A Multimodal Monitoring-Based Early Warning Method for Critical Neurosurgical Complications

By analyzing multimodal monitoring data, establishing dynamic coupling relationships and propagation verification mechanisms, the problem of insufficient adaptability to individual differences in the condition and changes in the disease among critically ill neurosurgical patients was solved, and more accurate early warning of complication risks was achieved.

CN123120175APending Publication Date: 2026-11-17XINYI CITY PEOPLES HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611576447.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are difficult to effectively adapt to the individual differences in the condition and dynamic changes in the illness of critically ill neurosurgical patients, resulting in inaccurate early warning of complication risks during monitoring.

Method used

By using a multimodal monitoring-based early warning method for critical neurosurgical complications, the state benchmark and initial degradation characteristics of each monitoring modality are determined, a dynamic coupling relationship is established, and the method is validated and updated by combining propagation contribution and prediction error to generate early warning information for complications.

Benefits of technology

It improves adaptability to individual differences in the condition and dynamic changes in the disease of critically ill neurosurgical patients, and enhances the accuracy and timeliness of complication risk prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120175A_ABST
    Figure CN123120175A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of medical monitoring data processing technology and provides a method for early warning of neurosurgical critical illness complications based on multimodal monitoring. The method includes determining the state baseline and initial degradation characteristics of each monitoring modality based on the patient's initial state information and multimodal monitoring data; determining candidate propagation relationships and grouping them based on state changes in different monitoring modalities to obtain dynamic coupling relationships; determining coupling degradation characteristics and state prediction characteristics of each propagation group based on propagation contribution; determining the verification results of the propagation groups by combining actual state characteristics and prediction errors, and updating the dynamic coupling relationships and coupling degradation characteristics accordingly; determining abnormal evolution characteristics based on the updated dynamic coupling relationships and coupling degradation characteristics; obtaining target complication risk prediction results by combining initial degradation characteristics and generating early warning information. This invention can improve the adaptability of the early warning process for neurosurgical critical illness complications to individual patient state differences and dynamic changes in the patient's condition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical monitoring data processing technology, and in particular to a method for early warning of critical neurosurgical complications based on multimodal monitoring. Background Technology

[0002] In neurosurgical intensive care patients, the physiological state may continuously change with disease progression, treatment measures, and recovery stages during postoperative monitoring or intensive care treatment. This typically requires continuous observation of various monitoring data, including intracranial pressure, cerebral oxygen partial pressure, blood oxygen saturation, electroencephalogram (EEG), arterial blood pressure, heart rate, and respiratory status. Different monitoring parameters have different ranges of variation, sampling frequencies, and significance. Furthermore, differences in lesion type, disease severity, treatment stage, and baseline physiological state among patients can cause the same monitoring parameters to exhibit different characteristics of change in different patients or at different stages of monitoring.

[0003] In actual continuous monitoring, medical staff need to simultaneously monitor the current status of multiple monitoring parameters and their changes over time. When multiple monitoring parameters change abnormally within a similar time range, relying solely on the current value of a single monitoring parameter or judging whether each parameter exceeds the reference range individually is insufficient to fully reflect changes in the patient's overall physiological state. Furthermore, different patients have different baseline conditions, and using uniform reference conditions to judge monitoring parameters also has certain limitations in adaptability to the actual conditions of different patients.

[0004] Furthermore, the condition of critically ill neurosurgical patients is characterized by continuous changes, and their monitoring status may also change at different stages of care. If subsequent risk analyses rely on fixed judgment criteria or existing analytical rules over a long period without fully incorporating new monitoring information generated during subsequent care, the analysis results may gradually deviate from the patient's current actual condition, thereby affecting the timely assessment of changes in the patient's condition and the risk of complications.

[0005] Therefore, how to improve the adaptability of multimodal monitoring data analysis to individual differences in the condition and dynamic changes in the disease of critically ill neurosurgical patients is a problem that needs to be further solved in the current continuous monitoring data processing of critically ill neurosurgical patients. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for early warning of critical neurosurgical complications based on multimodal monitoring, thereby improving the adaptability of multimodal monitoring data analysis to individual differences in the condition of critically ill neurosurgical patients and the dynamic changes in their condition.

[0007] This invention provides a method for early warning of critical neurosurgical complications based on multimodal monitoring, the method comprising the following steps:

[0008] Based on the initial status information and multimodal monitoring data of critically ill neurosurgical patients, the status benchmark and initial degeneration characteristics corresponding to each monitoring modality were determined;

[0009] Based on the initial degradation characteristics and corresponding state benchmarks of each monitoring mode, candidate propagation relationships between different monitoring modes are determined, and propagation groups are formed according to the overlapping characteristics between the candidate propagation relationships to obtain dynamic coupling relationships;

[0010] Based on the propagation contribution determined by the dynamic coupling relationship and the initial degradation characteristics, the coupling degradation characteristics are determined, and the state prediction characteristics corresponding to each propagation group are determined according to the coupling degradation characteristics.

[0011] Based on the actual state features corresponding to the state prediction features and the prediction error between them, the verification results corresponding to each propagation group are determined, and the dynamic coupling relationship and coupling degradation features are updated based on the verification results.

[0012] The abnormal evolution characteristics of the patient are determined by the updated dynamic coupling relationship and coupling degradation characteristics, and the risk prediction results corresponding to the target complication are determined based on the abnormal evolution characteristics and the initial degradation characteristics, thereby generating complication early warning information.

[0013] Preferably, the determination of the state benchmark and initial degradation characteristics includes:

[0014] The initial state characteristics corresponding to each monitoring mode are determined based on the multimodal monitoring data, and the preset reference range corresponding to each monitoring mode is determined based on the initial state information. The state benchmark corresponding to each monitoring mode is determined based on the initial state characteristics and the corresponding preset reference range.

[0015] Based on the multimodal monitoring data, the real-time state characteristics corresponding to each monitoring mode are determined, and based on the deviation relationship between the real-time state characteristics and the corresponding state benchmark, the initial degradation characteristics corresponding to each monitoring mode are determined.

