A method for grading severity of obstructive sleep apnea based on multi-modal physiological signals

Abstract

The present application relates to the medical brain signal processing and artificial intelligence auxiliary diagnosis technology field in sleep process, specifically relates to a kind of obstructive sleep apnea severity grading method based on multi-modal physiological signal.Method includes: the Pan-Tomkins algorithm is handled to the ECG signal collected, and the heart rate variability time domain and frequency domain features are extracted;The blood oxygen saturation signal collected synchronously is preprocessed, and the oxygen reduction index and T90 are calculated as hypoxia load characteristics;Two kinds of feature vectors are fused at feature level, and multi-modal physiological feature vector is formed;The one-dimensional convolutional neural network trained is used to analyze the multi-modal characteristics, and the OSA severity grading result meeting clinical standard is output.The present application overcomes the limitation of single mode method by fusing the complementary information of ECG and two kinds of physiological signals, realizes the end-to-end automatic analysis from original signal to grading result, and improves the accuracy and clinical practicability of OSA severity classification.

Description

Technical Field

[0001] This invention relates to the field of medical electroencephalogram (EEG) signal processing and artificial intelligence-assisted diagnosis during sleep, specifically to a method for grading the severity of obstructive sleep apnea based on multimodal physiological signals. Background Technology

[0002] Obstructive sleep apnea (OSA), a common sleep disorder closely related to sleep stages, is a prevalent sleep disease affecting nearly one billion people worldwide. It is characterized by recurrent partial or complete upper airway collapse during sleep, leading to apnea and hypopnea events. Clinically, the severity of OSA is primarily quantified using the Apnea-Hypopnea Index (AHI). This index directly relates to patient risk stratification and the development of individualized treatment strategies; therefore, accurate and objective severity assessment is of significant clinical importance.

[0003] Currently, research on automated OSA-assisted diagnosis based on physiological signals has made some progress, but there are still significant limitations. Existing methods mostly rely on a single physiological signal (such as ECG or...). Analysis using conventional methods fails to fully characterize the complex pathophysiological processes involved in OSA across multiple systems, limiting the robustness and generalization ability of the models. Furthermore, most studies focus only on the discrete detection of respiratory events, failing to achieve end-to-end continuous grading of OSA severity that meets clinical standards. This creates an analytical gap between event detection and final grading, hindering the clinical applicability of automated diagnostic systems. Summary of the Invention

[0004] This invention provides a method for grading the severity of obstructive sleep apnea based on multimodal physiological signals, characterized by the following steps: S1: ECG signal processing and heart rate variability (HRV) feature extraction. The Pan-Tomkins algorithm is used to preprocess the raw ECG signal, accurately detect the R wave, and calculate the interval sequence between adjacent R waves. Based on this interval sequence, time-domain and frequency-domain HRV features are extracted. Time-domain features include at least the standard deviation of all intervals (SDNN) and the root mean square (RMSSD) of the difference between adjacent intervals; frequency-domain features include at least the low-frequency power (LF, 0.04-0.15Hz), high-frequency power (HF, 0.15-0.4Hz), and the LF / HF ratio.

[0005] S2: Obtain the data from the original dataset Signal, for the original The signal was preprocessed, using a moving average to smooth high-frequency noise and removing invalid segments caused by motion artifacts or signal loss. The raw signal was then analyzed to calculate features characterizing blood oxygen fluctuations and the degree of hypoxia accumulation during sleep.

[0006] S3: Multimodal feature fusion, utilizing a feature-level fusion strategy to fuse ECG signals and By leveraging the complementary features of the signals and utilizing deep learning models to comprehensively assess the extracted ECG feature vectors, the accuracy of OSA severity grading is significantly improved.

[0007] S4: It integrates a network model based on a multi-branch convolutional neural network architecture. By constructing a multi-branch parallel structure, including four main parts: multi-scale transformation, 2D feature extraction, GDDR feature extraction and 1D-CNN classifier, it solves the problem of insufficient representation ability of traditional single-scale feature extraction methods in sleep breathing event detection, and improves the classification accuracy of normal breathing, hypoventilation and apnea events.

[0008] As a further aspect of the present invention: the steps of the Pan-Tomkins algorithm described in S1 include bandpass filtering (passband 5-15Hz), differentiation, squaring, and sliding window integration of the ECG signal, and then locating the QRS complex center through dynamic threshold detection and search backtesting mechanism, and finally accurately locating the R wave and calculating the NN interval sequence.

[0009] As a further aspect of the present invention: the 1D-CNN classifier in S4 contains multiple convolutional layers and uses ReLU as the activation function.

[0010] The beneficial effects of the method in this embodiment are as follows: Multimodal information complementarity: by fusing information reflecting autonomic nervous activity (ECG-HRV) and information directly reflecting respiratory events leading to hypoxia. These two key physiological signals can more comprehensively characterize the pathophysiological state of OSA, overcoming the limitations of single-modal signals.

