Multimodal monitoring of postoperative rehabilitation risk warning intelligent medical system and device for brain tumor patients

Abstract

The application relates to the technical field of postoperative monitoring of brain tumor patients, in particular to a multi-modal monitoring intelligent medical system and equipment for postoperative rehabilitation risk early warning of brain tumor patients. Through automatic data acquisition and standardized processing, the problem of long daily nursing time consumption is directly solved; through disease-specific feature extraction and risk prediction steps, the problems of high unplanned readmission rate and high incidence of serious adverse events are directly solved; through automatic matching and generation of rehabilitation programs, the problem of long rehabilitation program generation time is directly solved. Through automatic acquisition and standardized processing of multi-modal data, information islands are eliminated, and manual data collection time is reduced; through accurate early warning, medical staff can obtain early warning before the occurrence of risk events, so that intervention can be made in advance.

Description

Technical Field

[0001] This application relates to the field of postoperative monitoring technology for brain tumor patients, specifically to an intelligent medical system and equipment for early warning of postoperative rehabilitation risks in brain tumor patients using multimodal monitoring. Background Technology

[0002] Patients who have undergone craniotomy for brain tumors are in intensive care post-surgery, requiring continuous monitoring of their vital signs, intracranial status, neurological function recovery, and risk of complications. Healthcare staff (including doctors and nurses) need to be aware of each patient's risk status in real time and intervene promptly when the risk increases to prevent serious adverse events such as intracranial hemorrhage, seizures, infections, and deep vein thrombosis, and to reduce the rate of unplanned readmissions within 30 days. Unplanned readmissions refer to the need for readmission within 30 days of discharge due to worsening of the patient's condition. This rate directly reflects the quality of postoperative care and the effectiveness of rehabilitation programs. Existing general risk scoring systems cannot accurately identify the specific risk patterns of patients after brain tumor surgery, leading to high-risk patients not receiving timely intervention and ultimately requiring readmission.

[0003] Serious adverse events include intracranial hemorrhage, status epilepticus, severe infection, and deep vein thrombosis, which are often sudden and urgent. Existing intermittent manual monitoring cannot achieve continuous risk perception and early warning, forcing healthcare staff to respond passively only after an event occurs. Nurses need to manually collect and integrate patient data from multiple information systems (HIS, LIS, PACS, bedside monitors) for risk assessment and nursing records. This process is time-consuming, encroaching on time spent directly caring for patients. When a patient's risk status changes, physicians need to conduct a comprehensive assessment and manually develop or adjust the rehabilitation plan, a process that takes an average of more than 12 minutes, causing interventions to lag behind changes in risk. Summary of the Invention

[0004] To address the aforementioned technical problems in ICU or ward monitoring scenarios following brain tumor surgery, such as high rates of unplanned readmissions, high incidence of serious adverse events, long daily nursing hours, and lengthy time for generating rehabilitation plans, this application, as its first aspect, provides a method for obtaining postoperative rehabilitation plans for brain tumor patients. This method includes: obtaining physiological modality data, clinical laboratory modality data, behavioral state modality data, and postoperative imaging modality data of the brain tumor patient; standardizing the physiological modality data, clinical laboratory modality data, behavioral state modality data, and postoperative imaging modality data respectively to construct a standard dataset; and extracting temporal physiological features and imaging data from the standardized physiological data. The system integrates deep features, clinical structured features, and behavioral structured features; it fuses temporal physiological features, imaging deep features, clinical structured features, and behavioral structured features to generate a fused feature vector; it constructs a risk prediction model containing a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, and an adaptive weighted fusion layer, and trains, validates, and optimizes the risk prediction model; it inputs the fused feature vector into the trained risk prediction model to perform forward propagation calculations to output a comprehensive risk score, risk level, risk type, and interpretable evidence; and it generates a rehabilitation plan based on the comprehensive risk score, risk level, risk type, and interpretable evidence.