[0016] Preferably, the determination of the initial degradation characteristics includes:

[0017] Based on the real-time state characteristics corresponding to each monitoring mode and the state benchmark, determine the state deviation corresponding to each monitoring mode;

[0018] The deviation change amount corresponding to each monitoring mode is determined based on the difference between the state deviation amounts corresponding to adjacent monitoring times;

[0019] Based on the state deviation and the amount of deviation change, the initial degradation characteristics corresponding to each monitoring mode are determined.

[0020] Preferably, the step of grouping propagation relationships based on overlapping features among the candidate propagation relationships to obtain dynamic coupling relationships includes:

[0021] Based on the time sequence of state changes corresponding to each candidate propagation relationship, determine the propagation source, propagation target, expected propagation period, and association strength corresponding to each candidate propagation relationship;

[0022] Based on the propagation objectives and expected propagation periods corresponding to each candidate propagation relationship, candidate propagation relationships with the same propagation objectives and a common intersection of expected propagation periods are divided into the same propagation group, and candidate propagation relationships with the same propagation source are determined within each propagation group;

[0023] The association weights are determined based on the association strength of each candidate propagation relationship, and the coupling weights of each propagation group are determined based on the candidate propagation relationships and their association weights, thus obtaining the dynamic coupling relationship.

[0024] Preferably, determining the coupling degradation features based on the propagation contribution determined by the dynamic coupling relationship and the initial degradation features includes:

[0025] Based on the association weights corresponding to each candidate propagation relationship in the dynamic coupling relationship and the initial degradation characteristics corresponding to its propagation source, the candidate propagation contribution corresponding to each candidate propagation relationship is determined.

[0026] The maximum value among the candidate propagation contributions of the same propagation source within the same propagation group is determined as the in-group contribution of the corresponding propagation source.

[0027] Based on the intra-group contribution of each propagation source and the coupling weight of the corresponding propagation group, the propagation contribution of each propagation group is determined, and the initial degradation characteristics of the monitoring mode corresponding to the propagation target are adjusted according to the propagation contribution to obtain the coupling degradation characteristics.

[0028] Preferably, determining the state prediction features corresponding to each propagation group based on the coupling degradation features includes:

[0029] Based on the dynamic coupling relationship and coupling degradation characteristics, the target monitoring mode and prediction period corresponding to each propagation group are determined, and a preset prediction model is used to determine the complete prediction characteristics of the target monitoring mode within the prediction period based on the coupling degradation characteristics.

[0030] While keeping the preset prediction model, prediction start state and other propagation contributions unchanged, the propagation contribution corresponding to the target propagation group is not included. The coupling degradation characteristics are redefined, and the control prediction characteristics of the target monitoring mode in the prediction period are determined based on the redefined coupling degradation characteristics.

[0031] Based on the complete prediction features and the comparison prediction features, the state prediction features corresponding to the target propagation group are obtained.

[0032] Preferably, the step of determining the verification result corresponding to each propagation group based on the actual state characteristics corresponding to the state prediction characteristics and the prediction error between them, and updating the dynamic coupling relationship and coupling degradation characteristics based on the verification results, includes:

[0033] Obtain the actual state characteristics corresponding to the target monitoring mode within the prediction period, determine the first prediction error and the second prediction error between the actual state characteristics and the complete prediction characteristics and the control prediction characteristics, and determine the prediction improvement amount corresponding to the target propagation group based on the difference between the second prediction error and the first prediction error.

[0034] Based on the comparison between the predicted improvement amount and the preset contribution threshold, the verification result corresponding to the target propagation group is determined, and the corresponding coupling weight adjustment amount is determined based on the verification result.

[0035] The dynamic coupling relationship is updated according to the coupling weight adjustment amount, and the propagation contribution of each propagation group is re-determined according to the updated dynamic coupling relationship. Based on the re-determined propagation contribution and the initial degradation feature, the updated coupling degradation feature is obtained.

[0036] Compared with related technologies, the method for early warning of critical neurosurgical complications based on multimodal monitoring provided by this invention has the following beneficial effects:

[0037] This invention determines the state baseline and initial degradation characteristics corresponding to each monitoring modality based on the patient's initial state information and multimodal monitoring data. This allows abnormal states of different monitoring modalities to be characterized in conjunction with the patient's own initial state, improving the adaptability of monitoring data analysis to individual patient state differences. Furthermore, it establishes a dynamic coupling relationship based on the temporal relationship of state changes in different monitoring modalities and determines the coupling degradation characteristics based on propagation contributions, enabling the joint abnormal changes of multiple monitoring modalities during continuous monitoring to be characterized. Simultaneously, the dynamic coupling relationship is verified and updated through the prediction error between the state prediction results and subsequent actual state characteristics, allowing subsequent abnormal evolution analysis to be continuously adjusted in conjunction with newly generated monitoring information of the patient, thereby improving the adaptability of the complication risk prediction process to dynamic changes in the patient's condition. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of a method for early warning of critical neurosurgical complications based on multimodal monitoring, according to the present invention.

[0039] Figure 2 This is a schematic diagram of the candidate propagation relationship and propagation grouping of the present invention.

[0040] Figure 3 This is a schematic diagram illustrating the process of propagation verification and dynamic coupling relationship update of the present invention. Detailed Implementation

[0041] The following provides a detailed description of a method for early warning of critical neurosurgical complications based on multimodal monitoring, as provided by this invention. This embodiment is used to illustrate the data processing relationships and specific implementation process between the various technical features of this invention, and is not intended to limit the types of monitoring parameters, model structure, model parameters, threshold range, or specific types of target complication.

