A near-infrared function fusion method, system and device based on a brain-computer interface

Abstract

The application provides a near-infrared function fusion method, system and device based on a brain-computer interface, and belongs to the technical field of medical devices, and specifically comprises the following steps: determining blood oxygen characteristics of a blood oxygen signal change inflection point period of the infrared signal based on a fusion analysis processing strategy; when the extraction processing efficiency of the blood oxygen characteristics of a target user does not meet the requirements, determining the extraction strategy of the blood oxygen characteristics of the blood oxygen signal change inflection point period of the user based on the similarity of the blood oxygen characteristics in different target users and in combination with the recognition processing result of the brain-computer interface data of the user; performing the extraction processing of the blood oxygen characteristics of the user by using the extraction strategy; and determining the update processing target of an analysis target user according to the extraction processing result of the blood oxygen characteristics of the user and the recognition deviation user data, so that the reliability of sleep staging processing is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical device technology, and in particular relates to a near-infrared functional fusion method, system and device based on brain-computer interface. Background Technology

[0002] In the field of sleep monitoring and brain function assessment, current mainstream technologies mainly rely on the acquisition and analysis of physiological signals from a single modality. Polysomnography (PSG), as the clinical gold standard, simultaneously records multiple signals such as electroencephalography (EEG), electrooculography, electromyography, blood oxygenation, and respiration, enabling a more comprehensive assessment of sleep structure and quality.

[0003] Specifically, the invention patent application CN202511748062.5, "Sleep Diagnosis Method, Device and Equipment Based on Ventilator," uses electroencephalogram (EEG), electrooculogram (EOG), and electrocardiogram (ECG) information, combined with user input information obtained based on ISI, to diagnose insomnia type, determine the insomnia type, and add the severity of obstructive sleep apnea (OSA) to the insomnia type, generating a sleep diagnosis report, which further improves the accuracy of insomnia type diagnosis and sleep diagnosis report. However, it has the following drawbacks: There is a certain correlation between blood oxygenation characteristics and electroencephalogram (EEG) characteristics. Under normal physiological conditions, changes in neural activity precede blood oxygenation response, and the delay time (usually 2-6 seconds) is a core indicator for assessing the efficiency of neurovascular coupling. Therefore, determining the identification and processing strategy for characteristic signals of blood oxygenation signal changes based on the user's sleep staging based on EEG characteristics, and thus improving the reliability of sleep staging processing for users with identification deviations, has become an urgent technical problem to be solved.

[0004] Therefore, there is an urgent need for a near-infrared functional fusion method, system, and device based on brain-computer interface. Summary of the Invention

[0005] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides a near-infrared functional fusion method based on a brain-computer interface, which includes: S1 uses the sleep stage identification deviation data obtained based on brain-computer interface data to determine the users among the users who are combined with near-infrared function for fusion analysis and processing, and takes them as the analysis target users. Based on the analysis target user data and the brain-computer signal parsing results in the reliable identification process of the analysis target users, the fusion analysis and processing strategy of the analysis target users is determined. S2 determines the blood oxygen characteristics of the inflection point period of the blood oxygen signal change of the infrared signal based on the fusion analysis and processing strategy. When the extraction and processing efficiency of the blood oxygen characteristics of the target user does not meet the requirements, the extraction strategy of the blood oxygen characteristics of the user's blood oxygen signal change inflection point period is determined based on the similarity of blood oxygen characteristics in different target users and combined with the recognition and processing results of the user's brain-computer interface data. S3 uses the extraction strategy to extract the user's blood oxygen features, and based on the results of the extraction of the user's blood oxygen features and the identification deviation user data, determines the update processing target of the analysis target user.

[0006] The beneficial effects of this invention are as follows: Using sleep staging identification bias data obtained from brain-computer interface data, users are identified through fusion analysis combined with near-infrared spectroscopy. The number of deviations between a user's historical identification results and the gold standard (such as manual interpretation) is a key inverse indicator for measuring the quality of their EEG signals and the algorithm's adaptability. By utilizing the proportion of users with identification bias, the screening ratio for blood oxygen signal extraction is dynamically adjusted (from a preset ratio to a third preset ratio), allowing the system to achieve an optimal balance between "ensuring high data reliability" and "retaining a sufficient sample size." Finally, using the synchronized EEG-blood oxygen data of this highly reliable user group, the blood oxygen fluctuation characteristics at specific changes (inflection points) in the blood oxygen signal are extracted. This feature will serve as a new and stable biomarker to assist in identifying users whose staging is less reliable when relying solely on EEG characteristics, thereby improving the overall accuracy and robustness of the sleep staging system.

[0007] Based on the extraction and processing results of users' blood oxygenation features and the identification of users with deviations, the system determines the updated processing targets for the analysis target users. When the proportion of users with deviations is too high, it indicates an urgent need for more reliable data, and the strategy tends to be more proactive in incorporating such data (lowering the threshold); conversely, a stricter screening approach is adopted (raising the threshold). The core basis for screening is the depth and breadth of the matching between the blood oxygenation features of ordinary users and existing reliable blood oxygenation feature combinations (reliable templates extracted from the original analysis target users). In this way, the system can continuously and safely expand its core high-quality user group like a snowball, further improving the efficiency of blood oxygenation feature extraction and processing.

[0008] Furthermore, the identification deviation data includes the number of identification deviations made by the user.

[0009] Furthermore, the method for determining the target users in the analysis is as follows: Based on the identification deviation data of different users' sleep stages, users whose identification deviation number exceeds the preset deviation number threshold are identified and regarded as users with identification deviation. The identification deviation ratio is determined based on the proportion of users with identification deviations among those undergoing sleep staging. S13 determines the target user for analysis among the users based on the recognition deviation ratio and the number of recognition deviations for different users.

[0010] Furthermore, when the user belongs to the target user of the analysis, the blood oxygen characteristics of the blood oxygen fluctuation period obtained by near-infrared functional monitoring corresponding to the sleep stage characteristics are determined based on the user's sleep stage characteristics, that is, the blood oxygen characteristics of the blood oxygen signal inflection point period.

[0011] Furthermore, the analysis of target user data includes the number of target users and their proportion among all users.

[0012] Furthermore, the analysis results of brain-computer signals during the reliable identification process of the target user are determined based on the changes in sleep stage characteristics during the reliable identification process of the target user.

[0013] Furthermore, the reliable identification process is the sleep staging process, and the staging result is an accurate sleep staging process.

[0014] Furthermore, the method for determining the fusion analysis and processing strategy for the target users is as follows: Based on the analyzed target user data, determine the proportion of the analyzed target user among the users; Based on the analysis results of brain-computer signals in the reliable identification process of the target user, the reliable identification processes that meet the similarity requirements of sleep stage feature signals are grouped into the same sleep process combination. Based on the proportion of the target users in the user base and the different combinations of sleep processes of the target users, a fusion analysis and processing strategy for the target users is determined.

[0015] In a second aspect, the present invention provides a computer device comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the aforementioned near-infrared functional fusion method based on a brain-computer interface when running the computer program.

[0016] Thirdly, this application provides a near-infrared functional fusion system based on a brain-computer interface, employing the aforementioned near-infrared functional fusion method based on a brain-computer interface, specifically including: The analysis and processing module is responsible for determining the fusion analysis and processing strategy for the target users. The extraction module is responsible for determining the extraction strategy for blood oxygenation characteristics during the inflection point period of the user's blood oxygenation signal changes; The target identification module is responsible for determining the update processing targets of the target users being analyzed.

[0017] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart of a near-infrared functional fusion method based on a brain-computer interface; Figure 2 This is a flowchart illustrating the methods for identifying target users. Figure 3 This is a flowchart illustrating the method for determining the fusion analysis and processing strategy for target users; Figure 4 This is a framework diagram of a near-infrared functional fusion system based on a brain-computer interface. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0022] Example 1 like Figure 1 As shown, this application provides a near-infrared functional fusion method based on a brain-computer interface, specifically including: S1 uses the sleep stage identification deviation data obtained based on brain-computer interface data to determine the users among the users who are combined with near-infrared function for fusion analysis and processing, and takes them as the analysis target users. Based on the analysis target user data and the brain-computer signal parsing results in the reliable identification process of the analysis target users, the fusion analysis and processing strategy of the analysis target users is determined. S2 determines the blood oxygen characteristics of the inflection point period of the blood oxygen signal change of the infrared signal based on the fusion analysis and processing strategy. When the extraction and processing efficiency of the blood oxygen characteristics of the target user does not meet the requirements, the extraction strategy of the blood oxygen characteristics of the user's blood oxygen signal change inflection point period is determined based on the similarity of blood oxygen characteristics in different target users and combined with the recognition and processing results of the user's brain-computer interface data. S3 uses the extraction strategy to extract the user's blood oxygen features, and based on the results of the extraction of the user's blood oxygen features and the identification deviation user data, determines the update processing target of the analysis target user.