[0005] As a second aspect of this application, this application provides a virtual device for obtaining postoperative rehabilitation plans for brain tumor patients, comprising: a data acquisition and standardization unit for acquiring physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data of postoperative brain tumor patients, and standardizing the acquired data to obtain physiological standardized data, clinical laboratory standardized data, behavioral state standardized data, and imaging standardized data; a feature extraction unit for extracting temporal physiological features from the physiological standardized data, clinical structured features from the clinical laboratory standardized data, behavioral structured features from the behavioral state standardized data, and image depth features from the imaging standardized data; and a feature fusion unit for integrating the temporal physiological features, image depth features, and clinical structured features. The system integrates structured features and behavioral structured features to generate a fused feature vector. A risk identification unit sequentially performs orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping on the fused feature vector to obtain risk scores for five risks: epilepsy, intracranial hemorrhage, infection, falls, and deep vein thrombosis. Based on these scores, it generates risk levels and risk types. An interpretable evidence generation unit retrieves evidence from an authoritative medical knowledge base based on the marginal contributions of temporal physiological features, clinical structured features, behavioral structured features, and image depth features to the prediction results, dynamically generating interpretable evidence. Finally, a rehabilitation plan generation unit generates rehabilitation plans based on risk type, risk level, risk score, and interpretable evidence through knowledge graph matching and multi-hop reasoning. Attached Figure Description

[0006] Figure 1 The diagram shown is a flowchart illustrating a method for obtaining a postoperative rehabilitation plan for a brain tumor patient according to an embodiment of this application.

[0007] Figure 2 The diagram shown is a structural block diagram of a virtual device for obtaining postoperative rehabilitation plans for brain tumor patients, according to an embodiment of this application.

[0008] Figure 3 The diagram shown is a structural block diagram of a virtual device for obtaining postoperative rehabilitation plans for brain tumor patients, according to an embodiment of this application. Detailed Implementation

[0009] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] In response to the aforementioned technical problems in ICU or ward monitoring scenarios after brain tumor surgery, such as high rates of unplanned readmissions, high incidence of serious adverse events, long average daily nursing time, and long time required to generate rehabilitation plans, this application provides a method for obtaining postoperative rehabilitation plans for brain tumor patients. This method first collects and standardizes patient physiological modal data, clinical laboratory modal data, behavioral state modal data, and postoperative imaging modal data from multiple data sources in parallel through an automated interface. It then extracts temporal physiological features from the standardized physiological data, clinical structured features from the standardized clinical laboratory data, behavioral structured features from the standardized behavioral state data, and image depth features from the standardized imaging data. These features are then fused to generate a fused feature vector. The fused feature vector is then subjected to orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping to obtain comprehensive risk scores, risk levels, and risk types for five risks: epilepsy, intracranial hemorrhage, infection, falls, and deep vein thrombosis. Based on the marginal contribution of temporal physiological features, clinical structured features, behavioral structured features, and image depth features to the prediction results, authoritative medical knowledge bases are retrieved to obtain evidence-based support, and interpretable evidence is dynamically generated. Finally, a rehabilitation plan is generated based on the risk type, risk level, risk score, and interpretable evidence. The method provided in this application can solve the technical problem of how to effectively identify the risk status of patients after brain tumor surgery from multimodal data, and can also automatically generate corresponding rehabilitation plans based on the identification results.

[0011] By automating data collection and standardizing processing, the problem of "long daily nursing time" is directly solved; by extracting disease-specific features and implementing risk prediction steps, the problems of "high unplanned readmission rate" and "high incidence of serious adverse events" are directly solved; and by automatically matching and generating rehabilitation plans, the problem of "long rehabilitation plan generation time" is directly solved. Through automatic collection and standardized processing of multimodal data, information silos are eliminated, and the time spent on manual data collection is reduced; through precise early warning, medical staff can receive early warnings before risk events occur, thus enabling early intervention.

[0012] The technical solution of this application is deployed and run in the following hardware environment: Training server: configured with 4 NVIDIA A100 (40GB) GPUs, Intel Xeon Gold 6248R CPU @ 3.0GHz (48 cores), 512GB DDR4 RAM, and 2TB NVMe SSD storage; Inference server: configured with 1 NVIDIA T4 (16GB) GPU, Intel Xeon Silver 4214 CPU @ 2.2GHz (24 cores), 128GB RAM, and 500GB SSD storage; Data acquisition terminals: bedside monitor (Mindray BeneVision N series), transcranial Doppler device (Rimed Digi-Lite), and EEG acquisition device (Nihon Kohden EEG-1200); Communication protocols: HL7 / FHIR standard API, DICOM protocol, and MQTT IoT protocol.

