in the analysis.
In total, 126 unique ECG records and 64 unique EchoCG records were available. Among the final dataset, 112 ECG records and 64 EchoCG records corresponded to the 37 patients with complete data.
The ECG dataset initially contained 189 features, while the EchoCG dataset had 124. After filtering, 172 ECG parameters and 134 EchoCG parameters were retained for analysis.
Several preprocessing steps were performed: • Parameter exclusion: Non-informative fields such as patient identifiers, timestamps, and demographic attributes (e.g., gender, birthdate) were excluded. Additionally, features with only one unique value were removed. • Missing data handling: Features with missing values in more than 10% of patients were excluded.
Remaining missing values were imputed using feature-wise means. • Encoding and normalization: Categorical variables were binary-encoded. All numeric features were normalized to zero mean and unit variance. • Aggregation: When multiple records per patient were available, measurements were averaged to form a unified feature vector per patient.
After these steps, the dataset included 163 ECG and 109 EchoCG parameters for each of the 37 patients.
To evaluate the relationship between ECG and EchoCG features, Pearson correlation coeficients were computed for all pairwise combinations. The Pearson correlation coeficient between variables and is defined as: = √︀∑︀ =1(−¯) ∑︀=1(−¯)( −)¯ 2√︀∑︀ =1(−)¯ 2 where¯ and¯ are the sample means of and , respectively.
Several pairs demonstrated strong correlations ( > 0.90), such as: (1) (2) (3) (4)
These findings provided insights into potential functional and structural relationships between electrical and mechanical cardiac properties.
To predict continuous EchoCG parameters from ECG data, two regression models were trained: • Linear Regression: Assumes linear dependence between ECG features and each EchoCG parameter. • Random Forest Regressor: An ensemble of 200 decision trees trained using bootstrapped samples and random feature selection at each split.
Prior to training, features were normalized. Data was split into training (80%) and testing (20%) subsets using a random shufle split.
Performance was evaluated using the following metrics: • ECG parameter: Q/R amplitude ratio (lead AvF) and EchoCG parameter: Left ventricular systolic sphericity index ( = 0.99), • ECG parameter: T-wave symmetry (lead I) and EchoCG parameter: Mitral valve score ( = 0.99).
MAE = 1 ∑︁ | − ˆ |
=1 ⎯ RMSE = ⎷⎸⎸ 1 ∑=︁1 ( − ˆ )2 2 = 1 − ∑︀=1( − ˆ )2 ∑︀=1(−)¯ 2
Measures the average magnitude of errors between predicted and actual values:
Emphasizes larger errors more than MAE:
Represents the proportion of variance in the target variable explained by the model: where¯ is the mean of the observed values.
For several clinically important EchoCG parameters (e.g., LV end-diastolic diameter, LV end-systolic volume, LA diameter, LVEF), value ranges were discretized into 4–5 categorical classes reflecting clinical thresholds. This formulation transformed the prediction task into multi-class classification.
Due to class imbalance (e.g., most patients concentrated in one class), the analysis focused on the parameter with the most balanced class distribution: Left Atrial Anteroposterior Dimension. Two-class classification was performed.
A Random Forest classifier with 100 trees was used. Key settings included: • Class balancing using inverse frequency weighting, • Stratified train/test split to preserve class proportions, • Fixed random seed for reproducibility.
Classification performance was evaluated using: The harmonic mean of precision and recall, particularly useful for imbalanced datasets: Accuracy =
Number of correct predictions
Total number of predictions
Precision · Recall F1 = 2 · Precision + Recall (5) (6)
Where: TP • Precision = TP + FP TP • Recall = TP + FN
Here, TP, FP, and FN denote true positives, false positives, and false negatives, respectively. The regression task aimed to predict continuous echocardiographic (EchoCG) parameters using electrocardiographic (ECG) features. Two models were evaluated: linear regression and random forest regression. The performance was assessed on the test set using MAE, RMSE, and 2 score.