[0023] Furthermore, the identification deviation data includes the number of identification deviations made by the user.

[0024] Specifically, such as Figure 2 As shown, the method for determining the target user in the analysis is as follows: This embodiment precisely selects the user group with the highest reliability in sleep staging using EEG features from all users, i.e., the target users for analysis. Its underlying logic is "using deviation as a yardstick, layered filtering, and selecting the best." The number of deviations between a user's historical identification results and the gold standard (such as manual interpretation) is a key inverse indicator for measuring the quality of their EEG signals and the algorithm's adaptability. By setting dynamically adjusted screening ratios (from a preset ratio to a third preset ratio), the system can achieve an optimal balance between "ensuring high data reliability" and "retaining a sufficient sample size." Finally, using the synchronized EEG-blood oxygenation data of this high-reliability user group, the blood oxygenation variation characteristics at specific changes (inflection points) in the blood oxygenation signal are extracted. This feature will serve as a new and stable biomarker to assist in identifying users with lower reliability when staging solely based on EEG features, thereby improving the overall accuracy and robustness of the sleep staging system.

[0025] S11 identifies users whose number of identification errors exceeds a preset threshold based on the identification error data of different users' sleep stages, and treats them as users with identification errors. Number of identification biases: This refers to the number of times the system's automatic sleep staging results for a single user, based on EEG characteristics, differ from the validated gold standard staging results. It is typically accumulated over a certain period (e.g., one month). It directly quantifies the unreliability of the user's staging results.

[0026] Preset deviation threshold: A benchmark value used to define whether a user's installment payment results have "significant problems". Users whose installment payment reliability exceeds this threshold are considered unqualified.

[0027] Users with identified deviations: These are users whose number of identification deviations exceeds a preset threshold. They are a group of "problem users" who require special attention and handling.

[0028] This step is designed to perform a preliminary and direct dichotomous screening of the user group. It forms the basis of the entire method, aiming to quickly and clearly identify users whose phase selection results are obviously unreliable, isolating them from the "potentially reliable user pool" and providing a clear decision-making basis for subsequent, more refined proportion selection. It solves the problem of "where to begin screening."

[0029] The significance of this step lies in establishing the starting point and core criteria of the screening process. It transforms the abstract issue of "reliability" into specific, statistically comparable values ​​(number of deviations) and sets a clear "pass / fail" threshold. This makes the entire screening process objective and operable, avoiding subjective judgment.

[0030] In the above steps, if there are no users with identification bias, a preset proportion of users are selected as the target users for analysis. That is, the user quantity constraint is constructed based on the preset proportion of users, and the preset proportion of users with the fewest identification biases are selected as the target users for analysis.

[0031] Furthermore, if there are users with identification deviations, it is necessary to further determine whether the number of users with identification deviations is less than a preset threshold for the number of users with identification deviations. If so, then the constraint conditions for the number of users are constructed based on the second preset proportion of users, and the users with the fewest identification deviations are selected as the target users for analysis. If not, then proceed to step S12.

[0032] It is understood that the preset ratio is less than the second preset ratio.

[0033] Specific decision-making logic and examples: Hypothetical scenario: The platform has 1000 users (U1-U100). In the past 30 days, the system has generated 30 sleep period reports for each user. The reports are compared with the gold standard, and the number of deviations for each user is counted.

[0034] The preset deviation threshold is set to 5 times.

[0035] Execution of the filter: Statistics show that the number of deviations for 80 users, including U5, U23, and U41, are 7, 9, 6, etc., respectively, all greater than 5.

[0036] Results: The number of users with identification bias was identified as {U5, U23, U41, ...}, totaling 80 people.

[0037] Path determination: Due to the existence of users with identification bias (80 people > 0), proceed to the next step to determine their quantitative relationship.

[0038] If there are no users with identification bias (the number of biases for all users ≤ 5): this indicates that the overall user data quality is excellent and there are no obviously unreliable users.

[0039] Decision: To build a robust model, it is still necessary to select the most representative (i.e., most reliable) users. Using the strictest preset ratio (e.g., 15%), select the 150 users with the fewest deviations from the 1000 users as the target users for analysis. Because the foundation is strong, the best among the best can be selected.

[0040] If there are users with identification bias, and their number is less than the preset threshold for the number of users with bias (assuming threshold = 100): the number of problematic users is small (80 people < 100 in this example), which is an isolated phenomenon, and the overall data pool quality is acceptable.

[0041] Decision: In order to obtain a more diverse sample while controlling risks, a moderately strict second preset ratio (e.g., 20%) was adopted to select the 200 people with the fewest deviations from 1000 people as the target users for analysis.

[0042] If there are users with identification bias, and their number is not less than the preset threshold for the number of users with identification bias (assuming that 150 users with identification bias are greater than 100): Decision: Proceed to S12 to further assess the scope of the problem's impact.

[0043] S12 determines the identification deviation ratio based on the proportion of users with identification deviations among users undergoing sleep staging processing; Understandably, this includes: S121 If the identification deviation ratio is greater than the preset deviation ratio threshold in the above steps, then the user number constraint is constructed based on the third preset ratio of users, and the users with the fewest identification deviations in the third preset ratio are selected as the analysis target users among the users. Otherwise, proceed to step S13. It should be noted that the third preset ratio is greater than the second preset ratio.

[0044] Identification deviation ratio: This refers to the proportion of users with identification deviations out of the total number of users. The formula is: (Number of users with identification deviations / Total number of users). It reflects the prevalence of phased reliability issues among the user group.

[0045] Preset deviation percentage threshold: A critical value used to determine whether a problem is "common" or "severe". For example, setting it to 15% means that if more than 15% of users are users with identification deviations, then the problem is common.

[0046] This step is designed to further assess the severity of the problem after identifying problematic users. Simply knowing that problematic users exist is insufficient; it's also necessary to determine whether they are "isolated cases" or a "widespread phenomenon." Different severity levels directly determine the scope of high-reliability data needed to address the issue, thus dynamically adjusting the screening ratio. This upgrades the screening strategy from "based on the number of problematic users" to "based on the breadth of the problem's impact." This allows the system to intelligently perceive the overall state of current data quality. If the problem is widespread (high ratio), the maximum scope of high-reliability data (the third preset ratio) must be used to establish a stronger correction benchmark; if the problem is not widespread, data can be used more economically.

[0047] Specific decision-making logic and example (continued from S11): Total users: 1000; Users with identification deviation: 120. Identification deviation ratio: 12%. Preset deviation ratio threshold: 15%.

[0048] Judgment: 12% < 15%, the identification deviation rate is not greater than the threshold, the impact of the problem is within a controllable range, and it has not reached a generalized level. Therefore, instead of using the most aggressive screening ratio, we proceed to S13 for a more refined individual-level assessment.

[0049] S13 determines the target user for analysis among the users based on the recognition deviation ratio and the number of recognition deviations for different users.

[0050] Additionally, it is understandable that in the above steps, S131 obtains the number of identification deviations for different users, determines whether all users have at least one identification deviation record, and if so, determines to construct the user number constraint based on the third preset proportion of users, and selects the third preset proportion of users with the fewest identification deviations as the analysis target users among the users; otherwise, proceeds to step S132. S132 determines the percentage of users with no identification errors based on the number of identification errors among different users, and determines the identification reliability coefficient by combining the average number of identification errors among different users. It then determines whether the identification reliability coefficient is less than a preset reliability coefficient threshold. If so, it determines to construct user quantity constraints based on a third preset proportion of users, and selects the third preset proportion of users with the fewest identification errors as the analysis target users. If not, it constructs user quantity constraints based on a second preset proportion of users, and selects the second preset proportion of users with the fewest identification errors as the analysis target users.

[0051] Furthermore, when the user belongs to the target user of the analysis, the blood oxygen characteristics of the blood oxygen fluctuation period obtained by near-infrared functional monitoring corresponding to the sleep stage characteristics are determined based on the user's sleep stage characteristics, that is, the blood oxygen characteristics of the blood oxygen signal inflection point period.

[0052] Reliability coefficient: A composite indicator that comprehensively reflects the phased reliability of all users. It is determined by the percentage of users with no identification errors and the average number of identification errors for all users (e.g., percentage of users with no errors / (average number of errors + 1)). The higher the coefficient, the higher the overall reliability. A preset reliability coefficient threshold is a target value used to determine whether the overall reliability meets the standard.