[0042] This invention can be applied to continuous monitoring scenarios for critically ill neurosurgical patients. By acquiring multimodal monitoring data generated by different monitoring devices for patients, it analyzes the abnormal degradation of each monitoring modality relative to the patient's state baseline and the possible temporal propagation relationships between different monitoring modalities. It then uses subsequent actual monitoring data to verify and update the predictive role of the propagation relationships, ultimately forming the evolutionary characteristics of the patient's abnormal state and predicting the risk of target complications based on this.

[0043] Multimodal monitoring data may include one or more monitoring data such as intracranial pressure, brain tissue oxygen partial pressure, blood oxygen saturation, electroencephalogram (EEG) status, arterial blood pressure, heart rate, and respiratory status.

[0044] For continuously collected monitoring data, a corresponding monitoring sequence is formed according to the actual sampling time; for non-continuously collected data, the actual detection time is recorded and it participates in the determination of state characteristics within the corresponding valid data period.

[0045] In one specific implementation, corresponding multimodal monitoring data is acquired through a data acquisition interface connected to the patient monitoring equipment, and the monitoring modality, acquisition time, and monitoring value corresponding to each monitoring data are recorded. For monitoring data acquired by different monitoring devices, a unified monitoring time series is established based on their actual acquisition time, so that different monitoring modalities form state characteristics that can be jointly analyzed at their corresponding monitoring times.

[0046] For monitoring modes with different sampling frequencies, a unified analysis time interval can be preset, and the valid monitoring value for the corresponding analysis time can be determined based on the actual acquisition time of each monitoring data. When a monitoring mode has multiple valid monitoring values ​​at the corresponding analysis time, the state characteristics corresponding to that analysis time can be determined according to preset data aggregation rules; when there are no monitoring values ​​that meet the validity conditions at the corresponding analysis time, the monitoring mode is marked as a data missing state at the current analysis time.

[0047] When determining valid monitoring values, the validity of the monitoring data is judged based on the data validity identifier provided by the monitoring equipment, the collection time, and the preset data quality conditions. If the monitoring data is interrupted, the monitoring value is missing, or the corresponding data quality conditions are not met, the monitoring data will not be directly used as the current valid status feature.

[0048] For monitoring modalities with missing data, when the corresponding valid state features can be determined based on the preset valid data period, the valid state features are used in subsequent analysis; when the valid state features cannot be determined, the missing data of the monitoring modal at the current analysis time is not used to establish candidate propagation relationships.

[0049] The above method is used to form an effective state feature sequence corresponding to each monitoring mode. Based on the effective state feature sequence, the initial state features, real-time state features and state change sequence corresponding to each monitoring mode are determined, providing a data foundation for subsequent state benchmark determination, candidate propagation relationship analysis and prediction model input.

[0050] It should be noted that the “candidate propagation relationship” in this embodiment is used to describe the temporal sequence and predictive correlation between changes in different monitoring modalities, and does not mean that a medically certain causal relationship can be determined between different physiological states based solely on the candidate propagation relationship.

[0051] refer to Figure 1 As shown, the neurosurgical critical complication early warning method based on multimodal monitoring provided by the present invention includes, in sequence, the determination of state benchmark and initial degradation characteristics, the establishment of candidate propagation relationship and dynamic coupling relationship, propagation contribution and state prediction, propagation verification and dynamic coupling relationship update, as well as abnormal evolution analysis and complication early warning.

[0052] S1: Based on the initial state information and multimodal monitoring data of critically ill neurosurgical patients, determine the state baseline and initial degeneration characteristics corresponding to each monitoring modality.

[0053] After the patient enters the current monitoring phase, an initial observation period is set, and multiple state characteristic values ​​of each monitoring modality that meet the data quality conditions are obtained within the initial observation period.

[0054] For the For each monitoring mode, the median of multiple valid state feature values ​​within the initial observation period is used as the initial state feature corresponding to that monitoring mode. .

[0055] Based on the patient's initial state information, determine the preset reference range corresponding to this monitoring modality from the pre-established set of reference ranges:

[0056]

[0057] in, Indicates the first The lower limit of the preset reference range corresponding to each monitoring mode Indicates the first The upper limit of the preset reference range corresponding to each monitoring mode.

[0058] In one specific implementation, the corresponding state reference is... Represented as:

[0059]

[0060] That is, when the initial state feature is within a preset reference range, the initial state feature is used as the state reference; when the initial state feature exceeds the preset reference range, the reference range boundary closer to the initial state feature is used as the state reference. It is obtained by calculating the median of multiple valid state characteristic values ​​within the initial observation period.

[0061] Furthermore, the reference variation range for state scale unification is determined based on the preset reference range. :

[0062]

[0063] The reference range of variation is used to standardize the scale of state deviations from different monitoring modes, whereby... .

[0064] Subsequently, the real-time state characteristics corresponding to each monitoring mode are determined based on the real-time multimodal monitoring data.

[0065] Considering that different monitoring parameters correspond to different adverse change directions, for the first Each monitoring mode is set with state direction parameters. When the monitored value increases, corresponding to an increase in the degree of abnormality, let... When the monitored value decreases, corresponding to an increase in the degree of abnormality, let... .

[0066] The corresponding state deviation is:

[0067]

[0068] in, For the first Each monitoring mode at the monitoring time The corresponding real-time state characteristics, For the first Each monitoring mode at the monitoring time The corresponding state deviation.

[0069] Furthermore, the deviation change is determined based on the state deviation corresponding to adjacent monitoring times:

[0070]

[0071] in, Indicates the first Each monitoring mode at the monitoring time The corresponding deviation change amount indicates that when the deviation change amount is greater than 0, it means that the degree of deviation of the corresponding state has increased compared to the previous monitoring time, and when the deviation change amount is less than 0, it means that the degree of deviation of the corresponding state has decreased compared to the previous monitoring time.

[0072] The state deviation is used to characterize the degree of unfavorable deviation of the current state from the patient's state baseline, while the deviation change is used to characterize the increase or decrease of the degree of unfavorable deviation relative to the previous monitoring time.

