Intelligent Diagnosis Model for Stroke in Elderly Patients Based on Electrocardiogram Classification and MRI Image Segmentation Algorithms

2.1 Diagnosis of LVH on ECG

LVH on an ECG typically shows abnormal findings such as elevated QRS complex voltage, mild prolongation of the QRS interval, leftward deviation of the electrical axis, and ST-T changes [11]. The specific abnormal signals are presented in Table 1.

Table 1. Abnormal ECG findings in LVH

2.2 ECG data collection

This study selects 12-lead ECG signals, which are closer to clinical practice, for the classification study of LVH. The 12-lead ECG provides more comprehensive spatial information, covering electrical activities from various directions of the heart, making it more advantageous for classifying and diagnosing LVH. We retrospectively collected data from elderly patients (aged > 60 years) who underwent resting ECG examinations in the outpatient and inpatient departments of our hospital between January 1, 2023, and December 31, 2023. The 12-lead ECG data were batch-exported from the server in nECG format. The results were evaluated by two experienced medical technicians, and if there were conflicting opinions, a third party was invited to make the final decision. The screening variables primarily focused on ECG parameters, including RaVL, SV3, RI, RaVF, SV1, and RV5 (or RV6) voltages, as well as the QRS wave duration. LVH was diagnosed based on the Sokolow-Lyon criteria: SV1 + RV5 (or RV6 amplitude) ≥ 4.0 mV in men, ≥ 3.5 mV in women, RV5 ≥ 2.5 mV, RI + SIII ≥ 2.5 mV [12]. According to this standard, the study population was divided into two groups: the LVH group and the control group. A total of 2,065 patients were included, with 413 diagnosed with LVH (LVH group) and 1,652 without LVH (control group). All participants in this study provided informed consent. The study data and materials were anonymized and kept confidential, intended solely for research purposes.

2.3 ECG data preprocessing

Preprocessing involves several processes [13], such as heart rate segmentation, resampling, denoising, and normalization, to ensure the quality and applicability of the data.

2.3.1 Denoising

ECG data are susceptible to various interferences during the collection process, mainly including the following: 1) Power line and harmonic interference, with a frequency of 50 Hz; 2) Electromyographic noise and high-frequency noise such as power supply ripple introduced by the sampling circuit, typically with a frequency above 100 Hz; 3) Respiratory baseline drift and DC components introduced by sampling, typically distributed in the frequency range of 0-0.7 Hz.

To eliminate the interferences present in the raw data, this paper designs processing algorithms to filter the raw ECG data:

(1) Second-order IIR notch filter to eliminate power line interference

The IIR filter [14] effectively filters periodic interference signals. The transfer function of the IIR notch filter is given by:

$H(z)=\frac{1-2\cos {{\omega }_{0}}{{z}^{-1}}+{{z}^{-2}}}{1-2a\cos {{\omega }_{0}}{{z}^{-1}}+{{a}^{2}}{{z}^{-2}}}$               (1)

where, z represents the complex variable, and a represents the location of the pole, ω₀ represents the notch point, the calculation method is:

${{\omega }_{0}}=2p{{f}_{0}}/{{f}_{s}}$                      (2)

where, f₀ is the signal frequency and f_s are the sampling frequency. The notch filter is a band-stop filter, which takes advantage of its narrow stopband characteristics to quickly attenuate the input signal at a specified frequency point. For example, if the ECG waveform is severely interfered with at 60 Hz, it can be filtered by a 50 Hz notch filter to obtain a clearer waveform, as shown in Figure 1.

image003.png

(a) ECG signal with power line interference

image004.png

(b) Filtered signal after processing

Figure 1. Power line interference removal using a notch filter

(2) Mittag-Leffler filter to remove electromyographic interference

The Mittag-Leffler filter is a nonlinear filter [15] commonly used for signals with long memory characteristics. It performs excellently when dealing with non-gaussian noise and non-stationary signals. The Mittag-Leffler function is a part of fractional calculus theory and is defined in its single-parameter form as:

${{E}_{\alpha }}(z)=\sum\limits_{n=0}^{\infty }{\frac{{{z}^{n}}}{\Gamma (\alpha n+1)}}\alpha >0$                      (3)

where, Γ is the gamma function and 0<α<1. The double-parameter form is described by the generalized Mittag-Leffler function with parameters α and β, expressed as the following power series:

${{E}_{\alpha ,\beta }}(z)=\sum\limits_{n=0}^{\infty }{\frac{{{z}^{n}}}{\Gamma (\alpha n+\beta )}}$                  (4)

