Automatic sleep staging method based on multi-scale decision fusion

Abstract

The invention provides an automatic sleep staging method based on multi-scale decision fusion, which comprises the following steps: S1, performing multi-scale windowing processing on an original multi-modal physiological signal to obtain a long window stream and a short window stream; s2, performing feature learning on the signals of the long window flow and the short window flow by adopting a parallel double-flow feature learning architecture to obtain respective sleep staging prediction probabilities; and S3, performing decision-level fusion on the prediction probabilities of the long window stream and the short window stream to generate a final sleep staging result. According to the method, the physiological signals of the long time window and the short time window are processed in parallel, decision results of the physiological signals are fused, accurate recognition of the sleep stage is achieved, and the problem that in the prior art, the staging accuracy is limited due to the fact that macroscopic background and microcosmic event characteristics are difficult to capture at the same time through a single time window is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering and artificial intelligence, and more particularly to an automatic sleep staging method based on multi-scale decision fusion. Background Technology

[0002] Sleep staging is fundamental to assessing sleep structure and quality in sleep medicine and is crucial for diagnosing and treating sleep disorders. Traditional sleep staging, based on the American Academy of Sleep Medicine (AASM) guidelines, involves professionals manually interpreting polysomnography (PSG) signals, analyzing them in 30-second increments. However, this process is not only time-consuming and labor-intensive, but the interpretation results are also influenced by subjective factors, with limited consistency among different experts. With the development of artificial intelligence technology, researchers have developed various automated sleep staging algorithms based on machine learning and deep learning to improve the efficiency and accuracy of staging.

[0003] In recent years, deep learning models have made significant progress in the field of automatic sleep staging. For example, models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) can directly learn discriminative features from raw signals, significantly improving staging performance. Some studies have proposed different network architectures and methods to further improve staging accuracy. For example, Zhou et al. (2022) proposed SingleChannelNet, which uses a 30-second window combined with the context of adjacent epochs to improve the stability of the N1 stage; Dong et al. (2017) used a 5-second window combined with LSTM to enhance transient event detection; Zhao et al. (2025) proposed SE-MSResNet, which extracts features through multi-scale residual blocks; and Lu et al. (2024) introduced multi-scale convolutional kernels into Mast-GCN, focusing on depression identification.

[0004] However, most existing automatic sleep staging algorithms still use the clinically standard 30-second fixed window for analysis, which presents a core technical bottleneck: a single 30-second window is not optimal for capturing all relevant sleep physiological phenomena. On the one hand, stages N1 and N3 require longer time windows to observe continuous rhythmic changes; on the other hand, sleep spindles and K-complexes specific to stage N2, as well as rapid eye movements (REM) during REM sleep, are transient events lasting only a few tenths of a second to a few seconds. An excessively long 30-second window may cause these crucial microstructural features to be overwhelmed by background signals, thus affecting staging accuracy, especially for stage N2 identification. Furthermore, while some existing improved methods attempt to use shorter windows or fuse multiple adjacent 30-second windows, they typically optimize only a single time scale or process multi-scale information within a unified network path. This can lead to interference between features at different scales, failing to fully leverage the unique advantages of each scale's information. Therefore, a new sleep staging method that can effectively and collaboratively process and fuse multi-scale temporal information is urgently needed to overcome the limitations of existing technologies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an automatic sleep staging method based on multi-scale decision fusion. This invention achieves accurate identification of sleep stages by processing physiological signals from both long and short time windows in parallel and fusing their decision results. This effectively solves the problem in existing technologies where a single time window is insufficient to simultaneously capture macroscopic background and microscopic event characteristics, thus limiting staging accuracy.

[0006] The technical means employed in this invention are as follows: An automatic sleep staging method based on multi-scale decision fusion includes: S1. Perform multi-scale windowing processing on the original multimodal physiological signals to obtain long window streams and short window streams; S2. A parallel dual-stream feature learning architecture is used to learn features from the long window stream and the short window stream signals respectively, so as to obtain their respective sleep stage prediction probabilities. S3. The prediction probabilities of the long window stream and the short window stream are fused at the decision level to generate the final sleep staging result.