[0073] In one specific implementation, the initial degradation characteristics are determined based on the state deviation and the amount of deviation change:

[0074]

[0075] in Both are greater than or equal to zero. The weights corresponding to the state deviations. The weight corresponding to the deviation from the change. Indicates the first Each monitoring mode at the monitoring time The corresponding initial degradation characteristics.

[0076] The initial degradation characteristic is kept within the range of 0 to 1 by using an upper limit constraint. The initial degradation characteristic obtained in this way represents the degree of abnormal degradation formed by the current monitoring mode's own state before taking into account the propagation effect of other monitoring modes.

[0077] S2: Based on the initial degradation characteristics and corresponding state benchmarks of each monitoring mode, determine the candidate propagation relationships between different monitoring modes, and group the propagation based on the overlapping characteristics between the candidate propagation relationships to obtain dynamic coupling relationships.

[0078] Specifically, based on the state deviation, deviation change, and corresponding initial degradation characteristics formed at multiple consecutive monitoring times, a state change sequence corresponding to each monitoring mode is formed.

[0079] With the first The monitoring modality serves as the source of transmission, the first Using one monitoring modality as the propagation target, in multiple candidate time delays (Using the number of uniform sampling intervals) Determine the time delay correlation value between the propagation source state change sequence and the propagation target delayed state change sequence:

[0080]

[0081] in, This represents the state change corresponding to the propagation source. Indicates delay in propagation target The corresponding state change afterward This indicates the number of valid samples within the current analysis period. This represents the average value of the current propagation source state change sequence participating in the correlation calculation. This represents the average value of the current propagation target delay state change sequence participating in the association calculation.

[0082] When the propagation source state change sequence or propagation target state change sequence participating in the time delay correlation calculation does not have effective fluctuations, making the denominator of the time delay correlation value calculation formula zero, the candidate propagation relationship between the corresponding propagation source and propagation target is not determined based on the current analysis period.

[0083] The direction of state change for different monitoring modes is pre-unified according to the corresponding unfavorable change direction, so that the correlation value can reflect the positive temporal correlation between unfavorable state changes.

[0084] When the delay correlation value corresponding to one or more consecutive candidate delays reaches the preset correlation threshold, the corresponding propagation source and propagation target are determined as candidate propagation relationships.

[0085] For a set of consecutive candidate delays that meet the preset association threshold, the minimum delay and the maximum delay are determined as the lower limit and upper limit of the propagation delay, respectively, so as to determine the expected propagation period corresponding to the candidate propagation relationship.

[0086] For example, to illustrate the process of determining the expected propagation period, it is assumed that the time delay correlation values ​​corresponding to the 5th to 10th minutes after the state change of the propagation source all meet the preset correlation conditions. Then, the 5th to 10th minutes after the state change can be determined as the expected propagation period of the corresponding candidate propagation relationship. This time interval is only used to illustrate the determination of the expected propagation period and the subsequent propagation grouping process, and is not used to limit the actual propagation delay between different monitoring modalities.

[0087] The maximum value among the time-delay correlation values ​​within the expected propagation period is determined as the correlation strength corresponding to the candidate propagation relationship.

[0088] The same monitoring modality can generate multiple candidate propagation relationships pointing to the same propagation target based on different state change components. For example, candidate propagation relationships can be established based on the intracranial pressure deviation sequence and the deviation change sequence, respectively, and the subsequent state changes of brain tissue oxygen partial pressure.

[0089] After identifying multiple candidate propagation relationships, they are first classified according to the propagation target, and then sorted from earliest to latest according to the start time of the expected propagation period of each candidate propagation relationship.

[0090] Establish a propagation group based on the first candidate propagation relationship after ranking, and use its expected propagation period as the current common propagation period.

[0091] When the expected propagation period corresponding to the next candidate propagation relationship intersects with the current common propagation period, it is added to the current propagation group, and the corresponding intersection is updated to the new common propagation period.

[0092] When there is no intersection, a new propagation group is established using the current candidate propagation relationship.

[0093] By using the above methods, candidate propagation relationships within the same propagation group have the same propagation goals and their expected propagation periods have a common intersection, and the same candidate propagation relationship is assigned to only one propagation group in a single propagation analysis process.

[0094] The sources, targets, and expected durations of each candidate propagation relationship, as well as the process of forming propagation groups based on the common intersection of expected propagation durations (see reference). Figure 2 (As shown).

[0095] Furthermore, the association weights are determined based on the association strength corresponding to the candidate propagation relationships.

[0096] When the correlation strength has been normalized to the range of 0 to 1, the correlation strength can be directly used as the correlation weight.

[0097] For multiple candidate propagation relationships with the same propagation source in the same propagation group, the maximum value among their corresponding association weights is determined as the effective association weight corresponding to that propagation source.

[0098] If the propagation group include For each different propagation source, the corresponding coupling weight is expressed as:

[0099]

[0100] in, For the first The effective association weights corresponding to each propagation source Indicates the first The coupling weights corresponding to each propagation group.

[0101] The dynamic coupling relationship of the patient is formed by the transmission group, the candidate transmission relationships within the group, the association weights corresponding to the candidate transmission relationships, and the coupling weights corresponding to the transmission group.

[0102] S3: Determine the coupling degradation features based on the propagation contribution determined by the dynamic coupling relationship and the initial degradation features, and determine the state prediction features corresponding to each propagation group according to the coupling degradation features.

[0103] For candidate propagation relationships Candidate propagation contributions are determined based on their association weights and the initial degradation characteristics of the corresponding propagation sources:

[0104]

[0105] in, Indicates candidate propagation relationship Corresponding candidate propagation contributions Indicates candidate propagation relationship The corresponding association weights, This indicates the initial degradation characteristics of the corresponding propagation source.

[0106] For multiple candidate propagation contributions with the same propagation source within the same propagation group, the maximum value among them is determined as the in-group contribution corresponding to that propagation source:

[0107]

[0108] in Indicates the first [number]th [unit] within the same transmission group The group contribution corresponding to each source of transmission; They represent the first and second generations within the same propagation group. Each propagation source corresponds to a candidate propagation contribution for each candidate propagation relationship; This indicates the number of candidate propagation relationships corresponding to the propagation source within the current propagation group.