[0013] The technical solution of this application is deployed and run in the following software environment: Operating system: Ubuntu 20.04LTS; Deep learning framework: PyTorch 1.12.0 (CUDA 11.6); Machine learning libraries: LightGBM 3.3.2, scikit-learn 1.1.2; Knowledge graph construction: Neo4j 4.4.5; API service framework: FastAPI 0.85.0; Data processing: NumPy 1.21.5, Pandas 1.4.3, NiBabel 4.0.2 (medical image processing).

[0014] Detailed Implementation of the Method The following is a detailed description of a method for obtaining a postoperative rehabilitation program for brain tumor patients, provided by at least one embodiment of this disclosure. Figure 1 As shown, the method includes steps S100 to S500.

[0015] Step S100: Obtain physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data of patients after brain tumor surgery, and standardize the obtained data to obtain physiological standardized data, clinical laboratory standardized data, behavioral state standardized data, and imaging standardized data.

[0016] These multimodal data were extracted from the "MIMIC-III" intensive care database and the "eICU Collaborative Research Database" in the PhysioNet public database, as well as publicly available datasets on postoperative rehabilitation research for brain tumors published in *Nature Medicine*. The aforementioned publicly available data were obtained through standard API interfaces or offline downloads and converted to the unified FHIR standard format. The data content covers physiological modalities, clinical laboratory modalities, imaging modalities, and behavioral modalities to verify the effectiveness of the algorithm in this application under different data distributions.

[0017] Optionally, physiological modal data include: heart rate (HR), oxygen saturation (SpO2), body temperature (T), respiratory rate (RR), systolic blood pressure (SBP), diastolic blood pressure (DBP), non-invasive transcranial Doppler estimated intracranial pressure (eICP), cerebral perfusion pressure (CPP), drainage volume, muscle strength score, epileptic EEG characteristics, gait amplitude, and sleep stages. These data are stored in time-series format, with each record containing a device ID, original timestamp, and indicator value.

[0018] Optionally, clinical laboratory modal data include: white blood cell count (WBC), neutrophil percentage (Neut%), C-reactive protein (CRP), procalcitonin (PCT), serum potassium, serum sodium, serum chloride, blood glucose, and serum creatinine. Each record includes the test time, test item, and value.

[0019] Optionally, behavioral modal data include 24-hour activity steps, turning frequency, Montreal Cognitive Assessment (MoCA) cognitive score (MoCA, 0-30), medication adherence rate (0-1), and fall risk score (0-100). Each record includes the collection time, behavioral indicator type, and value.

[0020] Optionally, the image modality data includes postoperative computed tomography (CT) or magnetic resonance imaging (MRI) images. The raw image modality data is a DICOM format file containing a pixel matrix and metadata.

[0021] Specifically, step S100 may include the following steps.

[0022] Step S101: Timestamp alignment. Taking the end moment of the operation as T0, convert all multi-modal raw data into relative time coordinates with T0 as the origin. Specifically, for the data point x(t) collected at moment t, convert it to the relative time τ = t - T0. Then, uniformly resample all data to a frequency of once per minute. The specific operation is as follows: for high-frequency data such as heart rate (once per second), take the arithmetic mean of all sampling points within each minute as the sampling value for that minute; for low-frequency data such as test data (once every 6 hours), fill in null values between two samplings; for behavior data, distribute the cumulative value to each minute within that hour; for image data, retain the original sampling frequency for the image acquisition time points, and mark the non-image acquisition time points as missing.

[0023] Step S102: Differential filling of missing values. Adopt different filling strategies for different types of data. First, for time-series physiological data, use linear interpolation for filling. For the missing point t in the time series (where t1 < t < t2), the standard expression of its interpolation calculation formula is: ; when there are no valid values at both ends, fill with the global median.