[0011] High classification accuracy: Experiments show that the multimodal fusion classification model proposed in this invention significantly outperforms single-modal or existing benchmark models in classification accuracy, F1 score and Kappa coefficient on public datasets (such as MIT-BIH PSG), demonstrating its higher classification accuracy and consistency.

[0012] Highly clinically applicable: It directly outputs severity grading results that meet clinical standards (such as normal, mild, moderate, and severe), providing doctors with intuitive and reliable auxiliary diagnostic information, which helps to promote intelligent screening and graded diagnosis of OSA.

[0013] In summary, this invention integrates the heart rate variability characteristics of ECG signals with... Based on the hypoxic load characteristics of the signal, the constructed multimodal grading system demonstrates significant advantages in the OSA severity classification task. In tests on public datasets, its various indicators outperform existing single-modal methods. This system not only overcomes the information limitations and insufficient robustness of traditional methods due to the single signal, but also achieves end-to-end automated analysis from raw signals to clinical grading results, providing reliable technical support for the accurate and efficient auxiliary diagnosis of OSA. Attached image description: The invention will now be further described with reference to the accompanying drawings.

[0014] Figure 1 This is a schematic diagram of the overall process of a method for classifying the severity of obstructive sleep apnea based on multimodal physiological signals according to the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0016] Example 1 Step S1: Data Preparation. Obtain the overnight sleep ECG signals synchronously recorded in the dataset and Signal.

[0017] Step S2: ECG Feature Extraction. The ECG signal is processed using the Pan-Tomkins algorithm: first, a 5-15Hz bandpass filter is applied to remove baseline drift and high-frequency noise; then, differentiation, squaring, and sliding window integration are performed to enhance the QRS wave; finally, dynamic threshold monitoring and R-wave localization logic are used to accurately identify each R-wave and generate an NN interval sequence. Based on accurate NN interval sequences The core features in the time and frequency domains are extracted. To quantify the overall fluctuation of heart rate and the degree of variation between adjacent heartbeats, time-frequency features are calculated using SDNN (standard deviation of all NN intervals) and RMSSD (root mean square of the difference between adjacent NN intervals), respectively. The calculation methods are expressed as formulas (1) and (2): (1) (2) Where M represents the total number of NN intervals, represents the value of the i-th NN interval, and NN represents the average value of all NN intervals.

[0018] Then, the power spectral density (PSD) of the NN interval sequence was calculated using Fast Fourier Transform (FFT), and the power in the LF, HF, and LF / HF ratio bands was calculated. Here, LF represents the power in the 0.04 Hz–0.15 Hz band, reflecting the combined effects of the sympathetic and parasympathetic nervous systems; HF represents the power in the 0.15 Hz–0.4 Hz band, primarily reflecting parasympathetic activity; and the LF / HF ratio reflects the balance between sympathetic and parasympathetic activity.

[0019] Step S3: Feature extraction. This study selected two representative indicators, the Oxygen Depletion Index (ODI) and T90, for feature extraction. The signal undergoes feature calculation and analysis. The oxygen desaturation index is defined as the number of times per hour during sleep when blood oxygen saturation drops above a preset threshold (typically 3% or 4%) and persists for a certain period (e.g., 10 seconds). T90 is defined as the percentage of total sleep time with blood oxygen saturation below 90%.

[0020] Step S4: Multimodal signal fusion. This study employs a feature-level fusion strategy to fuse ECG and... The extracted features are concatenated and input into a deep learning model for classification.

[0021] Feature vector construction: Converting ECG feature vectors and Feature vector The features are concatenated to form the final multimodal feature vector. This can be expressed as formula (3): (3) Step S5: The model is built based on OSA severity grading. A one-dimensional convolutional neural network (1D-CNN) is used as the classifier. Multimodal feature vectors are then processed. As input, the network terminates at a fully connected layer, outputting four neurons, each corresponding to one of the four severity levels of OSA. The output is converted into a probability using the Softmax function, expressed as formula (4): (4) in, Indicates the use of multimodal features Determine the probability that a sample belongs to level j. This represents the raw score given by the model to level j, and K represents the total number of categories (K=4). This indicates that the score is exponentially increased, making larger scores larger and smaller scores smaller.

[0022] Then, the model is trained using a dataset containing severity labels pre-annotated by experts according to AHI, with the optimization objective being to minimize the cross-entropy loss function. This is expressed as formula (5): (5) Where L is the overall loss value and N is the total number of samples. It is a real label mark. It is the probability that the model predicts that sample i belongs to sample j. This represents taking the logarithm of the probability; the more accurate the probability, the smaller the negative value of this number.