[0109] This avoids the repeated accumulation of abnormal states when the same propagation source forms multiple candidate propagation relationships due to different state change components.

[0110] The propagation contribution is determined based on the intra-group contribution of each propagation source and the coupling weight of the propagation group:

[0111]

[0112] in Indicates the first The dissemination contribution corresponding to each dissemination group; Indicates the coupling weights of the corresponding propagation group; Indicates the first The group contribution corresponding to each source of transmission; This indicates the number of different transmission sources within the current transmission group.

[0113] When the propagation contribution calculated according to the intra-group contribution of each propagation source and the coupling weight of the propagation group is greater than 1, the propagation contribution is determined to be 1, thereby keeping the propagation contribution in the range of 0 to 1.

[0114] Furthermore, the initial degradation characteristics of the monitoring modality corresponding to the propagation target are adjusted based on the propagation contribution.

[0115] In one implementation:

[0116]

[0117] in, This represents the initial degradation characteristics corresponding to the propagation target. Indicates the propagation contribution of the propagation group. This indicates the coupling degradation characteristics corresponding to the propagation target.

[0118] When multiple propagation groups exist for the same propagation target, propagation analysis can be performed separately according to the prediction time period. For multiple propagation groups with overlapping prediction time periods, coupling degradation characteristics can be determined group by group according to the order of propagation time periods, thereby avoiding the unconstrained repeated accumulation of contributions from different propagation groups.

[0119] Furthermore, the predicted time period corresponding to the propagation group is determined based on the common intersection of the expected propagation time periods of each candidate propagation relationship in the propagation group, and the target monitoring mode is determined based on the corresponding propagation target.

[0120] For the preset prediction model

[0121] In one specific implementation, a corresponding preset prediction model is established for different target monitoring modes. Each preset prediction model is used to predict the state characteristics of the corresponding target monitoring mode within the prediction period based on the historical multimodal monitoring state characteristics related to the corresponding target monitoring mode.

[0122] In one specific implementation, the preset prediction model is constructed using a gated recurrent unit network, which includes an input layer, a gated recurrent unit layer, a fully connected layer, and an output layer connected in sequence.

[0123] The input layer is used to receive state feature vectors corresponding to multiple consecutive monitoring times in chronological order.

[0124] The state feature vector corresponding to each monitoring moment includes at least the real-time state features, initial degradation features, propagation contribution of each propagation group, and coupling degradation features of each target monitoring mode.

[0125] Considering that the number of effective propagation groups may differ at different monitoring times, in one specific implementation, a maximum number of propagation groups is preset. .

[0126] Sort the current effective propagation groups in descending order of coupling weight. When the number of effective propagation groups is greater than... When selecting the previous option Each propagation group participates in the model input; when the number of effective propagation groups is less than [number missing], [the number of effective propagation groups is less than [number missing]]. When there is insufficient space, fill the missing space with zero values.

[0127] When determining the control prediction features corresponding to the target propagation group, the input positions corresponding to each propagation group remain unchanged under the condition of complete prediction. The propagation contribution in the input position corresponding to the target propagation group is set to zero. Other propagation groups are not reordered or their input positions are not adjusted because the propagation contribution of the target propagation group is set to zero. The coupling degradation features affected by the propagation contribution of the target propagation group are re-determined.

[0128] This keeps the model input dimensions fixed for different monitoring times.

[0129] The same propagation group sorting and zero-filling rules are used in both the model training phase and the real-time patient monitoring phase. The first propagation group input position corresponds to the first effective propagation group after sorting according to the coupling weight from largest to smallest, the second propagation group input position corresponds to the second effective propagation group after sorting, and so on, so that the sorting meaning corresponding to each propagation group input position remains consistent.

[0130] In one specific implementation, the state feature vectors corresponding to the 12 consecutive monitoring times prior to the prediction start time are used to form an input sequence.

[0131] The gated recurrent unit layer is set to one layer with a hidden state dimension of 32. The output of the gated recurrent unit layer is connected to the fully connected layer, and the fully connected layer outputs the predicted state features of the target monitoring modality during the prediction period.

[0132] When the prediction period includes multiple target monitoring times, the preset prediction model outputs the corresponding prediction state feature sequence according to the time order of each target monitoring time. Each prediction state feature in the prediction state feature sequence corresponds to the actual state feature of the corresponding actual monitoring time within the prediction period.

[0133] During the training phase, the above input sequence is formed using historical multimodal continuous monitoring data, and the target monitoring modal state features actually acquired within the corresponding prediction period are used as training labels.

[0134] During the training process of the pre-set prediction model, historical monitoring data is divided into patient units so that the monitoring data corresponding to the same patient does not appear in the training data and independent validation data at the same time.

[0135] For each training sample, the effective state features before the start time of the corresponding risk prediction are extracted according to the predetermined input sequence length, and training labels are formed based on the actual state features of the corresponding target monitoring modality in the subsequent prediction period.

[0136] In both the training and independent validation data, the same monitoring modality order, propagation group sorting rules, and missing position filling rules are used to form the model input, ensuring that the meaning of the corresponding input positions remains consistent between the model training and actual monitoring phases.

[0137] The network parameters of the preset prediction model are adjusted based on the prediction loss corresponding to the training data, and the model training termination time and the model parameters to be saved are determined using the prediction loss corresponding to the independent validation data. After training is completed, the preset prediction model corresponding to the target monitoring modality is fixed, and the fixed model is used for state prediction during the actual monitoring phase.

[0138] The model training loss uses mean squared error:

[0139]

[0140] in, This represents the predicted state features output by the model. This indicates the corresponding actual state characteristics. This represents the training loss corresponding to the preset prediction model. This indicates the number of the effective predicted state feature.