[0007] Further, step S1 includes: S11. Receive raw multimodal physiological signals, including electroencephalogram (EEG) signals, electrooculogram (EOG) signals, and electromyogram (EMG) signals; S12. The original multimodal physiological signals are segmented according to the clinical standard 30-second long window to obtain a long window stream; S13. The original multimodal physiological signal is segmented into shorter window segments to obtain a short window stream. The window length of the short window stream is selected according to the specific physiological event to be captured, and can be 1 second, 2 seconds, 5 seconds, 10 seconds or 15 seconds.

[0008] Further, step S12 includes: S121. Determine the start and end points of the long window, and continuously divide it into 30-second segments. S122. Assign a corresponding sleep stage label to each long window. The sleep stage labels are marked by professionals according to clinical standards to obtain the long window stream.

[0009] Further, step S13 includes: S131. Select the corresponding short window length based on the characteristics of the test data; S132. Determine the start and end points of the short window, and continuously divide it according to the selected short window length; S133. Apply the Hamming window to truncate the signal within each short window to reduce the abrupt changes caused by signal truncation, thereby reducing spectral leakage. S134. Assign a corresponding sleep stage label to each short window. The sleep stage label is inherited from the long window to which it belongs.

[0010] Further, step S2 includes: S21. Perform independent feature extraction on each signal channel to avoid premature fusion that could lead to information loss; S22. The signal of each channel is extracted by a feature extractor consisting of multiple cascaded two-stream convolutional blocks. As the network layer deepens, the dimension of the feature map increases and the temporal resolution decreases, thereby learning multi-level abstract features from local details to global contours. S23. Concatenate the deep feature vectors extracted from each channel to form an aggregated feature vector; S24. The aggregated feature vector is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each input window.

[0011] Further, step S21 includes: S211. Perform independent feature extraction on each signal channel of the long window stream to obtain the feature representation of the long window stream; S212. Perform independent feature extraction on each signal channel of the short window stream to obtain the feature representation of the short window stream.

[0012] Further, step S22 includes: S221. For the signal of each channel of the long window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the long window stream. S222. For the signal of each channel of the short window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the short window stream.

[0013] Further, step S23 includes: S231. The deep feature vectors extracted from each channel of the long window stream are concatenated to form the aggregated feature vector of the long window stream. S232. The deep feature vectors extracted from each channel of the short window stream are concatenated to form the aggregated feature vector of the short window stream.

[0014] Further, step S24 includes: S241. The aggregated feature vector of the long window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each long window input window. S242. The aggregated feature vector of the short window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each short window input window.

[0015] Further, step S3 includes: S31. For each 30-second sleep segment, collect the predicted probability vectors of all corresponding short windows, and perform element-wise averaging to obtain a probability vector of a 30-second sleep segment after short window aggregation. S32. The probability vector after short window aggregation is weighted and summed with the probability vector of the long window corresponding to the 30-second segment to obtain the fused probability vector. S33. Take the category with the highest probability value in the fusion probability vector as the final sleep staging result for the 30-second segment.

[0016] Compared with the prior art, the present invention has the following advantages: 1. The present invention provides an automatic sleep staging method based on multi-scale decision fusion. By using multi-scale signal processing technology, the original multimodal physiological signal is divided into signal streams of two time scales, long and short, to capture macro background and micro event features respectively, thereby improving the accuracy of sleep staging and enhancing the ability to identify specific sleep stages.

[0017] 2. The present invention provides an automatic sleep staging method based on multi-scale decision fusion, which adopts a parallel dual-stream feature learning architecture to enable long and short window streams to independently learn the morphological features of signals at corresponding time scales, avoiding mutual interference between features at different scales, thereby giving full play to the unique advantages of information at each scale and further improving sleep staging performance.

[0018] 3. This invention provides an automatic sleep staging method based on multi-scale decision fusion. By employing decision-level fusion strategy techniques, it weighted and fused the prediction results independently output by long and short window streams. This method can simultaneously capture the macroscopic background rhythms required to define sleep stages (long window advantage) and the key instantaneous events that distinguish stages (short window advantage). By integrating macroscopic and microscopic information, it generates more accurate and robust sleep staging results, providing more refined analytical tools for clinical diagnosis and scientific research. This effectively solves the problems of limited staging accuracy and insufficient identification of specific sleep stages caused by single time windows in existing technologies.

