1.

INTRODUCTION

According to data from the World Health Organization, cardiovascular disease has always been one of the main causes of death worldwide [1-3]. And the electrocardiogram signal is received and generated through non-invasive sensing devices [4-6]. Doctors have been using electrocardiogram signals to detect various heart diseases such as arrhythmia and myocardial infarction [7]. It reflects the multidimensional information of the heart activity of the detector and is one of the important means to evaluate the heart condition [8, 9].

The electrocardiogram (ECG) signal is usually composed of P-waves QRS complex Composed of four waveforms: T wave and U wave [10-13]. Some related research involves processing these signals, mining features, and combining statistical methods for analysis to achieve auxiliary diagnosis of various heart diseases [14-17]. A method based on fractional Fourier transform (FrFT) is used for denoised ECG signals in the time frequency plane. The results show that this method can easily detect QRS complex waves in clean ECG signals, and it has been validated using MIT-BIH arrhythmia database data [18]. Li Wei et al.’s study used a dataset containing demographic data, complications, laboratory tests, 12 lead resting electrocardiogram reports and 180 electrocardiogram features, as well as responsive CAG reports, to extract 14 feature training models. The results showed that the model performed better than internal medicine resident physicians, medical interns, and emergency physicians [19]. It is obvious that this type of research has high requirements for the quality of annotation. Lahcen El Bouny et al. proposed a new QRS complex classification method based on stationary wavelet transform (SWT), along with two classifiers, support vector machine (SVM) and K-nearest neighbor (KNN). SWT is used to extract discriminative features from useful sub bands of each QRS complex category. The extracted features are used as inputs for SVM and KNN for classification. The results show acceptable performance [20]. In addition, due to differences in the knowledge and experience of doctors themselves, it is easy to cause errors in signal recognition, and the actual clinical value of the results is limited.

Recent studies have focused on using deep learning methods to learn and recognize electrocardiogram signals, like convolutional neural network (CNN) and RNN [21-25]. Researchers have proposed a QRS wave feature detection algorithm based on multi-resolution wavelet transform, which uses neural networks and SVM to classify and analyze the extracted features, and demonstrates good classification performance [26]. Researchers used the MIT-BIH arrhythmia database and four different CNN architectures to classify and evaluate signal data represented by Short Time Fourier Transform (STFT), as well as bispectral and third-order cumulant processing, and discussed the model results [20]. In another study, a new wavelet sequence model DBLSTM-WS based on deep bidirectional LSTM network was proposed for classifying electrocardiogram signals. Meanwhile, a new method based on wavelet layers for generating ECG signal sequences was studied and implemented, and the results showed that the model has high classification performance [27].

In this study, we consider the need for doctors and related wearable detection devices to quickly identify abnormal beats in clinical diagnosis. Calculated and extracted high-dimensional features of the corresponding signal based on the performance of the time series signal itself. And based on these high-dimensional features of the signals, train a nomogram model to achieve rapid recognition and interpretation of differential heart beats in electrocardiogram signals. This model can significantly reduce the pressure of early screening and form a complete workflow, providing new ideas for clinical monitoring and other engineering needs.

2.

MATERIAL AND METHOD

3.

RESULTS

3.2

Feature extraction and selection

We extracted a total of 783 features, including 275 time-domain features (such as statistical, descriptive, peak, and change features), 367 frequency-domain features (including Fourier transform features, power spectrum features, FFT coefficient features), and 141 other features (autocorrelation, complexity, periodicity). Initially, all features were normalized using the scale function. After performing analysis of variance (ANOVA), 523 features remained. We then conducted correlation analysis, removing 119 highly correlated features. The remaining features were ranked using the minimum redundancy maximum relevance (mRMR) method, selecting the top 50. LASSO regression analysis with cross-validation was then applied to these 50 features in the training cohort to determine the optimal subset (Figure 1). Ultimately, the 10 most important features were identified: variance_larger_than_standard_deviation, cid_ce__normalize_True, fft_coefficient__attr_.real.__coeff_1, fft_coefficient__attr_.imag.__coeff_1, fft_coefficient__attr_.imag.__coeff_2, fft_coefficient__attr_.imag.__coeff_8, fft_coefficient__attr_.abs.__coeff_2, fft_coefficient__attr_.abs.__coeff_8, fft_aggregated__aggtype_.variance, and permutation_entropy__dimension_7__tau_1. These included 1 time-domain feature, 7 frequency-domain features, and 2 other features, with their correlations shown in Figure 2.