[0141] In one specific implementation, the Adam optimization algorithm is used to update the network parameters, with the learning rate set to 0.001 and the batch size set to 64.

[0142] When the prediction loss corresponding to the validation data no longer decreases within a preset number of consecutive training rounds, training is stopped, and the corresponding model structure and model parameters are saved.

[0143] After the model training is completed, the model structure and model parameters are fixed during the real-time patient monitoring phase.

[0144] When making a complete prediction, the complete state feature sequence, including the propagation contribution of the target propagation group, is input into the preset prediction model to obtain the complete prediction features corresponding to the target monitoring mode.

[0145] When performing comparative prediction, the preset prediction model, the prediction starting state, and the corresponding propagation contributions of other propagation groups remain unchanged. Only the propagation contribution of the target propagation group is set to zero, and the coupling degradation characteristics affected by this propagation contribution are redefined.

[0146] The newly determined state feature sequence is input into the same preset prediction model to obtain the corresponding control prediction features.

[0147] Therefore, the main difference between the complete prediction and the control prediction lies in whether the target propagation group's corresponding propagation contribution is included in the prediction.

[0148] The complete prediction features and the control prediction features together form the state prediction features corresponding to the target propagation group.

[0149] S4: Based on the actual state features corresponding to the state prediction features and the prediction error between them, determine the verification results corresponding to each propagation group, and update the dynamic coupling relationship and coupling degradation features based on the verification results.

[0150] After the predicted period for the propagation group ends, the actual state characteristics corresponding to the target monitoring modality are obtained.

[0151] When the prediction period includes multiple actual monitoring times, the average absolute error between the actual state characteristics, the complete prediction characteristics, and the control prediction characteristics is calculated separately.

[0152] The first prediction error is:

[0153]

[0154] in, The first prediction error represents the difference between the actual state characteristics and the complete predicted characteristics. Indicates the first [number]th ... The actual state characteristics corresponding to each valid monitoring moment; Indicates the first condition under complete prediction conditions Predicted state characteristics corresponding to each valid monitoring time; This indicates the number of valid monitoring moments used in error calculation within the prediction period.

[0155] The second prediction error is:

[0156]

[0157] in, This represents the second prediction error between the actual state characteristics and the control prediction characteristics; Indicates the first under the control prediction condition The predicted state characteristics corresponding to each effective monitoring time.

[0158] The predicted improvement amount corresponding to the target propagation group is determined based on the difference between the two. :

[0159]

[0160] When the predicted improvement is greater than zero, it means that the complete prediction result after taking into account the propagation contribution of the target propagation group is closer to the actual monitoring state than the control prediction result.

[0161] Set a preset contribution threshold. Step size for single-cycle coupling weight adjustment and the number of consecutive valid validations that fail to provide predictive support .

[0162] when At that time, determine whether the target propagation group receives effective forecast support within the current forecast period, and then... ,in, This represents the amount of coupling weight adjustment determined during the current propagation verification process.

[0163] when At that time, it is determined that the current improvement in prediction is insufficient to adjust the corresponding coupling weights, and then... .

[0164] If the prediction improvement of the target propagation group in the current effective validation period is not greater than zero, but the number of consecutive effective validation periods without prediction support has not yet reached the preset number of effective validations. At that time, keep the corresponding coupling weights unchanged, and let .

[0165] When the target propagation group is continuous If the predicted improvement for each valid validation period is not greater than zero, then the propagation group is determined to have continuously failed to obtain predictive support, and this is set to... .

[0166] When the number of consecutive valid validation periods in which the target propagation group fails to obtain predictive support reaches After reducing the corresponding coupling weights, the number of valid verification periods that have not received prediction support for the propagation group is reset to zero.

[0167] When actual monitoring data is missing or the data quality does not meet the preset requirements during the forecast period, the corresponding forecast period will not be counted as a valid verification number.

[0168] Updated coupling weights for:

[0169]

[0170] This keeps the updated coupling weights within the range of 0 to 1.

[0171] Subsequently, the propagation contribution corresponding to the propagation group is re-determined based on the updated dynamic coupling relationship, and the updated coupling degradation characteristics are obtained again based on the re-determined propagation contribution and the corresponding initial degradation characteristics.

[0172] By comparing the prediction errors among complete predicted features, control predicted features, and actual state features, the verification results of the target propagation group are determined. Based on these verification results, the dynamic coupling relationship, propagation contribution, and coupling degradation features are updated, thus forming a closed loop for propagation verification and dynamic coupling relationship updating (see reference). Figure 3 (As shown).

[0173] S5: Determine the abnormal evolution characteristics of the patient based on the updated dynamic coupling relationship and coupling degradation characteristics, and determine the risk prediction results corresponding to the target complication based on the abnormal evolution characteristics and the initial degradation characteristics, and generate complication early warning information.

[0174] The abnormal evolution characteristics of the patient are determined by the updated dynamic coupling relationship and coupling degradation characteristics.

[0175] Specifically, in continuous Within a valid monitoring period, the coupling degradation characteristics of each target monitoring mode and the coupling weights of each propagation group are obtained.

[0176] For the For each target monitoring mode, determine the corresponding coupling degradation change:

[0177]

[0178] in, Indicates the first The target monitoring mode is in continuous The corresponding changes in coupling degradation within each effective monitoring period This indicates the updated coupling degradation characteristics corresponding to the current monitoring period. This indicates the coupling degradation characteristics corresponding to the first effective monitoring period involved in the current anomaly evolution analysis.

[0179] When determining the change in coupling weight for each propagation group, the propagation group corresponding to the current propagation group in the historical effective monitoring period is determined based on the propagation target and candidate propagation relationships within the group. The change in coupling weight is then determined based on the coupling weight of the corresponding propagation group in different effective monitoring periods. For cases newly formed in the current monitoring period where no corresponding propagation group exists in the historical effective monitoring period, the change in coupling weight is set to zero.