[0053] This step is set up for the final, most refined global health assessment. Under the premise that the scope of the problem's impact (S12) is controllable, this step delves into two key details: 1) whether there are "zero-error" perfect users, and 2) even if everyone has errors, what is the overall error level? By calculating and identifying reliability coefficients, the system can quantify the overall "reliability" of the current user pool. This allows for the final decision at the most critical decision-making juncture: whether to adopt a moderate or maximum proportion. Its significance lies in achieving a comprehensive quantitative assessment from "whether there is a problem" to "how deep and wide the problem is." It ensures that the determination of the final selection ratio is a scientific decision based on three dimensions: the number of problematic users, the scope of the problem's impact, and the overall quality of user data. This maximizes the purity (high reliability) and sufficiency (meeting modeling requirements) of the selected target user group for analysis.

[0054] Specific decision-making logic and examples (continued from S12): The current identification deviation ratio of 8% has not triggered the maximum ratio, proceed to S13.

[0055] S131: Check the existence of "zero error" users. Judgment logic: Check whether there has been at least one identification deviation in the historical records of all 1000 users. Example: Suppose that 150 users such as U1, U2, and U10 have a deviation count of 0, while the remaining 850 users have a deviation count ≥ 1.

[0056] Result: Not all users exhibited identification bias. Therefore, the extreme case of "all users have bias" was not triggered, and the process proceeded to S132. S132: Calculate the identification reliability coefficient and make a decision: The percentage of users with no identification errors = 150 / 1000 = 0.15, and the average number of identification errors for all users = (total number of errors) / 100. Assume the calculated average is 2.1 times. Design the identification reliability coefficient = percentage of users with no errors / (average number of errors + 1) = 0.15 / (2.1 + 1) ≈ 0.15 / 3.1 ≈ 0.048. Judgment: Set the preset reliability coefficient threshold to 0.05. The calculated 0.048 < 0.05. Decision: The overall identification reliability coefficient is slightly lower than the safety threshold, indicating that although the proportion of problematic users is not high, the overall average error level of users is not optimistic, and the data foundation has a slight vulnerability.

[0057] Therefore, a more stringent third preset ratio (e.g., 30%) was used as the basis for screening, selecting the 30 users with the fewest deviations as the target users for analysis. This was to establish a larger and more robust highly reliable user base, even with slightly insufficient reliability.

[0058] Assuming another scenario (where the coefficient meets the standard): If the calculated identification reliability coefficient is 0.06 > 0.05, it indicates good overall reliability. The decision will be made to select 200 users as the target users for analysis using the second preset ratio (20%), thus saving data resources while ensuring quality.

[0059] It should be noted that the blood oxygen characteristics include temporal characteristics and the rate of change of blood oxygen amplitude characteristics between adjacent monitoring times.

[0060] Understandably, it involves calculating the time delay between specific phase points of EEG slow-wave oscillations (such as the down-state start point and up-state peak point) and inflection points of fNIRS blood oxygenation signal changes (such as the HbO rise start point and peak point). Under normal physiological conditions, changes in neural activity precede blood oxygenation response, and this delay (usually 2-6 seconds) is a core indicator for assessing neurovascular coupling efficiency. Analyzing the cross-correlation or coherence between the EEG slow-wave envelope and the HbO / HbR concentration change curve within a single slow-wave oscillation cycle is crucial. In healthy deep sleep, both should exhibit high waveform synchronicity. Furthermore, by screening blood oxygenation characteristics during periods of blood oxygenation fluctuation, a foundation is laid for further improving the reliability of sleep staging identification for users with low accuracy in sleep staging.

[0061] Furthermore, the analysis of target user data includes the number of target users and their proportion among all users.

[0062] Furthermore, the analysis results of brain-computer signals during the reliable identification process of the target user are determined based on the changes in sleep stage characteristics during the reliable identification process of the target user.

[0063] Furthermore, the reliable identification process is the sleep staging process, and the staging result is an accurate sleep staging process.

[0064] Specifically, such as Figure 3 As shown, the method for determining the fusion analysis and processing strategy for the target user is as follows: This embodiment analyzes a target user group (i.e., users with high reliability in sleep staging using EEG features) to determine their fusion analysis strategy. Specifically, it determines when and during which sleep stages the combined analysis of EEG and near-infrared spectroscopy (fNIRS) features is necessary. The core logic is "setting the tone based on group size, mitigating risks based on pattern stability, and dynamically optimizing execution effectiveness." This method recognizes that even for highly reliable users, some sleep stages may still have potential staging bias risks due to physiological changes, signal interference, and other factors. By analyzing the pattern stability of users' sleep EEG signals, the method balances the reliability of blood oxygen feature extraction and identification bias based on differences in stability. While ensuring the auxiliary effect on low-reliability users and the efficiency of blood oxygen feature extraction, this method achieves optimal staging quality by accurately identifying and covering high-risk processes.

[0065] S21 uses the analyzed target user data to determine the proportion of the analyzed target user among the users; The percentage of target users in the overall user base: This refers to the proportion of target users with high EEG staging reliability verified through historical data to the total number of users on the platform. This proportion reflects the scale of the high-quality data base that can be used to extract robust EEG-blood oxygenation correlation models. Preset proportion: A lower threshold used to determine if the high-quality user base is too weak. Third preset proportion: A higher threshold used to determine if high-quality users have become the mainstream group on the platform. Second set proportion: An early warning threshold used to monitor the coverage of fusion analysis. When the actual fusion proportion is too low, it indicates that the current focusing strategy may have missed too many potentially risky sleep processes, requiring an expanded fusion scope to comprehensively eliminate biases.

[0066] This step is designed to establish the tone for risk control from a macro-level data foundation perspective. If highly reliable users are extremely few (ratio ≤ preset percentage), there will be a severe shortage of high-quality data available for modeling. In this case, a full fusion analysis of all sleep processes must be performed to build the most robust correlation model possible and eliminate biases to the greatest extent possible—a "comprehensive defense" strategy. If highly reliable users constitute the vast majority (ratio ≥ third preset percentage), it indicates that the platform possesses sufficient high-quality data to build a very robust model, allowing for the attempt of a "precise defense" strategy. This involves implementing fusion analysis only in the most typical and common sleep patterns (with the lowest stage risk), while closely monitoring the execution effect. The significance of this step lies in establishing a match between the risk control strategy and the data foundation capabilities. It ensures that the most conservative and comprehensive risk elimination plan is adopted when high-quality data is scarce, while a more refined plan is adopted when high-quality data is abundant, defining the scope for subsequent precise decision-making based on individual pattern risks.

[0067] The above steps also include the following: S211 Determine whether the proportion of the target user in the user group is greater than a preset proportion. If yes, proceed to step S212. If no, determine that the fusion analysis processing strategy for the target user is to perform fusion analysis processing in all sleep processes. S212 determines whether the proportion of the analysis target among users is not less than the third preset proportion. If so, the fusion analysis processing strategy for the analysis target user is determined to be to perform fusion analysis processing only in the combination of sleep processes with the most sleep processes. That is, if the sleep stage feature signal of the sleep process falls into the combination of sleep processes with the most sleep processes, then the fusion analysis processing of the sleep process is required. And when the proportion of sleep processes that have been fused in the most recent preset time period is not greater than the second preset proportion, then the fusion analysis processing of the sleep process will be performed in all sleep stages in the future preset time period. If not, proceed to step S22. Specific decision-making logic and examples Assume the platform has a total of 1000 users (U1-U1000). Through preliminary screening, 150 of them were identified as the target users for analysis.

[0068] The target user percentage is calculated as 150 / 1000 = 15%, with a preset percentage of 10% and a third preset percentage of 30%.

[0069] S211: Determine if the percentage is greater than the preset percentage: Determine: 15% > 10%, the condition is met, proceed to S212.

[0070] S212: Determine whether the percentage is not less than the third preset percentage: Judgment: 15% < 30%, the proportion is no greater than the third preset proportion. Decision: High-quality users have reached a certain scale (more than 10%), but have not yet become the mainstream (less than 30%). At this point, we cannot completely rely on high-quality data and relax risk control, nor should we indiscriminately merge all processes. We need to further examine the stability of individual user sleep patterns to accurately identify and cover high-risk processes, proceeding to step S22.

[0071] In another scenario: if the target users only account for 8% (≤10%), the direct decision is to perform a fusion analysis on all sleep processes of all target users. Because the foundation of high-quality data is weak, comprehensive fusion is necessary to build a usable model and minimize the risk of stage-based bias.