Figure 1.

LASSO Regression Model for Radiomic Feature Selection. (a) LASSO coefficient profiles for 100 radiomic features. (b) Optimal feature selection based on AUC value. The dotted vertical lines indicate the optimal λ values based on the minimum criteria and 1 standard error of the minimum criteria. AUC, area under the curve; LASSO, least absolute shrinkage and selection operator.

Figure 2.

Correlation Map of the Final 10 Extracted Features

3.3

Nomogram construction, validation, and performance

We analyzed features related to identifying abnormal beats from the training set using multiple logistic regression analysis. The resulting predictive multivariate logistic regression equation is as follows:

The nomogram (Figure 3) incorporating selected features based on multivariate logistic regression analysis was constructed using the training dataset and validated with the testing dataset. The model performed well in both datasets. In the training set, the AUC was 0.958 (95% CI, 0.851-0.964) with 0.930 specificity and 0.881 sensitivity; accuracy was 0.961, precision 0.971, recall 0.986, and F1 score 0.979. In the test set, the AUC was 0.947 (95% CI, 0.944-0.949) with 0.851 specificity and 0.892 sensitivity; accuracy was 0.951, precision 0.959, recall 0.988, and F1 score 0.973 (Figure 4, Table 1). The calibration chart for the training set showed that both the Bias-corrected line and the Apparent line were very close to the Ideal line, with a mean square error of 0.002, indicating high prediction accuracy (Figure 5). The test set calibration chart also showed high consistency between predicted probabilities and observed proportions, further validating the model’s robustness. Decision curve analysis (Figure 6) from both datasets confirmed that the nomogram model provided a higher net benefit compared to “no anomalies” and “all anomalies” strategies, demonstrating significant clinical application value.

Figure 3.

Nomogram for Predicting Abnormal Beat Risk (The nomogram is constructed based on ECG signal features. ECG, electrocardiogram.)

Figure 4.

Model Performance in Training and Testing Datasets. (a) ROC curve performance of the nomogram model in the training dataset. (b) ROC curve performance of the nomogram model in the testing dataset. (c) ROC curves of CNN models in training and testing datasets. (d) ROC curves of the VGG model in training and testing datasets.

Figure 5.

Nomogram-Predicted Survival (Predicted versus actual probability of abnormal beats in the (a) training dataset. (b) testing dataset.)

Figure 6.

Clinical Net Benefit of Nomogram Models (Decision curves for predicting the risk of abnormal beats in the (a) training dataset. (b) testing dataset. The black line represents the assumption that all beats are normal (“treat-none”). The gray line represents the assumption that all beats are abnormal (“treat-all”). The red line indicates the net benefit of the nomogram model.)

Table 1.

The performance of models in datasets

	precision	recall	F1	acc
Proposed Model in Training dataset	0.971	0.986	0.979	0.961
Proposed Model in Testing dataset	0.960	0.988	0.973	0.951
CNN in Testing dataset	0.925	0.951	0.938	0.937
VGG In Testing dataset	0.874	0.979	0.924	0.919

4.

DISCUSSION

This study developed and validated a nomogram model targeting the characteristics of electrocardiogram signals. This model can quickly and accurately identify and describe abnormal beats, thereby helping to improve the diagnostic efficiency of clinical and related examination equipment. In addition, the process designed in this study (starting from segmenting electrocardiogram signals, using tsfresh software package to calculate features, and finally using nomograms for evaluation) has strong reusability and can be well embedded in real-time electrocardiogram monitoring strategies of various daily portable examination devices, playing a good auxiliary role.

Compared with the results of two deep learning models, it can be seen that the Nomogram model even outperforms their performance, indicating that the model in this study has good clinical application ability. Of course, it is undeniable that as the number of iterations increases, deep learning models will eventually become better. However, while ensuring performance, this research model can better interpret the patient’s ECG signal performance and heart condition through the analysis of signal characteristics.