[0180] For the Each propagation group determines the corresponding change in coupling weights:

[0181]

[0182] The current coupling degradation feature, the amount of coupling degradation change, the coupling weight corresponding to the effective propagation group, and the amount of coupling weight change are collectively used as the abnormal evolution feature.

[0183] Risk prediction model

[0184] In one specific implementation, a corresponding risk prediction model is established for each target complication.

[0185] Before establishing a risk prediction model for the target complication, the types of target complications that require risk warning and the corresponding preset risk prediction time range are determined in advance. Based on the clinical confirmation criteria for the target complication, the confirmation records and occurrence times related to the target complication in the historical monitoring data are determined.

[0186] For each valid risk prediction moment in the historical monitoring data, that moment is taken as the starting point of the current risk analysis, and the corresponding initial degradation characteristics, dynamic coupling relationships and abnormal evolution characteristics are determined based on the multimodal monitoring data obtained at that moment and before.

[0187] The preset risk prediction time range after the risk prediction time is used as the confirmation range for the occurrence status of the target complication, and the occurrence of the target complication is determined based on the corresponding clinical confirmation records within this range.

[0188] Specifically, monitoring records in which the target complication has been confirmed to have occurred before the risk prediction time will not be directly used as training samples for risk prediction before the occurrence of the target complication; and monitoring records in which the occurrence of the target complication cannot be determined within the preset risk prediction time range will not be directly used as training samples with a clear complication label.

[0189] This ensures a clear temporal correspondence between the risk prediction input features corresponding to the target complication and its subsequent occurrence, and avoids using monitoring information that can only be obtained after the risk prediction moment to construct the current risk prediction input during the prediction model training process.

[0190] To keep the model input dimensions fixed, the set of monitoring modalities participating in risk prediction and the maximum number of effective propagation groups are pre-defined. .

[0191] Each effective propagation group is sorted from largest to smallest according to the updated coupling weight. When the number of propagation groups exceeds Before selecting One; when the number of propagation groups is insufficient, zero values ​​are used for padding.

[0192] The initial degradation features, updated coupled degradation features, and changes in coupled degradation corresponding to each monitoring modality, along with the coupling weights and changes in coupling weights corresponding to each effective propagation group after sorting, are sequentially combined to form a fixed-dimensional risk prediction input vector:

[0193]

[0194] in, This represents the risk prediction input vector corresponding to the target complication; Indicates the first The initial degradation characteristics corresponding to each monitoring mode; Indicates the first The updated coupling degradation features corresponding to each target monitoring mode; This indicates the amount of coupling degradation change of the corresponding target monitoring mode within a continuous monitoring period; Indicates the first The coupling weights corresponding to each propagation group; Indicates the corresponding propagation group in continuous Changes in coupling weights within a valid monitoring period; This indicates the number of monitoring modalities involved in risk prediction; This indicates the number of target monitoring modes involved in risk prediction; This indicates the number of propagation group input locations pre-set in the risk prediction model.

[0195] The risk prediction model consists of an input layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence.

[0196] In one specific implementation, the first fully connected layer includes 32 neurons, the second fully connected layer includes 16 neurons, both fully connected layers use the ReLU activation function, the output layer sets an output node corresponding to the current target complication, and uses the Sigmoid function to output a risk prediction value in the range of 0 to 1.

[0197] For historical monitoring data, at each historical risk prediction point, the risk prediction input vector is determined using the multimodal monitoring data obtained at that point and before.

[0198] Then, based on the medical records and monitoring records within the preset risk prediction time range after the historical risk prediction time, the corresponding training labels are determined.

[0199] When there are confirmed target complication records within the preset risk prediction time range, the corresponding training label is set to 1; when there are no target complication records, the corresponding training label is set to 0.

[0200] This establishes a correspondence between the risk prediction input and the actual occurrence of the target complication.

[0201] During the training process of the risk prediction model, historical training data and independent validation data are divided into patient-based data sets, so that different monitoring periods of the same patient do not participate in model training and independent validation at the same time.

[0202] For each training sample, the risk prediction input vector is determined using only the valid monitoring data obtained before the corresponding risk prediction time, and the occurrence status of the target complication within the preset risk prediction time range after the corresponding risk prediction time is used as the training label.

[0203] In both the training and independent validation data, the same monitoring modality arrangement rules, propagation group sorting rules, and zero-value filling rules are used to ensure that the same positions in the risk prediction input vector have consistent data meaning.

[0204] The risk prediction model parameters are updated based on the model loss corresponding to the training data, and the training termination time is determined based on the model loss corresponding to the independent validation data. After training is completed, the model structure and model parameters are fixed, and the risk warning threshold is determined using the correspondence between the risk prediction values ​​in the independent validation data and the actual occurrence of the target complication.

[0205] The model training uses binary cross-entropy as the loss function:

[0206]

[0207] in, This represents the training loss corresponding to the target complication risk prediction model. Indicates the first Risk prediction value corresponding to each training sample Indicates the corresponding training label. This indicates the training sample number.

[0208] In one specific implementation, the Adam optimization algorithm is also used for parameter updates, the learning rate is set to 0.001, the batch size is set to 64, and the training termination time is determined based on the loss corresponding to the independent validation data.

[0209] Once the model training is complete, fix the model structure and model parameters.

[0210] During real-time patient monitoring, the risk prediction input vector corresponding to the current patient is input into the trained risk prediction model to obtain the risk prediction value corresponding to the target complication. .

[0211] Risk warning thresholds are pre-determined based on the correspondence between predicted risk values ​​and actual records of target complications in independently validated samples. .

[0212] when At that time, early warning information corresponding to the target complication is generated.

[0213] Early warning information may include the target complication type, current risk prediction value, corresponding prediction time range, and relevant monitoring modalities.

[0214] Risk prediction results are used as auxiliary risk indicators during the monitoring of critically ill neurosurgical patients and are not directly used as clinical diagnostic results for target complications.