Linear regression was able to capture linear relationships between ECG and EchoCG features. However, only a limited subset of parameters achieved satisfactory performance. Table 1 presents the top EchoCG parameters predicted with the highest 2 scores.
Random forest models outperformed linear regression across most predicted parameters due to their ability to capture non-linear relationships. Table 2 summarizes the best-performing predictions based on 2 scores.
The classification experiment targeted the prediction of clinically meaningful ranges of selected EchoCG parameters. Due to data imbalance, we focused on the binary classification of the parameter Left Atrial Anteroposterior Dimension (LAAPD), which had the most balanced class distribution.
The random forest classifier achieved the following results: • Accuracy: 78% • F1 Score (majority class): 0.88 • F1 Score (minority class): 0.00
Despite attempts to balance the dataset using class weights and stratified sampling, the model failed to correctly classify any minority class instances. This result highlights the dificulty of applying standard classifiers on highly imbalanced clinical datasets.
Pearson correlation analysis identified several ECG–EchoCG parameter pairs with strong linear relationships ( > 0.9), suggesting high predictive potential. Notable examples include: • ECG: Q/R amplitude ratio (lead AvF) ↔ EchoCG: L systolic sphericity index ( = 0.99) • ECG: T-wave symmetry (lead I) ↔ EchoCG: Mitral valve score ( = 0.99) • ECG: Heart rhythm abnormality score ↔ EchoCG: Tricuspid regurgitation grade ( = 0.95) These correlations validate the feasibility of ECG-driven estimation of certain mechanical heart characteristics.
This study explored the feasibility of predicting echocardiographic parameters using electrocardiographic features via machine learning methods. Both regression and classification tasks were evaluated to assess the potential for non-invasive, ECG-based estimation of EchoCG measurements.
Our regression results show that both linear and non-linear models can predict a subset of EchoCG parameters with reasonable accuracy. However, Random Forest regression consistently outperformed linear regression, especially for parameters with known non-linear relationships to ECG markers.
The highest 2 scores (above 0.80) were achieved for: • KST LeA, potik cherez lehenevu arteriiu (2 = 0.896) • VTLSh, aortalnyi potik (2 = 0.830)
Other EchoCG parameters such as tricuspid regurgitation grade, LV hypertrophy markers, and LV ejection fraction (FVLSh) were also predicted with acceptable performance (2 > 0.6), demonstrating that ECG signals contain information reflective of structural and hemodynamic cardiac states.
Linear regression, while interpretable, was limited in its predictive power for most parameters. It only achieved moderate 2 (around 0.5) for a few features, indicating that non-linear modeling is essential for capturing complex ECG–EchoCG relationships.
In the classification setting, EchoCG parameters were discretized into clinically meaningful ranges. The model’s performance was significantly limited by strong class imbalance in the dataset. In the binary classification task (e.g., predicting left atrial diameter class), the model achieved a high F1 score for the majority class (0.88) but completely failed to identify the minority class (F1 = 0.00), despite class weighting. 4.1. Limitations • The dataset size (n = 37 patients) was small, limiting the generalizability and statistical power of models. • Many EchoCG parameters exhibited significant class imbalance, limiting the applicability of standard classifiers. • The features were extracted from structured ECG and EchoCG datasets; waveform-based deep learning was not explored. • Ensemble methods such as XGBoost or LightGBM designed for imbalanced data were not employed.
This study confirms that machine learning models can predict several echocardiographic parameters from ECG features with promising accuracy, particularly when using ensemble methods like Random Forests. However, data limitations — especially in size and class distribution — currently constrain the reliability and scope of these predictions. With further development and clinical validation, ECGdriven estimation of EchoCG parameters could become a valuable, low-cost tool for cardiac screening and monitoring. Moreover, the developed method undoubtedly has significant potential outside the cardiology clinic, for example, for an objective assessment of the functional state of military personnel.
Support for this research was provided by the National Research Foundation of Ukraine under project No. 2023.04/0094, titled "Development of technology for objective monitoring of functional capabilities and stress of military personnel based on miniature electrocardiographs and machine learning."
During the preparation of this work, the authors used GPT-4o in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.