[0019] 4. This invention provides an automatic sleep staging method based on multi-scale decision fusion, which significantly enhances the ability to identify N2 stage. The key to defining N2 stage lies in the sleep spindle wave and K-complex wave, and these microstructures are more easily detected under short window conditions. This invention specifically learns these features through short window branching, effectively solving the problem that the traditional 30s model has difficulty accurately identifying N2 stage.

[0020] 5. This invention provides an automatic sleep staging method based on multi-scale decision fusion, which not only improves the accuracy of sleep staging but also provides more valuable information for clinical diagnosis due to its sensitivity to sleep microstructure. For example, by analyzing the prediction results of short windows, the occurrence of events such as sleep spindle waves can be quantified, which is of great significance for studying neurodegenerative diseases and assessing cognitive function. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a diagram of the overall architecture of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0025] like Figure 1 As shown, this invention provides an automatic sleep staging method based on multi-scale decision fusion, comprising: S1. Perform multi-scale windowing processing on the original multimodal physiological signals to obtain long window streams and short window streams; S2. A parallel dual-stream feature learning architecture is used to learn features from the long window stream and the short window stream signals respectively, so as to obtain their respective sleep stage prediction probabilities. S3. The prediction probabilities of the long window stream and the short window stream are fused at the decision level to generate the final sleep staging result.

[0026] In a specific implementation, as a preferred embodiment of the present invention, step S1 includes: S11. Receive raw multimodal physiological signals, including electroencephalogram (EEG) signals, electrooculogram (EOG) signals, and electromyogram (EMG) signals; S12. The original multimodal physiological signal is segmented according to the clinical standard 30-second long window to obtain a long window stream; Step S12 includes: S121. Determine the start and end points of the long window, and continuously divide it into 30-second segments. S122. Assign a corresponding sleep stage label to each long window. The sleep stage labels are marked by professionals according to clinical standards to obtain the long window stream.

[0027] S13. The original multimodal physiological signal is segmented into shorter window segments to obtain a short window stream. The window length of the short window stream is selected according to the specific physiological event to be captured, and can be 1 second, 2 seconds, 5 seconds, 10 seconds or 15 seconds. Step S13 includes: S131. Select the corresponding short window length based on the characteristics of the test data; S132. Determine the start and end points of the short window, and continuously divide it according to the selected short window length; S133. Apply the Hamming window to truncate the signal within each short window to reduce the abrupt changes caused by signal truncation, thereby reducing spectral leakage. S134. Assign a corresponding sleep stage label to each short window. The sleep stage label is inherited from the long window to which it belongs. In this implementation, in order to detect sleep spindle waves lasting 0.5-2 seconds, a short window length of 2 seconds can be selected. All short windows segmented from the same 30-second long window inherit the manually labeled sleep stage label of that long window for subsequent supervised learning of the model.

[0028] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Perform independent feature extraction on each signal channel to avoid premature fusion leading to information loss; Step S21 includes: S211. Perform independent feature extraction on each signal channel of the long window stream to obtain the feature representation of the long window stream; S212. Perform independent feature extraction on each signal channel of the short window stream to obtain the feature representation of the short window stream.

[0029] S22. The signal of each channel is processed by a feature extractor consisting of multiple cascaded two-stream convolutional blocks. As the network layers deepen, the dimension of the feature map increases and the temporal resolution decreases, thereby learning multi-level abstract features from local details to global contours; Step S22 includes: S221. For the signal of each channel of the long window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the long window stream. S222. For the signal of each channel of the short window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the short window stream.

[0030] S23. Concatenate the deep feature vectors extracted from each channel to form an aggregated feature vector; Step S23 includes: S231. The deep feature vectors extracted from each channel of the long window stream are concatenated to form the aggregated feature vector of the long window stream. S232. The deep feature vectors extracted from each channel of the short window stream are concatenated to form the aggregated feature vector of the short window stream.