[0023] To verify the effectiveness of this invention, we conducted the following experiments: Experimental setup: Dataset: This invention uses the MIT-BIH PSG dataset, a publicly available dataset collected by the MIT Biomedical Engineering Laboratory for sleep research. It contains over 80 hours of four-, six-, and seven-channel PSG recordings from 16 subjects, each recording including beat-by-beat annotated electrocardiogram (ECG) signals. ECG signals were extracted from the PSG recordings in this dataset. The signals were analyzed and categorized into normal, low breathing, and apnea classes based on corresponding labels. Each duration of the electrocardiogram signal was sampled for 30 seconds at 250 Hz, with a total of 7500 samples recorded per session.

[0024] Evaluation metrics: This invention uses accuracy (Acc), macro F1 score (F1), Cohen's kappa coefficient (KP), and macro sensitivity (Sen), and calculates performance metrics from matrices to evaluate the classification performance of the TriGNet model.

[0025] Experimental results: To verify the fusion of ECG and To assess the effectiveness of the OSA severity grading model based on multimodal physiological signals, this study selected the MIT-BIH PSG sleep physiological signal public dataset as the test set. Accuracy (ACC), F1 score, and Kappa coefficient were used as core evaluation metrics. The proposed model was compared with the benchmark model TriGNet in the same field. The experimental results are shown in Table 3. The data in the table show that the multimodal grading model (Ours) constructed in this study significantly outperforms TriGNet in all metrics: the overall classification accuracy reached 88.13%, an improvement of 1.95 percentage points compared to TriGNet's 86.18%; the F1 score was 0.8172, higher than the benchmark model's 0.7879, indicating that the model has better class balance recognition ability in the imbalanced OSA four-level classification task; the Kappa coefficient was 0.7832, an improvement of 3.17% compared to TriGNet's 0.7591, verifying that the model's classification results are more consistent with the true labels.

[0026] Table 1. Experimental results of the model on the MIT-BIH PSG test set.

[0027] The above results fully demonstrate that, compared to the hierarchical scheme of single-mode signals, fusing the HRV characteristics of ECG with... The multimodal strategy for blood oxygen fluctuation characteristics can effectively integrate complementary information from different physiological signals, thereby improving the accuracy and stability of OSA severity grading and providing reliable technical support for intelligent grading diagnosis of clinical OSA.

[0028] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for grading the severity of obstructive sleep apnea based on multimodal physiological signals, characterized in that, Includes the following steps: S1. Preprocessing of ECG signals and extraction of heart rate variability features: The Pan-Tomkins algorithm is used to detect R waves and calculate the NN interval sequence, and time-domain and frequency-domain features are extracted based on this sequence; S2. Preprocessing and hypoxia feature extraction of blood oxygen saturation signals: The signal was smoothed and effective segments were filtered. The oxygen reduction index and T90 were calculated as characteristics of low oxygen load. S3. Multimodal Feature Fusion: The ECG feature vector extracted in step S1 is combined with the feature vector extracted in step S2. The feature vectors are concatenated to form a multimodal feature vector; S4. Severity classification based on deep learning: Input the multimodal feature vectors into a trained one-dimensional convolutional neural network model and output the OSA severity classification results.

2. The method according to claim 1, characterized in that, The time-domain features mentioned in step S1 include at least the standard deviation of all NN intervals and the root mean square of the difference between adjacent NN intervals; the frequency-domain features include at least the low-frequency power in the 0.04-0.15Hz band, the high-frequency power in the 0.15-0.4Hz band, and their ratios.

3. The method according to claim 1, characterized in that, The oxygen reduction index mentioned in step S2 is the number of events per hour in which blood oxygen saturation drops above a set threshold and continues for a set time during sleep; T90 is the percentage of time with blood oxygen saturation below 90% in total sleep time.

4. The method according to claim 1, characterized in that, The one-dimensional convolutional neural network described in step S4 contains multiple convolutional layers, employs the ReLU activation function, and the output layer converts the scores into probability distributions corresponding to different severity levels through the Softmax function.

5. The method according to claim 1, characterized in that, The loss function used for model training in step S4 is the cross-entropy loss function.

6. A severity grading system for obstructive sleep apnea based on multimodal physiological signals, used to implement the method according to any one of claims 1-5, characterized in that, include: The signal processing module is used to process the input ECG signal and... The signal undergoes preprocessing and feature extraction, outputting ECG feature vectors and Feature vector; The feature fusion module is used to combine the ECG feature vector and Feature vectors are concatenated to form a multimodal feature vector; The intelligent classification module has a built-in pre-trained one-dimensional convolutional neural network model, which is used to receive the multimodal feature vectors and output the OSA severity classification results.

7. The system according to claim 6, characterized in that, The ECG processing unit in the signal processing module uses the Pan-Tomkins algorithm to realize R-wave detection and NN interval sequence extraction.

8. The system according to claim 6, characterized in that, The output of the intelligent classification module corresponds to the four clinical severity levels of OSA: normal, mild, moderate, and severe.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-5.

Images