[0072] If the target user percentage is 35% (≥30%), the direct decision is to adopt a preliminary, precise defense strategy: only when the EEG characteristic signals during sleep fall into the combination of sleep processes with the largest number of users (representing the most stable and lowest-risk pattern) will fusion analysis be suspended (or only low-intensity fusion be performed). Simultaneously, monitoring is initiated: if the proportion of users actually triggering fusion analysis in the recent period (e.g., within a week) is ≤ a second predetermined proportion (e.g., 15%), then a broader fusion approach will be adopted in the next cycle to prevent overlooking potential risk processes outside the established patterns.

[0073] S22 Based on the analysis results of the brain-computer signals in the reliable identification process of the target user, the reliable identification processes that meet the similarity requirements of the sleep stage feature signals are divided into the same sleep process combination. Identifying reliable processes: This specifically refers to analyzing the sleep staging processes in the target user's history that have been verified as absolutely accurate. These processes are the gold standard for extracting reliable EEG-blood oxygenation correlation models and also represent "risk-free" or "low-risk" sleep patterns.

[0074] Sleep stage characteristic signal: EEG feature vectors extracted from the above reliable process that characterize the sleep stage.

[0075] Sleep process combination: Multiple reliable sleep process combinations that meet certain similarity requirements (such as dynamic time regularization distance less than a threshold) are grouped into the same set and called a sleep process combination. A user has multiple different sleep process combinations, representing multiple "reliable" sleep patterns existing under different physiological states.

[0076] Matching process combination: For a single target user, the combination containing the most reliable processes among all their sleep process combinations. This represents the user's most common and stable "low-risk" sleep EEG pattern.

[0077] Matching Factor Analysis: This metric quantifies the concentration of "low-risk" sleep patterns in a single target user. The formula is: (Number of sleep processes in the matching process combination) / (Total number of reliable processes identified by the user). A higher factor indicates that the user's reliable sleep processes are more concentrated in a specific "low-risk" pattern, suggesting relatively stable sleep characteristics.

[0078] This step is designed to assess the stability of sleep characteristics at the micro-individual sleep pattern level. Not all sleep processes of the target users are stable. Analyzing users with low matching factors indicates that their reliable sleep patterns are scattered, with multiple "low-risk" patterns, or a large number of unique and difficult-to-classify sleep processes. These scattered or unique sleep processes indicate unstable sleep characteristics; therefore, a fusion analysis of these processes is needed to eliminate risk.

[0079] The significance of this step lies in accurately identifying the sleep processes within each user that may have a staged deviation risk. It categorizes users' sleep processes into two types: "high-risk" (unstable sleep characteristics) and "low-risk" (stable sleep characteristics), providing a direct basis for implementing differentiated fusion analysis.

[0080] The above steps also include the following: S221 Determine whether the number of sleep process combinations for different target users is greater than the preset threshold for the number of process combinations. If yes, determine that the fusion analysis processing strategy for the target users is to perform fusion analysis processing in all sleep processes. If no, proceed to step S222. S222 Based on the number of different sleep process combinations of the target users, determine the sleep process combination with the most sleep processes among the sleep process combinations of the target users, and use it as the matching process combination of the target users in the phased analysis. Based on the proportion of sleep processes in the matching process combination of the target users, determine the analysis matching factor of the target users. Determine whether the average value of the analysis matching factor of different target users is less than the preset matching factor threshold. If so, determine that the fusion analysis processing strategy of the target users is to perform fusion analysis processing in all sleep processes. If not, proceed to step S23. Specific decision-making logic and examples (continued from S21): The platform analyzes the reliable identification process of 150 target users and forms a sleep process combination.

[0081] S221: Assess the degree of dispersion of user patterns: Judgment logic: Check whether the number of sleep process combinations for all target users is greater than the preset threshold for the number of process combinations (e.g., threshold = 4). If each user has many different sleep pattern combinations (a large number of combinations), it indicates that the user's internal sleep patterns are extremely variable and the sleep characteristics are not stable enough. This means that almost all users have unstable sleep characteristics and need to be widely integrated to avoid the technical problem of low extraction efficiency of blood glucose characteristics caused by using a single sleep process combination.

[0082] Example: Suppose that, statistically, 120 users have a sleep process combination count > 4, and 30 users have a count ≤ 4. Therefore, not "all users" meet the condition (> 4), so proceed to S222.

[0083] S222: Calculate and analyze matching factors and assess overall risk concentration. Calculation and Judgment: For each user, identify the combination of matching processes and calculate and analyze the matching factors.

[0084] The average of the matching factors for all 150 users was calculated. The average value reflects the concentration of the "low-risk" pattern in the overall user group with stable sleep characteristics. A low average value means that the sleep patterns of the group are dispersed as a whole, and the efficiency of extracting blood oxygen characteristics using a single combination is lower, requiring a large-scale fusion. A high average value means that there is a clear dominant "low-risk" pattern in the group, and their sleep characteristics are relatively stable, so the fusion of processes outside the pattern can be more targeted.

[0085] Example: Assume the calculated average analytical matching factor is 0.60. Set the preset matching factor threshold to 0.75.

[0086] Judgment: 0.60 < 0.75, the average value is less than the threshold, and the overall concentration is not high. This indicates that there are no highly clustered sleep process combinations within the high-quality user group, and the extraction efficiency of using a single sleep process combination is slow. Therefore, the fusion analysis processing strategy is determined to be to perform fusion analysis processing on all sleep processes to comprehensively eliminate risks.

[0087] Assuming another scenario (average value meets the standard): if the calculated average analysis matching factor is 0.80 (>0.75), it indicates that the overall "low-risk" mode has a good concentration. Therefore, it is possible to focus on a certain sleep process combination for blood oxygen feature extraction and processing, and proceed to step S23 for final user segmentation and strategy formulation.

[0088] S23 determines the fusion analysis and processing strategy for the target users based on the proportion of the target users among the users and the different combinations of sleep processes of the target users.

[0089] The above steps specifically include: S231 Based on the analysis matching factor of the analysis target user, determine whether the analysis matching factor of the analysis target user is greater than the preset matching factor threshold. If yes, determine that the fusion analysis processing strategy of the analysis target user is to perform fusion analysis processing in all sleep processes. If no, proceed to step S232. S232 determines whether the proportion of analysis target users whose matching factor is greater than the preset matching factor threshold among all analysis target users is greater than the preset proportion threshold. If so, it determines that the remaining analysis target users will only undergo fusion analysis processing in the sleep process combination with the most sleep processes. That is, if the sleep stage feature signal of the sleep process falls into the sleep process combination with the most sleep processes, then sleep process fusion analysis processing is required. And if the proportion of sleep processes that have undergone fusion analysis processing in the most recent preset time period is not greater than the second preset proportion, then sleep process fusion analysis processing will be carried out in all sleep stages in the future preset time period. If not, it determines that the remaining analysis target users will undergo fusion analysis processing in the sleep process combination with the most sleep processes. If the proportion of sleep processes that have undergone fusion analysis processing in the most recent preset time period is not greater than the set proportion, then sleep process fusion analysis processing will be carried out in all sleep stages in the future preset time period.

[0090] It should be noted that the second set ratio is less than the set ratio.

[0091] Preset percentage threshold: The standard used to stratify the risk among target users. This step distinguishes the size of the "high-risk" user group (with unstable sleep characteristics) to address the heterogeneity risk within the high-quality user group. Even if the overall average matching factor meets the standard, there may still be some "high-risk" users with low matching factors, i.e., users with unstable sleep characteristics. The goal of this step is refined stratification management: a comprehensive analysis strategy is used for "low-risk" users (with high matching factors), whose sleep characteristics are relatively stable; for "high-risk" users (with low matching factors) whose sleep characteristics are unstable, the control intensity is determined based on the group size—if "low-risk" users are a minority, a comprehensive analysis strategy is used for them; if "low-risk" users constitute a significant proportion, a focused analysis strategy is required.

[0092] Its significance lies in achieving precise and differentiated risk control. It avoids a "one-size-fits-all" approach to all high-quality users, instead allocating appropriate fusion strength based on the stability of each user's sleep characteristics (reflected through analysis of matching factors), and embedding effect monitoring and dynamic adjustment mechanisms during execution to ensure the reliability of blood oxygen feature extraction and balanced control of extraction efficiency.

[0093] Specific decision-making logic and final example (continuing from S22 "average target met" scenario): Assuming the average matching factor for 150 users is 0.80 (> the threshold of 0.75), we need to stratify each user.