The Mittag-Leffler filter is applied using the MATLAB function "ML_filter (t, y, sigma, alpha, beta)", which applies the Mittag-Leffler filter with exponential forgetting to a time series. The comparative effect of filtering out electromyographic interference is shown in Figure 2.

image005.png

(a) Electromyographic interference in ECG signal

image006.png

(b) Filtered signal after processing

Figure 2. Electromyographic interference removal using the Mittag-Leffler filter

(3) Wavelet denoising to eliminate baseline drift interference

Wavelet transform [16] inherits and develops the localized idea of the short-time Fourier transform while overcoming the drawback of fixed window size not changing with frequency. It provides a "time-frequency" window that changes with frequency and is an ideal tool for time-frequency analysis and processing of ECG signals. The main feature of wavelet transform is that it can fully highlight certain aspects of the problem’s characteristics through transformation, performing localized analysis of time (or space) frequency. By using scaling and translation operations, the signal (or function) can be gradually refined in multiple scales, achieving fine time subdivision at high frequencies and frequency subdivision at low frequencies. The wavelet transform formula used in this article is as follows (5). The Morlet wavelet is selected as the wavelet basis function, which has better localized characteristics in the frequency domain and is suitable for processing electrocardiogram vibration signals.

$WT(a,\tau )=\frac{1}{\sqrt{a}}\int_{-\infty }^{\infty }{f}(t)*\psi \left( \frac{t-\tau }{a} \right)dt$                       (5)

where, a(scale) controls the scaling of the wavelet function, and τ(transition) controls the translation of the wavelet function. The comparison effect of filtering baseline drift interference is shown in Figure 3.

image007.png

(a) Baseline drift in ECG signal

image008.png

(b) Filtered signal after processing

Figure 3. Wavelet denoising to remove baseline drift interference

2.3.2 Resampling

ECG signals in the dataset may have different sampling frequencies, leading to inconsistent lengths when inputting them. Therefore, resampling is necessary during preprocessing to ensure that all data have the same sampling frequency, making it easier to input data consistently for further analysis.

2.3.3 Heartbeat segmentation

In some datasets, a single ECG signal may be several minutes long, which cannot be directly input into the algorithm due to its large size. Therefore, after locating the R-wave, heartbeat segmentation is used to cut a long ECG segment into several shorter heartbeats. The dataset usually provides the location of the R-wave. After identifying the R-wave, heartbeat segmentation divides the signal between adjacent R-waves into individual heartbeat cycles, providing clear signal segments for subsequent analysis. If the dataset does not provide R-wave locations, localization methods must be applied to locate the R-wave before proceeding with further operations.

2.3.4 Normalization

Normalization is generally performed after heartbeat segmentation and denoising. It involves scaling the data to a small specific range to eliminate amplitude differences in the signal, which helps improve the stability and comparability of the algorithm. Various methods of normalization include z-score normalization, min-max normalization, and standardization.

2.4 Pre-trained ECG intelligent diagnosis model construction

Deep learning technology, through algorithms like deep neural networks, can learn feature representations from large-scale and complex ECG signals and gradually improve its ability to recognize abnormal ECGs through continuous adjustment and optimization. To fully utilize the multi-angle feature information of the twelve-lead ECG, this paper constructs a model based on CNN and Transformer networks, which extracts ECG features from both local waveform characteristics and global long-range dependency features, and integrates these features to detect the abnormal manifestations of LVH in ECG signals. This improves the recall rate for leads with specific LVH arrhythmia patterns. Meanwhile, the self-attention mechanism of the Transformer [17] can capture long-distance dependencies in the ECG signal, providing global contextual information for the model. The architecture of the ECG intelligent diagnosis model and the ECG signal classification process are shown in Figure 4.