[0031] S24. Pass the aggregated feature vector through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each input window. Step S24 includes: S241. The aggregated feature vector of the long window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each long window input window. S242. The aggregated feature vector of the short window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each short window input window.

[0032] In this embodiment, a parallel dual-branch structure is employed to process long and short window data streams from the multi-scale window module separately. The long and short window branches have the same network structure, but their weights are trained independently, allowing them to focus on learning the optimal features at their respective scales. To preserve the unique features of each physiological signal (such as EEG and EOG from different leads), the network performs independent feature extraction for each signal channel in the early stages of the module, avoiding premature fusion that could lead to information loss. Each channel's signal passes through a feature extractor composed of multiple cascaded "dual-stream convolutional blocks (DSCBs)". DSCBs effectively extract morphological features of the signal through parallel convolutional paths. As the network layers deepen, the dimension of the feature maps increases, and the temporal resolution decreases, thereby learning multi-level abstract features from local details to global contours. The deep feature vectors extracted from each channel are then concatenated to form an aggregated feature vector. This vector is finally passed through a fully connected classification head and a Softmax function, outputting a predicted probability vector for each input window containing five sleep stages (Wake, N1, N2, N3, REM).

[0033] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. For each 30-second sleep segment, collect the predicted probability vectors of all corresponding short windows, and perform element-wise averaging to obtain a probability vector of a 30-second sleep segment after short window aggregation. S32. The probability vector after short window aggregation is weighted and summed with the probability vector of the long window corresponding to the 30-second segment to obtain the fused probability vector. S33. Take the category with the highest probability value in the fusion probability vector as the final sleep staging result for the 30-second segment.

[0034] In this embodiment, for each 30-second sleep segment, CSFM first collects the predicted probability vectors of all corresponding short windows. For example, if the short window is 2 seconds, there will be 15 short window probability vectors. The module averages these 15 vectors element-wise to obtain an aggregated short window probability vector, which represents the comprehensive judgment of short-scale analysis within these 30 seconds. The aggregated short window probability vector is then weighted and summed with the long window probability vector corresponding to the 30-second segment (in this invention, a simple average is used, i.e., each weight is 0.5). This fused probability vector P fusion It simultaneously incorporates both long-term macroscopic information and short-term microscopic event information. In the final decision-making process, the category with the highest probability value in the fused probability vector is taken as the final sleep stage result for that 30-second segment, i.e., Y. fusion =argmax(P fusion ).

[0035] Example: Sleep stage scoring in patients with sleep apnea (hospital setting) A hospital admitted a patient complaining of "nocturnal snoring with sleep apnea." Clinically, a sleep stage rating scale (PSG) is needed to assess the degree of sleep structure fragmentation. Traditional methods require sending PSG data to a professional physician for manual scoring. The method of this invention involves recording the patient's 8-hour sleep signal using a PSG device, inputting the data into a pre-trained MDFSleepNet model, and the model outputting the patient's sleep stage information. Based on this result, a sleep phase diagram can be drawn, or further medical diagnostic information can be provided to the doctor. The implementation steps of this invention can be standardized as follows: 1. Dataset Construction: Collect hospital PSG data (including expert annotations), divide it into training set / validation set in an 8:2 ratio, and perform bandpass filtering (0.3-45Hz), 50Hz notch denoising, and downsampling (200Hz→100Hz) on the signal. 2. Model training: Initialize MDFSleepNet, Adam optimizer (batch size 32, initial learning rate 1e-3, cosine annealing), train for 50 epochs, and save the model weights with the highest accuracy on the validation set. 3. Clinical Deployment: The trained model is integrated into the hospital system. After inputting PSG signals, it automatically outputs a scoring report (including sleep structure diagram, percentage of each stage, and transient event statistics). The report conforms to the DICOM standard and can be directly used for clinical diagnosis.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic sleep staging based on multi-scale decision fusion, characterized in that, include: S1. Perform multi-scale windowing processing on the original multimodal physiological signals to obtain long window streams and short window streams; S2. A parallel dual-stream feature learning architecture is used to learn features from the long window stream and the short window stream signals respectively, so as to obtain their respective sleep stage prediction probabilities. S3. The prediction probabilities of the long window stream and the short window stream are fused at the decision level to generate the final sleep staging result.