[0094] S231: Initial risk stratification for individual users: Judgment logic: Iterate through each target user and determine whether its analysis matching factor is greater than the preset matching factor threshold (0.75). Based on the threshold, the user is directly divided into two categories: "low-risk" users (matching factor > 0.75) and "high-risk" users (matching factor ≤ 0.75).

[0095] S232: Determine the final strategy based on the proportion of "low-risk" users. Judgment logic: Calculate the proportion of "low-risk" users (matching factor ≤ 0.75) among all target users in the analysis.

[0096] If the proportion of "low-risk" users is relatively high, it means that even when the overall data is good, the extraction efficiency of blood oxygenation features is high, and a focused extraction strategy can be adopted for high-risk users.

[0097] Example: Suppose that out of 150 users, 110 users have an analysis matching factor greater than 0.75 ("low-risk" users), accounting for 110 / 150≈73.3%. Assuming a preset percentage threshold of 70%, the judgment is: 73.3% > 70%, the percentage of "low-risk" users is greater than the threshold.

[0098] Final decision: For the 110 "low-risk" users with a matching factor > 0.75: adopt a comprehensive extraction strategy, that is, perform fusion analysis in all sleep processes.

[0099] For the 40 "high-risk" users with a matching factor ≤ 0.75: Due to the high risk and considerable size of this group, a monitoring-based precise defense strategy will be adopted, meaning that fusion analysis will only be performed when their sleep process signals do not fall into the matching process combination. However, stricter monitoring will be initiated at the same time: if the proportion of sleep processes processed by fusion analysis in the most recent preset period does not exceed a second preset proportion (e.g., 15%), then full fusion will be switched in the future period.

[0100] This method constructs an adaptive fusion analysis strategy decision-making system with the core objective of eliminating staging bias risk. Risk-oriented: all strategies are formulated to improve the reliability of staging results and the efficiency of blood oxygen feature extraction. Through a three-level analysis of "group size - individual pattern - user stratification," it achieves refined risk assessment from macro to micro and from group to individual. It can distinguish between users and sleep processes with "high sleep feature stability" and "low sleep feature stability," and apply differentiated fusion strengths. The strategy incorporates effect monitoring and dynamic adjustment mechanisms (triggered by thresholds such as a second set ratio and a set ratio) to ensure that the strategy can self-correct based on actual execution results (fusion coverage), preventing insufficient risk elimination or resource waste due to overly conservative or aggressive strategies.

[0101] This method ultimately ensures that when using high-reliability user data to build an EEG-blood oxygenation correlation model and applying it to assist low-reliability users, the data foundation itself has been rigorously screened and risk-excluded, thereby maximizing the robustness and accuracy of the entire sleep staging system.

[0102] It is understandable that the efficiency of the extraction and processing of the blood oxygenation features of the target user is not up to standard, specifically including: The fundamental goal of this method is to evaluate whether the "processing efficiency" of extracting blood oxygen features from the target user group meets the requirements, thereby determining whether it is necessary to expand the feature extraction scope to ordinary users. The core logic is to "measure efficiency by the breadth of pattern coverage and determine expansion needs by a threshold." True extraction efficiency is reflected not only in the number of features extracted but also in the breadth of user coverage of the extracted feature patterns. By performing coverage analysis on the existing blood oxygen feature combinations, we determine whether the blood oxygen feature patterns provided by the current target user group are sufficiently representative. If each feature pattern is matched by only a very small number of target users, it indicates low feature extraction efficiency, requiring an expansion of the collection scope; conversely, if there are common feature patterns matched by a large number of target users, it indicates that the current extraction is already efficient, and there is no need to expand the collection scope to ordinary users.

[0103] Based on the similarity of blood oxygen features among different target users, blood oxygen features that meet the similarity coefficient requirements are grouped into the same blood oxygen feature combination. The system identifies target users whose sleep processes match different combinations of blood oxygenation features. Based on the number of target users whose sleep processes match different combinations of blood oxygenation features, it determines whether the efficiency of extracting and processing the blood oxygenation features of the target users meets the requirements.

[0104] It is understandable that if the number of target users whose analysis results in matching sleep processes (i.e., sleep processes in which blood oxygen features fall within the blood oxygen feature combination) is less than the target user number threshold, then the efficiency of the blood oxygen feature extraction process for identifying target users is determined to be insufficient.

[0105] Assume the platform has 150 target users. Blood oxygenation features are extracted from their historical data and clustered to form 15 blood oxygenation feature combinations (C1-C15). Set the target user number threshold to 20.

[0106] Check criteria: Whether the number of matching users for all blood oxygen feature combinations is less than the target user number threshold (20). The number of matched users for the three combinations C1(35>20), C4(45>20), and C15(22>20) is greater than the threshold. Therefore, not all combinations are less than the threshold.

[0107] Conclusion: The extraction and processing efficiency meets the requirements. The presence of multiple feature combinations (C1, C4, C15) matched by more than 20 (threshold) target users indicates that these blood oxygenation patterns are quite prevalent and representative among high-quality user groups.

[0108] Specifically, when the efficiency of the extraction and processing of the blood oxygen features of the target user meets the requirements, it is not necessary to extract and process the blood oxygen features during the inflection point period of the user's blood oxygen signal change.

[0109] Specifically, the method for determining the extraction strategy of blood oxygenation features during the inflection point period of the user's blood oxygenation signal change is as follows: The fundamental goal of this method is to form reliable, standardized blood oxygen saturation feature templates through cluster analysis, based on high-quality blood oxygen saturation features extracted from target users (high-reliability users). Using these templates, the method determines whether and how to extract blood oxygen saturation features from ordinary users, thereby building a richer and more reliable blood oxygen saturation feature reference library for identifying users with sleep delinquencies. The core logic is: "Using reliable user features as a benchmark, forming standard templates through clustering; verifying template reliability through statistical reproducibility; and expanding data collection under the premise of risk control." The method first uses high-quality blood oxygen saturation features from target users for clustering to form candidate feature templates; then, it uses rigorous statistical tests to screen out highly reproducible and reliable templates; finally, based on the maturity of the template library and the individual stage risk of ordinary users, it determines whether to extract features from ordinary users and what data collection strategy to adopt. The entire process ensures that the blood oxygen saturation feature references provided for identifying users with sleep delinquencies are highly reliable and representative.

[0110] S31 classifies blood oxygen features that meet the similarity requirements among different target users into the same blood oxygen feature combination based on the degree of similarity of their blood oxygen features. It should be noted that the similarity coefficient is determined based on the average of the similarity coefficients of the temporal characteristics between blood oxygen features and the rate of change of blood oxygen amplitude characteristics between adjacent monitoring times.

[0111] It should be noted that the users mentioned are those who are neither identified as biased users nor identified as target users.

[0112] Blood oxygen feature combination: This refers to clustering all blood oxygen change inflection point features extracted from the target users (high-reliability users) according to time-series characteristics (such as inflection point waveform shape, upward / downward trend) and the similarity of blood oxygen amplitude change rate sequences between adjacent monitoring times. Features with similarity coefficients that meet a set threshold are grouped into the same combination. Each combination represents a representative "standard template" of blood oxygen change extracted from high-quality data.

[0113] Similarity coefficient: A comprehensive index measuring the degree of similarity between two blood oxygenation characteristics. It is calculated as a weighted average of temporal feature similarity (such as the normalized inverse of dynamic time-normalized distance) and amplitude variation rate sequence similarity (such as cosine similarity). This coefficient focuses on the dynamic pattern of blood oxygenation changes rather than absolute amplitude, and can effectively standardize individual differences.

[0114] This step is designed to extract representative standard patterns of blood oxygenation changes from limited but high-quality data of the target users. The staging of the target users is highly reliable, and their corresponding blood oxygenation features are of relatively high quality with a good signal-to-noise ratio. By clustering these high-quality features, we can uncover universally applicable prototypes of blood oxygenation changes associated with stable sleep stages. These prototypes will serve as benchmark templates for judging whether the blood oxygenation features of ordinary users are "normal" or "reliable."

[0115] The significance of this step lies in establishing a standardized feature reference system based on high-quality data. It elevates high-quality features extracted from a small number of reliable users into a series of standard templates using pattern recognition methods, providing an objective basis for subsequently finding similar features among ordinary users and evaluating feature reliability. This is the first and most important step in building a reliable blood oxygen feature database.

[0116] Specific examples: Assume the platform has extracted blood oxygenation inflection point features from the historical data of 150 target users, specifically the 8-second intervals before and after each user's transition from the lucid phase to N1, resulting in 5000 high-quality blood oxygenation features. The system calculates the similarity coefficients between each pair of these 5000 features and uses a clustering algorithm (such as DBSCAN) to divide them into 25 different blood oxygenation feature combinations, such as "Combination T1: rapid rise followed by slow decline" and "Combination T2: step-like rise." Features within each combination exhibit highly similar morphologies and dynamic change patterns.