The 10 key signal features extracted in the study include 1 time domain feature, 7 frequency domain features, and 2 other features: The feature ‘variance_lager_than_standard-deviation’ is mainly used to ensure the quality and consistency of electrocardiogram signal data. It checks the validity of the data by verifying whether the variance is greater than the standard deviation, thereby reflecting the stability and noise level of the signal. The feature ‘cid_ce_ normalize.True’ is a complexity density estimation feature that has been normalized. This feature evaluates the regularity of cardiac activity and health status by analyzing the complexity and degree of variation of ECG signals. Feature ‘fftcofefficent_attr_. real.__ Coeff_1 ‘is one of the frequency domain features obtained by performing Fourier transform on ECG signals, specifically referring to the first real part coefficient of the frequency domain signal. This coefficient mainly evaluates the DC component or average value of the signal throughout the entire sampling period, and is mainly used to detect the overall level and baseline offset of the signal. Feature ‘ffft_coefficent_attr_.image.__Coeff_1 ‘is similar to the previous feature, but mainly reflects the phase information of the electrocardiogram signal at the lowest frequency component, showing the low-frequency changes and phase characteristics of the electrocardiogram signal, and reducing noise interference. Feature ‘ffft_coefficent_attr_.image.__Coeff_2 ‘reflects the phase information of the electrocardiogram signal on the second frequency component. This feature reflects the phase characteristics and periodic changes of the electrocardiogram signal, and to some extent indicates the periodic changes and cardiac cycle activity in the electrocardiogram signal. Feature ‘ffft_coefficent_attr_.image.__Coeff_8 ‘reflects the phase information of the electrocardiogram signal on the 8th frequency component, reflects the detailed changes in the electrocardiogram signal and the phase characteristics of the high-frequency component, reflects the rapid electrical activity of the heart, and also reduces noise interference to a certain extent. Feature ‘fft_cofefficent_ attr_. abs.__ Coeff_2 ‘reflects the energy and intensity information of the electrocardiogram signal at the second frequency component, indicating the amplitude of periodic changes in the heart. Feature ‘fft_cofefficent_ attr_. abs.__ Coeff_8 ‘reflects the energy and intensity information of the electrocardiogram signal at the 8th frequency component. This feature focuses more on describing the mid to high frequency changes and detailed features of the electrocardiogram signal. Feature ‘fft_Aggregated_aggtype_.variance.’ reflects the variance of the amplitude of the electrocardiogram signal in the frequency domain, represents the degree of dispersion of the energy distribution of different frequency components, and evaluates the complexity of the electrocardiogram signal. The feature ‘permutation_entropy_dimension_7_tau_1’ reflects the arrangement entropy of the electrocardiogram signal, measuring the complexity and irregularity of the signal. The features extracted from ECG signals provide a comprehensive and robust analysis, reducing the need for specialized knowledge to identify different segments of the heartbeats. Covering time domain, frequency domain, and complexity measures, these features allow for a holistic understanding of ECG signals.

This approach ensures higher stability, stronger resistance to interference, and reduced noise impact. The detailed insights gained from these features enhance their applicability, making them suitable for embedding in various detection devices. By optimizing the diagnostic process, these features enable a more focused detection of abnormal heartbeats, ultimately improving the accuracy and efficiency of cardiac diagnostics.

This study had some limitations. Firstly, the data comes from a single public database and lacks support from multi center data. Therefore, the trained nomogram model may have a certain degree of bias. In future research, we will expand the sample size and collect recent clinical samples for training and validation to ensure the model’s generalization ability. Secondly, the EEG signal characteristics of the research center were only calculated using an open-source software package, so there may be a certain degree of tilt. In future research, we will also consider more methods for calculating signal features. In addition, the data volume of the two types of electrocardiogram signals was balanced, which may disrupt the authenticity of the data distribution and have a certain impact on the model.

5.

CONCLUSIONS

In summary, this study proposes a column chart based on the characteristics of electrocardiogram signals to detect abnormal pulsation in patients. The results confirm that the column chart can help clinical doctors diagnose and optimize treatment, and has clinical application value.