Claims

1. A neurosurgical critical complication early warning method based on multi-modal monitoring, characterized in that, The early warning method includes the following steps: Based on the initial status information and multimodal monitoring data of critically ill neurosurgical patients, the status benchmark and initial degeneration characteristics corresponding to each monitoring modality were determined; Based on the initial degradation characteristics and corresponding state benchmarks of each monitoring mode, candidate propagation relationships between different monitoring modes are determined, and propagation groups are formed according to the overlapping characteristics between the candidate propagation relationships to obtain dynamic coupling relationships; Based on the propagation contribution determined by the dynamic coupling relationship and the initial degradation characteristics, the coupling degradation characteristics are determined, and the state prediction characteristics corresponding to each propagation group are determined according to the coupling degradation characteristics. Based on the actual state features corresponding to the state prediction features and the prediction error between them, the verification results corresponding to each propagation group are determined, and the dynamic coupling relationship and coupling degradation features are updated based on the verification results. The abnormal evolution characteristics of the patient are determined by the updated dynamic coupling relationship and coupling degradation characteristics, and the risk prediction results corresponding to the target complication are determined based on the abnormal evolution characteristics and the initial degradation characteristics, thereby generating complication early warning information.

2. The method for predicting neurosurgical critical complications based on multi-modal monitoring according to claim 1, characterized in that, The determination of the state benchmark and initial degradation characteristics includes: The initial state characteristics corresponding to each monitoring mode are determined based on the multimodal monitoring data, and the preset reference range corresponding to each monitoring mode is determined based on the initial state information. The state benchmark corresponding to each monitoring mode is determined based on the initial state characteristics and the corresponding preset reference range. Based on the multimodal monitoring data, the real-time state characteristics corresponding to each monitoring mode are determined, and based on the deviation relationship between the real-time state characteristics and the corresponding state benchmark, the initial degradation characteristics corresponding to each monitoring mode are determined.

3. The method for early warning of critical neurosurgical complications based on multimodal monitoring according to claim 2, characterized in that, The determination of the initial degradation characteristics includes: Based on the real-time state characteristics corresponding to each monitoring mode and the state benchmark, determine the state deviation corresponding to each monitoring mode; The deviation change amount corresponding to each monitoring mode is determined based on the difference between the state deviation amounts corresponding to adjacent monitoring times; Based on the state deviation and the amount of deviation change, the initial degradation characteristics corresponding to each monitoring mode are determined.

4. The method for early warning of critical neurosurgical complications based on multimodal monitoring according to claim 1, characterized in that, The step of grouping propagation relationships based on overlapping features among the candidate propagation relationships to obtain dynamic coupling relationships includes: Based on the time sequence of state changes corresponding to each candidate propagation relationship, determine the propagation source, propagation target, expected propagation period, and association strength corresponding to each candidate propagation relationship; Based on the propagation objectives and expected propagation periods corresponding to each candidate propagation relationship, candidate propagation relationships with the same propagation objectives and a common intersection of expected propagation periods are divided into the same propagation group, and candidate propagation relationships with the same propagation source are determined within each propagation group; The association weights are determined based on the association strength of each candidate propagation relationship, and the coupling weights of each propagation group are determined based on the candidate propagation relationships and their association weights, thus obtaining the dynamic coupling relationship.

5. The method for early warning of critical neurosurgical complications based on multimodal monitoring according to claim 4, characterized in that, The determination of coupling degradation features based on the propagation contribution determined by the dynamic coupling relationship and the initial degradation features includes: Based on the association weights corresponding to each candidate propagation relationship in the dynamic coupling relationship and the initial degradation characteristics corresponding to its propagation source, the candidate propagation contribution corresponding to each candidate propagation relationship is determined. The maximum value among the candidate propagation contributions of the same propagation source within the same propagation group is determined as the in-group contribution of the corresponding propagation source. Based on the intra-group contribution of each propagation source and the coupling weight of the corresponding propagation group, the propagation contribution of each propagation group is determined, and the initial degradation characteristics of the monitoring mode corresponding to the propagation target are adjusted according to the propagation contribution to obtain the coupling degradation characteristics.

6. The method for early warning of critical neurosurgical complications based on multimodal monitoring according to claim 5, characterized in that, The step of determining the state prediction features corresponding to each propagation group based on the coupling degradation features includes: Based on the dynamic coupling relationship and coupling degradation characteristics, the target monitoring mode and prediction period corresponding to each propagation group are determined, and a preset prediction model is used to determine the complete prediction characteristics of the target monitoring mode within the prediction period based on the coupling degradation characteristics. While keeping the preset prediction model, prediction start state and other propagation contributions unchanged, the propagation contribution corresponding to the target propagation group is not included. The coupling degradation characteristics are redefined, and the control prediction characteristics of the target monitoring mode in the prediction period are determined based on the redefined coupling degradation characteristics. Based on the complete prediction features and the comparison prediction features, the state prediction features corresponding to the target propagation group are obtained.

7. The method for early warning of critical neurosurgical complications based on multimodal monitoring according to claim 6, characterized in that, The step of determining the verification result corresponding to each propagation group based on the actual state characteristics corresponding to the state prediction characteristics and the prediction error between them, and updating the dynamic coupling relationship and coupling degradation characteristics based on the verification results, includes: Obtain the actual state characteristics corresponding to the target monitoring mode within the prediction period, determine the first prediction error and the second prediction error between the actual state characteristics and the complete prediction characteristics and the control prediction characteristics, and determine the prediction improvement amount corresponding to the target propagation group based on the difference between the second prediction error and the first prediction error. Based on the comparison between the predicted improvement amount and the preset contribution threshold, the verification result corresponding to the target propagation group is determined, and the corresponding coupling weight adjustment amount is determined based on the verification result. The dynamic coupling relationship is updated according to the coupling weight adjustment amount, and the propagation contribution of each propagation group is re-determined according to the updated dynamic coupling relationship. Based on the re-determined propagation contribution and the initial degradation feature, the updated coupling degradation feature is obtained.