[0117] S32 determines the reliable blood oxygen feature combination in the blood oxygen feature combination based on the extracted data of blood oxygen feature in different blood oxygen feature combinations, and determines the number of recognition deviations of the user's sleep stage results based on the recognition processing results of the user's brain-computer interface data. It should be noted that the reliable blood oxygen feature combination is a blood oxygen feature combination in which the number of sleep processes from which blood oxygen features are extracted is greater than a preset threshold for the number of sleep processes.

[0118] A reliable combination of blood oxygenation features refers to a combination of blood oxygenation features in which the number of independent sleep processes from which the blood oxygenation features originate is greater than a preset threshold (e.g., 80 sleep processes). This requires that the template be observed in a large number of independent, high-quality sleep events to ensure that it has sufficient statistical reproducibility and is not a random pattern of a few users.

[0119] Preset threshold for the number of trusted feature combinations: This determines whether the reliable template library built based on the analysis of target users has sufficient diversity (e.g., 8 combinations). Reaching this threshold means that a relatively complete standard template library covering multiple situations has been formed.

[0120] Potential identification feature combinations: These refer to combinations that, while not reaching the "credible" standard, exceed a lower second preset threshold (e.g., 30). These templates have shown some reproducibility and have the potential to develop into reliable templates. Preset potential proportion threshold: This is the standard for assessing whether "potential templates" occupy the mainstream direction among all templates (e.g., 35%).

[0121] This step is set up to validate the statistical power of standard templates extracted from high-quality data. Even if a template comes from reliable users, if it is supported by only a very small number of sleep processes, it may only reflect the specificity of individual users rather than a general pattern. By setting a high recurrence threshold, we can filter out those patterns that appear repeatedly in the reliable user group and have high consistency. These patterns are truly qualified as "gold standard" templates and can be safely used to guide feature extraction for ordinary users and as a reference for identifying biased users.

[0122] Its significance lies in ensuring the statistical robustness of the standard template library. It prevents individual phenomena from being mistaken for general laws, ensuring high confidence in subsequent feature matching and reference applications based on these templates. This is a crucial step in transforming "patterns extracted from high-quality data" into "reliable knowledge that can be generalized and applied."

[0123] It is understood that the above steps include the following: S321 Determine whether there is a reliable blood oxygen feature combination. If yes, proceed to step S322. If no, a reliable blood oxygen feature combination cannot be determined at this time. In order to avoid interference from the blood oxygen feature extraction results of users with low accuracy of sleep staging, it is determined that all users do not need to perform blood oxygen feature extraction processing during the inflection point period of blood oxygen signal change. S322 determines whether the number of the reliable blood oxygen feature combinations is greater than the preset threshold for the number of reliable feature combinations. If yes, then for all users, the blood oxygen feature extraction process for the inflection point period of blood oxygen signal change is performed according to the preset extraction strategy. If no, then proceed to step S323. It is understood that the preset extraction strategy requires sleep process fusion analysis, i.e. blood oxygen feature extraction, only when the sleep stage feature signal of the sleep process falls into the combination of sleep processes with the most users.

[0124] S323 uses the number of sleep processes extracted from different blood oxygen feature combinations to identify blood oxygen feature combinations with a number of sleep processes greater than a second preset threshold as potential identification feature combinations. It then determines whether the proportion of the potential identification feature combination in the blood oxygen feature combination is greater than a preset potential proportion threshold. If so, it performs blood oxygen feature extraction processing for all users during the inflection point period of blood oxygen signal change according to the preset extraction strategy. If not, it proceeds to step S33. Specific decision-making logic and examples: Continuing from the previous example, the system has generated 25 blood oxygenation characteristic combinations (T1-T25). The preset threshold for the number of sleep processes is set at 80.

[0125] S321: Determine if a reliable combination of blood oxygenation characteristics exists: Check: Count the number of independent sleep processes contained in each combination. Assume that the number of combinations T3 (150 times), T7 (120 times), T11 (95 times), T15 (85 times), and T20 (100 times) exceeds 80.

[0126] Result: A reliable combination of blood oxygenation characteristics {T3, T7, T11, T15, T20} exists. Proceed to S322.

[0127] If not found: This indicates that even from high-quality user data, a pattern with sufficient statistical reproducibility could not be extracted. This may mean that the correlation between blood oxygenation signals and sleep stages is very weak or unstable in the current data. In this case, feature extraction for ordinary users would lack reliable evaluation criteria and easily introduce noise. The decision is: do not perform blood oxygenation feature extraction on any ordinary users to avoid building an unreliable feature library.

[0128] S322: Determine if the number of trusted templates is sufficient: Judgment: The number of trustworthy templates is 5. Set the preset threshold for the number of trustworthy feature combinations to 4.

[0129] Result: 5 is greater than 4, indicating reliable data. This means that the current reliable template library may cover a sufficient diversity of sleep transition scenarios. Therefore, based on this, we can directly extract data from ordinary users to determine whether the blood oxygen features in the above scenarios are truly reliable. Assuming that the preset threshold for the number of reliable feature combinations is 6, we proceed to S323.

[0130] S323: Assess the development trend of potential templates: Judgment: Let the second preset threshold be 30. Check the remaining combinations. Assume that 10 combinations, namely T1 (65 times), T5 (50 times), T8 (45 times), T12 (40 times), T18 (35 times), and T22 (32 times), meet the conditions and are potential identification feature combinations. Potential combination percentage = 10 / 25 = 40%.

[0131] Judgment: Assume the preset potential percentage threshold is 35%. 40% > 35%, so the condition is met.

[0132] Results: Although the number of absolutely reliable templates is insufficient, a considerable proportion of templates show the potential to become reliable templates (with reproducibility rates between 30 and 80). This indicates that continued data accumulation is expected to upgrade more templates to reliable ones. The decision is: for all ordinary users, initiate blood oxygen feature extraction according to the preset extraction strategy to accelerate the maturation and enrichment of the template library. If the proportion is insufficient, it indicates that most of the patterns extracted from the target user data currently being analyzed have very low reproducibility and lack development potential. In this case, large-scale data collection is not advisable. Proceed to S33 and adopt a more conservative, individualized data collection strategy.

[0133] S33 determines the blood oxygen feature extraction strategy for the inflection point period of the user's blood oxygen signal change based on the number of recognition deviations and the reliable blood oxygen feature combination data in the blood oxygen feature combination.

[0134] It is understood that the above steps are based on the number of recognition errors of the user. If the number of recognition errors of the user does not meet the requirements, the risk of recognition error is relatively high. Therefore, it is determined that the user does not need to perform blood oxygen feature extraction processing during the inflection point period of blood oxygen signal change. Otherwise, the blood oxygen feature extraction processing during the inflection point period of blood oxygen signal change is performed for the user according to the preset extraction strategy.

[0135] Refined data collection based on template library maturity and user risk: Preset extraction strategy: This refers to a feature acquisition rule with controllable risk. Specifically, blood oxygenation inflection point features are extracted only when the EEG stage feature signal of a certain sleep process falls into the combination of sleep processes most frequently encountered by the user (representing their most common and stable sleep EEG patterns). The assumption of this strategy is that in the most stable EEG pattern, the accompanying blood oxygenation response is most likely to match the standard template extracted from the target user, thereby acquiring high-quality features that can be used to strengthen the template library.

[0136] This step is designed to continue collecting data from ordinary users in the safest and most efficient way to enrich and validate the template library, even before the reliable template library is fully mature. By screening ordinary users (collecting only data from their most stable sleep patterns), the matching probability of new features with existing standard templates can be maximized, improving data utilization. Simultaneously, it avoids introducing noise from collecting data in unstable patterns, which could contaminate the feature library under construction. Its significance lies in achieving "precise targeting" and "risk isolation" in data collection. It ensures that during the template library construction phase, every piece of new data is used to strengthen existing reliable patterns or validate potential patterns, rather than being scattered across a large number of uncontrollable explorations. This is crucial for efficiently building a clean and reliable blood oxygen feature reference library.

[0137] Specific decision-making logic and final example: Following the results in S32, assuming the potential combination percentage is 30% (<35%), the global activation condition is not met, and the process proceeds to S33 for individualized decision-making. Let the threshold for determining that a regular user's "number of identification deviations does not meet the requirements" be: more than 12 deviations in the last 50 periods.