2. The automatic sleep staging method based on multi-scale decision fusion according to claim 1, characterized in that, Step S1 includes: S11. Receive raw multimodal physiological signals, including electroencephalogram (EEG) signals, electrooculogram (EOG) signals, and electromyogram (EMG) signals; S12. The original multimodal physiological signals are segmented according to the clinical standard 30-second long window to obtain a long window stream; S13. The original multimodal physiological signal is segmented into shorter window segments to obtain a short window stream. The window length of the short window stream is selected according to the specific physiological event to be captured, and can be 1 second, 2 seconds, 5 seconds, 10 seconds or 15 seconds.

3. The automatic sleep staging method based on multi-scale decision fusion according to claim 2, characterized in that, Step S12 includes: S121. Determine the start and end points of the long window, and continuously divide it into 30-second segments. S122. Assign a corresponding sleep stage label to each long window. The sleep stage labels are marked by professionals according to clinical standards to obtain the long window stream.

4. The automatic sleep staging method based on multi-scale decision fusion according to claim 2, characterized in that, Step S13 includes: S131. Select the corresponding short window length based on the characteristics of the test data; S132. Determine the start and end points of the short window, and continuously divide it according to the selected short window length; S133. Apply the Hamming window to truncate the signal within each short window to reduce the abrupt changes caused by signal truncation, thereby reducing spectral leakage. S134. Assign a corresponding sleep stage label to each short window. The sleep stage label is inherited from the long window to which it belongs.

5. The automatic sleep staging method based on multi-scale decision fusion according to claim 1, characterized in that, Step S2 includes: S21. Perform independent feature extraction on each signal channel to avoid premature fusion that could lead to information loss; S22. The signal of each channel is extracted by a feature extractor consisting of multiple cascaded two-stream convolutional blocks. As the network layer deepens, the dimension of the feature map increases and the temporal resolution decreases, thereby learning multi-level abstract features from local details to global contours. S23. Concatenate the deep feature vectors extracted from each channel to form an aggregated feature vector; S24. The aggregated feature vector is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each input window.

6. The automatic sleep staging method based on multi-scale decision fusion according to claim 5, characterized in that, Step S21 includes: S211. Perform independent feature extraction on each signal channel of the long window stream to obtain the feature representation of the long window stream; S212. Perform independent feature extraction on each signal channel of the short window stream to obtain the feature representation of the short window stream.

7. The automatic sleep staging method based on multi-scale decision fusion according to claim 5, characterized in that, Step S22 includes: S221. For the signal of each channel of the long window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the long window stream. S222. For the signal of each channel of the short window stream, feature extraction is performed by a feature extractor composed of multiple cascaded dual-stream convolutional blocks to obtain multi-level abstract features of the short window stream.

8. The automatic sleep staging method based on multi-scale decision fusion according to claim 5, characterized in that, Step S23 includes: S231. The deep feature vectors extracted from each channel of the long window stream are concatenated to form the aggregated feature vector of the long window stream. S232. The deep feature vectors extracted from each channel of the short window stream are concatenated to form the aggregated feature vector of the short window stream.

9. The automatic sleep staging method based on multi-scale decision fusion according to claim 5, characterized in that, Step S24 includes: S241. The aggregated feature vector of the long window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each long window input window. S242. The aggregated feature vector of the short window stream is passed through a fully connected classification head and a Softmax function to output a predicted probability vector containing five sleep stages for each short window input window.

10. The automatic sleep staging method based on multi-scale decision fusion according to claim 1, characterized in that, Step S3 includes: S31. For each 30-second sleep segment, collect the predicted probability vectors of all corresponding short windows, and perform element-wise averaging to obtain a probability vector of a 30-second sleep segment after short window aggregation. S32. The probability vector after short window aggregation is weighted and summed with the probability vector of the long window corresponding to the 30-second segment to obtain the fused probability vector. S33. Take the category with the highest probability value in the fusion probability vector as the final sleep staging result for the 30-second segment.

Images