The entire network structure in the figure above consists of three main parts: the lead-aware hierarchical Transformer module for extracting long-range dependencies and lead features across heartbeats, the attention convolution module for extracting local waveform features, and the feature fusion classification module for combining the outputs of these two modules from different views and outputting classification results. The classification process is as follows: the module inputs a multi-lead ECG signal, where m represents the number of leads and n represents the signal length. First, the signal is passed through the lead-aware hierarchical Transformer module to obtain the Transformer output, where C represents the patch embedding feature dimension. The ECG signal is then sent through the convolution attention module to obtain the convolution output. The lead-aware hierarchical Transformer module uses a self-attention mechanism to capture long-range dependency features across multiple heartbeats. Additionally, the lead-aware mechanism utilizes a window-based method to calculate self-attention scores from two views: the time dimension and the lead dimension, thereby focusing on the specific differences in the ECG signals from different leads. The original electrocardiogram signal X $\in$ R^m×n is first segmented into $m\times \frac{n}{4}$ non overlapping patches by the EGG Patch embedding layer, and each patch is converted into a vector of length Cy(0)$\in$R^C. Next, there is a hierarchical structure of 5 stages, with the first four stages consisting of several window based Transformer blocks and a patch merging layer, and the fifth stage consisting only of window based Transformer blocks. Among them, the window based Transformer blocks for the first two stages are time lead Transformer blocks, while the other stages are time Transformer blocks. The convolution attention module leverages the inductive bias of the local ECG signal waveform's time invariance, using convolutional blocks to extract local waveform features. The spatial and channel attention module (CBAM) [18] can focus on signal regions that are significant for distinguishing heart rhythms. Then, generalized mean (GeM) pooling [19] is applied to reduce the dimensionality of both the Transformer and convolution outputs. Finally, the lead-aware hierarchical Transformer output, after generalized mean pooling, and the attention convolution module output, after generalized mean pooling, are concatenated and passed through a fully connected layer. After activation by the Sigmoid function, the classification prediction results are obtained.

transformerjie_gou_.jpg

Figure 4. ECG signal classification process

2.5 Comparison of pre-training results of different models

In this section, we compare the performance of the model in Section 2.4 with other neural network models, and perform experimental validation using the PTB-XL public database ECG data. PTB-XL is currently the largest public ECG dataset [20], containing rich 12-lead ECG records, aimed at solving the problems of small-scale open datasets and the lack of benchmark tasks. The dataset contains 21,837 clinical 12-lead ECG records, all signals collected through 12 standard leads (I, II, III, AVL, AVR, AVF, V1, …, V6) connected to the subject's right arm and a reference electrode. According to the original labels, the 21,837 ECG records in the dataset are classified as normal (non-LVH) or abnormal ECGs (LVH, including QRS wave voltage elevation, slight prolongation of the QRS interval, left axis deviation, ST-T changes, etc.), and the dataset is divided into training, validation, and test sets in a 6:1:3 ratio. The experiments use the divided training, validation, and test sets to evaluate the performance of the ECG intelligent diagnostic model in this paper, comparing the performance of this model with that of Support Vector Machine (SVM) and CNN. The model in this article uses Adam optimization algorithm to derive network parameters, and the hyperparameter setting selects a network learning rate of 0.001, batch size of 128, epochs of 200. The model performance evaluation metrics include sensitivity, specificity, F1 score, accuracy, and AUC (Area Under the Curve, namely the ROC curve area), and the specific experimental results are shown in Figure 5.

image010.jpg

Figure 5. Comparison of pre-trained results of three different models

2.6 Transfer learning model classification result evaluation

To solve the problem of limited sample size in self built datasets and help optimize the generalization performance of the model, we have done the following work: firstly, the dataset is divided into training set, validation set, and test set, and randomly divided into 60% training set, 10% validation set, and 30% test set. The model is trained on the training set, the model parameters are adjusted on the validation set, and finally the model's real-world performance is evaluated on the test set. Then compare the performance of the PTB-XL dataset and the self built dataset to determine whether the model is underfitting or overfitting. Compared with the benchmark model, draw the accuracy of the training set and validation set for different training set sizes, and determine overfitting or underfitting through the learning curve. If overfitting is found, adjust the regularization parameters, feature set, or polynomial features. Simultaneously transfer learning methods [21] are used to construct the LVH ECG intelligent diagnosis model. By transferring the learned model parameters through weight transfer, a new model is trained, accelerating and optimizing the learning efficiency of the LVH ECG intelligent diagnosis model based on the self-built dataset. First, the self-built dataset is randomly divided into 60% training set, 10% validation set, and 30% test set. Based on the pre-trained model, the classification layer is adjusted to recognize LVH and non-LVH ECGs. The transferred LVH ECG intelligent diagnosis model is then trained on the self-built dataset, and compared with the accurate results marked by doctors. The model's sensitivity, specificity, accuracy, and F1 score on the test set were found to be 0.81, 0.92, 0.87, and 0.91, respectively, with an AUC of 0.91. The specific evaluation metrics and model performance are shown in Figure 6.