[0138] The decision-making process for ordinary user U_P is as follows: Obtain the number of identification deviations for U_P: Query its historical installment records. Assuming that U_P has 6 deviations in the last 50 installments, the individual installment risk is judged as follows: 6 times < 12 times, which meets the requirements (i.e., the user's own installment results are relatively reliable and the data quality is guaranteed to a certain extent).

[0139] Final decision: For user U_P, blood oxygen feature extraction is performed according to the preset extraction strategy. That is, in U_P's subsequent sleep monitoring, the system will only trigger blood oxygen inflection point feature extraction when the EEG features of a certain sleep cycle fall into its most stable and common sleep pattern. The extracted features will then be matched with the existing standard template library to update the template's statistical information or discover clues to new patterns.

[0140] For another ordinary user, U_Q: If the number of identification errors is 18 (>12), then its own staging reliability is judged to be low and the risk of data noise is high. The decision is: U_Q will not undergo blood oxygen feature extraction to prevent its low-quality data from interfering with the construction process of a reliable feature library.

[0141] This method constructs a progressive data acquisition strategy decision-making system with the core objective of building a highly reliable blood oxygen feature reference library for identifying users with biases. The system is reliable from the outset: all analysis and decisions begin with high-quality blood oxygen features from the target users, ensuring a high signal-to-noise ratio and physiological relevance of the benchmark data. Through a three-level progressive mechanism of "high-quality feature clustering → statistical reproducibility verification → data acquisition strategy risk control," it ensures that every blood oxygen feature template and newly added feature in the final database undergoes rigorous reliability screening, greatly improving the overall quality of the feature library. The system can intelligently adjust the data acquisition strategy for ordinary users (from prohibiting acquisition, to individualized conditional acquisition, and then to global conditional acquisition) based on the maturity of the template library extracted from reliable users (from scarce to abundant), always promoting the expansion and improvement of the reliable feature library in the most stable and efficient way.

[0142] This method ultimately ensures that the blood oxygen feature reference library built to assist in identifying users with deviations is based on the most reliable source data, the most rigorous statistical tests, and the most prudent expansion strategies. This not only provides a high-confidence auxiliary reference for identifying users with deviations in sleep staging, but also provides a scalable and iterative reliable feature discovery and application framework for the continuous optimization and knowledge accumulation of the entire sleep staging system.

[0143] Specifically, the method for determining the update processing target of the target user in the analysis is as follows: During system operation, some high-performing ordinary users are dynamically updated as new analysis target users to expand the data source of high-quality EEG-blood oxygenation features. This allows for a more efficient construction and improvement of a reliable blood oxygenation feature reference library, ultimately serving to enhance the staging accuracy of users with identification deviations. The core logic is "setting the tone based on the severity of system problems, using blood oxygenation pattern matching as the core, and dynamically adjusting the entry threshold." When the proportion of users with identification deviations is too high, it indicates an urgent need for more reliable data, and the strategy tends to actively absorb more data (lowering the threshold); conversely, strict screening is adopted (raising the threshold). The core basis for screening is the matching depth and breadth of ordinary users' blood oxygenation features with existing reliable blood oxygenation feature combinations (reliable templates extracted from the original analysis target users). In this way, the system can continuously and safely expand its core high-quality user group like a "snowball."

[0144] S41 uses the aforementioned user data with identification deviation to determine the identification deviation ratio; It is understood that the above steps include the following: If the identification deviation ratio is greater than the preset deviation ratio threshold in the above steps, all users with matching reliable feature combinations will be used as the target users for update processing. Otherwise, the process will proceed to step S42.

[0145] It should be noted that users whose blood oxygenation features fall within the number of sleep processes in the trusted blood oxygenation feature combination and whose number of processes exceeds a preset value are considered as matched users. If the number of matched users in the trusted blood oxygenation feature combination exceeds a preset threshold for the number of matched users, then the blood oxygenation features of the trusted blood oxygenation feature combination are used as identification features, and combined with EEG signal features, a fusion analysis is performed to identify the sleep stages of users with identification deviations.

[0146] Identification Deviation Ratio: This refers to the proportion of users with identification deviations (users whose installment results are unreliable) out of the total number of users. This is a global indicator that measures the severity of the installment quality problems currently facing the system.

[0147] Preset deviation ratio threshold: A critical value used to trigger "emergency response mode". Exceeding this threshold indicates that the system's auxiliary capabilities face significant challenges.

[0148] Matching users: These are ordinary users whose blood oxygenation characteristics fall into a reliable combination of blood oxygenation characteristics, and whose corresponding number of sleep processes in that combination is greater than a preset value. This ensures that the user not only matches occasionally, but also has a certain statistical reproducibility on this reliable pattern.

[0149] This step is set up to determine the aggressiveness of the update strategy based on the overall health of the system. When identification bias is widespread (high proportion), it means that the existing feature library provided by the target users may be insufficient to cover or solve the current problem, and there is an urgent need to expand high-quality data sources. At this time, a more proactive strategy should be adopted, relaxing the criteria and widely recruiting any users who perform well on existing reliable templates to quickly expand the team of "reliable data producers." This is a problem-driven resource mobilization.

[0150] The significance of this step lies in giving the system strategy a holistic perspective and a sense of urgency. It links micro-level user update decisions with macro-level system performance indicators, enabling the activation of a "green channel" when the system encounters significant challenges, accelerating the construction of a reliable data ecosystem, and demonstrating the system's adaptive and self-enhancing capabilities.

[0151] Specific decision-making logic and examples: Assume the platform has 1000 users, of whom 80 have identification biases.

[0152] Identification deviation ratio = 8%.

[0153] Set the preset deviation ratio threshold to 5%.

[0154] Judgment: 8% > 5%, the identification deviation rate exceeds the threshold, indicating a serious systemic bias problem. An active inclusion strategy should be initiated immediately. All matching users on the platform (i.e., ordinary users whose blood oxygenation patterns are highly consistent with existing reliable templates and are stable) will be listed as target users for analysis and updated processing. The sleep data of these newly added target users will be directly used for high-quality blood oxygenation feature extraction and model enhancement, aiming to quickly improve the ability to assist users with identification deviations.

[0155] S42 determines the similarity of blood oxygen features with different target users based on the extraction results of the user's blood oxygen features, determines the overlap with credible blood oxygen feature combinations based on the similarity, determines the credible blood oxygen feature combinations to which the user's blood oxygen features fall based on the overlap, and uses them as matching credible feature combinations. It is understood that the above steps include the following: S421 Determine whether the number of matching reliable feature combinations of the user is greater than the preset threshold for the number of matching feature combinations. If yes, then the user is used as the target for updating the analysis target user. If no, proceed to step S422. S422 determines whether there exists a matching trusted feature combination with a number of sleep processes greater than a preset value based on the number of sleep processes of the user in different matching trusted feature combinations. If yes, proceed to step S423; otherwise, determine that the user does not belong to the update processing target of the analysis target user. S423 uses the matching reliable feature combination with a number of sleep processes greater than the preset value of the number of processes as the filtering combination, and determines whether the number of the filtering combination is greater than the preset threshold of the number of filtering combinations. If yes, it is determined that the user belongs to the update processing target of the analysis target user. If no, proceed to step S43.

[0156] S43 determines whether the user is the target for update processing of the target user based on the identification deviation ratio and the matching reliable feature combination.

[0157] Matching reliable feature combinations: For a given user, the reliable combinations of blood oxygenation features that their blood oxygenation characteristics can fall into. This represents which reliable templates match the user's blood oxygenation pattern.

[0158] Preset threshold for the number of matching feature combinations: a standard that requires users to match at least a number of different reliable templates to assess the breadth of their blood oxygenation patterns.

[0159] Process Count Preset: Within a single matching combination, the standard for the number of sleep processes that the user is required to contribute, used to evaluate the depth (stability) of their matching on that pattern.

[0160] Filtering combination: refers to the combination of reliable matching features that the user not only matches, but also contributes more than the preset value of the number of sleep processes.

[0161] Preset filter combination quantity threshold: requires users to have a quantity standard of deep matching on at least multiple reliable templates.

[0162] Combination Quantity Threshold (Dynamic): In S43, this is a threshold dynamically calculated based on the recognition deviation ratio. The higher the ratio, the lower this threshold, meaning a more lenient entry threshold.

[0163] When the overall system deviation is within a controllable range, the update strategy should focus on ensuring the quality of new members and avoiding diluting the high reliability of the original target user group. Therefore, a progressive screening standard from "breadth matching" to "depth matching" was established: Breadth (S421): Users must match multiple reliable templates to demonstrate the diversity of their blood oxygenation patterns, rather than a single coincidence. Depth (S422-S423): Users must have a sufficient number of successful cases on multiple templates to demonstrate that their matching is stable and repeatable, rather than accidental. Dynamic calibration (S43): The final admission threshold is fine-tuned according to the overall system status, maintaining flexibility while remaining robust. Its significance lies in establishing a rigorous "data quality" admission mechanism. It ensures that only ordinary users whose blood oxygenation signals are highly consistent with the existing reliable knowledge base and whose behavior is stable are eligible to become core data contributors. This maintains the brand value of the "analysis target user" group—that is, extremely high data reliability, which is the cornerstone of the system's credibility.

[0164] It is understood that in the above steps, based on the identification deviation ratio, a threshold for the number of matching trustworthy feature combinations is determined, and when the number of matching trustworthy feature combinations of a user is greater than the threshold for the number of combinations, the user is determined to belong to the update processing target of the analysis target user.

[0165] It should be noted that the higher the identification deviation ratio, the smaller the threshold for the number of combinations.

[0166] Specific decision-making logic and examples: Assuming a 10% identification bias, proceed to the fine-tuning process. Examine ordinary user U_Z.

[0167] S42: Fine-tuning process: Background: The blood oxygenation characteristics of U_Z were compared and found to fall into three different reliable combinations of blood oxygenation characteristics: combination A, combination B, and combination C.

[0168] S421: Check the matching breadth: The number of matching combinations is 3. The preset threshold for the number of matching feature combinations is 4.

[0169] Judgment: 3 < 4, not satisfied. Proceed to S422.

[0170] S422 & S423: Check match depth: Count the number of sleep processes in each matching combination for U_Z: combination A (12 times), combination B (5 times), combination C (8 times). Assume the preset value for the number of processes is 10.

[0171] Judgment: Only combination A satisfies "number of sleep processes > 10" (12 > 10). Therefore, the selected combination is {A}.

[0172] Number of filter combinations = 1. Set the preset threshold for the number of filter combinations to 2.

[0173] Judgment: 1 < 2, not satisfied. Proceed to S43.

[0174] S43: Final decision based on dynamic threshold: Premise: The current system's identification error rate is known to be 10%.

[0175] Dynamic rule: Set the dynamic combination quantity threshold as: Initial value - (Identification deviation ratio × Adjustment coefficient). For example, if the initial value is 5 and the adjustment coefficient is 20, then the current threshold is 5 - (10% × 20) = 3.

[0176] Judgment: The number of reliable feature combinations for user U_Z is 3, which is equal to the dynamic threshold of 3.

[0177] Final decision: User U_Z is determined not to be an update processing target for the target user in the analysis.

[0178] This method constructs a dynamic, dual-mode target user update mechanism. The system can keenly perceive the overall phase quality (by identifying the deviation ratio) and switch update strategies between "crisis mode" and "robust mode". In the fine screening mode, through the dual verification of "breadth + depth", it ensures that the blood oxygenation mode of newly added target users not only "similar" to the reliable template, but also reaches the high standard of "similarity" and "stability", effectively maintaining the data purity of the core user group. The dynamic threshold mechanism in S43 allows the admission threshold to be fine-tuned according to the real-time performance of the system, realizing the flexibility and adaptability of the strategy, and enabling the system to always maintain the best balance between "expanding data sources" and "ensuring data quality".

[0179] This method ultimately enables the organic updating of the target user group. It is no longer static but can be dynamically updated based on system needs and data performance. This not only accelerates the enrichment and improvement of the reliable blood oxygen feature database, providing a more powerful and diverse reference for identifying users with deviations, but also gives the entire sleep staging system continuous self-optimization and evolution capabilities.

[0180] Example 2 In a second aspect, the present invention provides a computer device comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the aforementioned near-infrared functional fusion method based on a brain-computer interface when running the computer program.

[0181] Example 3 Thirdly, such as Figure 4 As shown, this application provides a near-infrared functional fusion system based on a brain-computer interface, employing the aforementioned near-infrared functional fusion method based on a brain-computer interface, specifically including: The analysis and processing module is responsible for determining the fusion analysis and processing strategy for the target users. The extraction module is responsible for determining the extraction strategy for blood oxygenation characteristics during the inflection point period of the user's blood oxygenation signal changes; The target identification module is responsible for determining the update processing targets of the target users being analyzed.

[0182] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0183] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0184] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A near-infrared functional fusion method based on brain-computer interface, characterized in that, Specifically, it includes: Using the sleep stage recognition deviation data obtained from brain-computer interface data, the target user for analysis is determined. Based on the target user data and the analysis results of brain-computer signals in the reliable recognition process of the target user, the fusion analysis processing strategy of combining the target user with near-infrared function is determined. The reliable identification process is the sleep staging process, and the staging result is an accurate sleep staging process. Based on the aforementioned fusion analysis and processing strategy, the blood oxygen characteristics of the inflection point period of the infrared signal change are determined. When the blood oxygen pattern of the target user is not representative, the extraction strategy of the blood oxygen characteristics of the user's blood oxygen signal change inflection point period is determined based on the similarity of blood oxygen characteristics among different target users and in combination with the recognition and processing results of the user's brain-computer interface data. The blood oxygen characteristics of the user are extracted using the extraction strategy. Based on the results of the blood oxygen characteristics extraction and the identification bias user data, users are identified as new target users for analysis.

2. The near-infrared functional fusion method based on brain-computer interface as described in claim 1, characterized in that, The recognition deviation data includes the number of recognition deviations made by the user.

3. The near-infrared functional fusion method based on brain-computer interface as described in claim 1, characterized in that, The method for determining the target users of the analysis is as follows: Based on the identification deviation data of different users' sleep stages, users whose identification deviation number exceeds the preset deviation number threshold are identified and regarded as users with identification deviation. The identification deviation ratio is determined based on the proportion of users with identification deviations among those undergoing sleep staging. Based on the identification deviation ratio and the number of identification deviations for different users, the target users for analysis are determined among the users.

4. The near-infrared functional fusion method based on brain-computer interface as described in claim 3, characterized in that, If there are no users with identification bias, then a preset proportion of users are selected as the target users for analysis. That is, the user quantity constraint is constructed based on the preset proportion of users, and the preset proportion of users with the fewest identification biases are selected as the target users for analysis.

5. The near-infrared functional fusion method based on brain-computer interface as described in claim 3, characterized in that, When the user belongs to the target user of the analysis, the blood oxygen characteristics of the blood oxygen fluctuation period obtained by near-infrared functional monitoring corresponding to the sleep stage characteristics are determined based on the user's sleep stage characteristics, that is, the blood oxygen characteristics of the blood oxygen signal inflection point period.

6. The near-infrared functional fusion method based on brain-computer interface as described in claim 1, characterized in that, The blood oxygen characteristics include temporal characteristics and the rate of change of blood oxygen amplitude characteristics between adjacent monitoring times.

7. The near-infrared functional fusion method based on brain-computer interface as described in claim 1, characterized in that, The method for determining the fusion analysis strategy for the target user combined with near-infrared function is as follows: Based on the analyzed target user data, determine the proportion of the analyzed target user among the users; Based on the analysis results of brain-computer signals in the reliable identification process of the target user, the reliable identification processes that meet the similarity requirements of sleep stage feature signals are grouped into the same sleep process combination. Based on the proportion of the target users in the user base and the different sleep process combinations of the target users, a fusion analysis strategy is determined for the target users to be analyzed using near-infrared spectroscopy.

8. The near-infrared functional fusion method based on brain-computer interface as described in claim 1, characterized in that, The determination that the blood oxygenation pattern of the target user in the analysis is not representative includes: Based on the similarity of blood oxygenation features among different target users, blood oxygenation features that meet the similarity coefficient requirements are grouped into the same blood oxygenation feature combination. The analysis target users are identified as having matching sleep processes across different combinations of blood oxygenation characteristics. Based on the number of analysis target users with matching sleep processes across different combinations of blood oxygenation characteristics, the representativeness of the blood oxygenation patterns of the analysis target users is determined.

9. A computer device, comprising: A memory and processor connected by communication, and a computer program stored in the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes a near-infrared functional fusion method based on a brain-computer interface as described in any one of claims 1-8.

10. A near-infrared functional fusion system based on a brain-computer interface, employing the near-infrared functional fusion method based on a brain-computer interface as described in any one of claims 1-8, characterized in that, Specifically, it includes: The analysis and processing module is responsible for determining the fusion analysis and processing strategy for the target users. The extraction module is responsible for determining the extraction strategy for blood oxygenation characteristics during the inflection point period of the user's blood oxygenation signal changes; The target identification module is responsible for determining the update processing targets of the target users being analyzed.

Images