[0024] Secondly, for clinical test data, fill with the mean value of the same disease type and the same number of days after surgery, and attach the isimputedclin flag (0 represents the original value, 1 represents the filled value). Then, for behavior data, use forward valid value filling, that is, fill the subsequent missing values with the most recent valid observation value.

[0025] Finally, for image data, use forward filling with the most recent image and attach a time decay weight. Among them, the acquisition of image data also includes: adjusting the window width and window level of the original CT image (brain window: window width 80, window level 30; bone window: window width 1500, window level 300), automatically segmenting the lesion area using the U-Net++ network, normalizing the image size to 224×224×3 through bilinear interpolation, and linearly mapping the pixel intensity to the [0,1] interval to obtain the image observation value as the image data.

[0026] For any moment t, the corresponding image observation value I(t) is taken from the most recent valid image moment t last , and its calculation formula is: The time decay weight w img (t) is calculated as follows: where t is the current time, the target time point for which the weight needs to be calculated or data needs to be filled. I(t) is the filled image value, the image data value used at moment t, and actually equals tlast The value at time t. last The most recent valid image time is the time point before (and including) the current time t, the closest time that image data was actually acquired. valid The effective observation set is the set of all time points in the dataset where image data was actually acquired. img (t) represents the time decay weight, indicating the credibility or importance of the currently filled data, which decreases exponentially as the time interval increases. λ is the decay coefficient, a parameter controlling the rate of weight decrease, which is fixed at 0.1 here. The time interval Δt = tt last The current time t and the time of the most recent image acquisition t last The time difference between them.

[0027] Forward padding assumes that the state remains unchanged when no new data is updated.

[0028] Exponential decay: The time decay weight is introduced to quantify the "freshness of the data". As Δt increases (i.e., the longer the time since the last image was captured), the time decay weight will quickly approach 0, indicating that the algorithm or model may no longer accurately reflect the current state of the image data and its reference value should be reduced.

[0029] Example: If 10 hours have passed since the last image (Δt=10) and λ=0.1, then the weight wimg=e −0.1 ×10=e −1 ≈0.368, which means that the weight of this data is only about 36.8%.

[0030] Different data types have different missing patterns. Differential imputation strategies can preserve the true distribution characteristics of the data to the greatest extent, while providing the model with the clinical meaning inherent in the missing information itself through flags and decay weights (for example, testing for missing information may mean that the condition is stable and does not require frequent testing).

[0031] Step S103: Outlier Removal and Correction. A two-stage filtering mechanism is used for outlier processing. The first stage is clinical range filtering: reasonable value ranges for each indicator are set based on prior clinical knowledge. For example, heart rate is within [30, 220] beats / min, and non-invasive intracranial pressure is within [0, 60] mmHg. Values ​​outside these ranges are marked as outliers. The second stage is statistical filtering: for data that passes the clinical range filtering, the three-standard-deviation principle is further applied. That is, for indicator X, its mean μ and standard deviation σ are calculated. If |x-μ|>3σ, it is marked as an outlier. For values ​​marked as outliers, local median correction is used: the median of the effective values ​​at two time points before and after the outlier is taken as the replacement value.

[0032] Step S104: Normalization. Using prior medical knowledge (clinical normal range) rather than statistical characteristics of the data (such as maximum and minimum values), the scaling boundary is defined to linearly map the data to the interval [−1,1].

[0033] For any eigenvalue x, its normalized value x norm The calculation formula is: Where x is the original feature value, which is the original clinical observation data to be processed (such as the specific value of a certain biochemical indicator).

[0034] x norm The normalized value is the mapped value, which theoretically ranges between [−1, 1] (if x is within the clinical range).

[0035] This is the lower limit of clinical normal; it is the minimum value of the normal reference range defined in medical clinical standards.

[0036] This represents the upper limit of clinical normal, which is the maximum value of the normal reference range defined in medical clinical standards for this characteristic.

[0037] For example, regarding heart rate, =60 times / minute =100 beats / min, normal heart rate (80 beats / min) is mapped to approximately -0.5, tachycardia (120 beats / min) to approximately 0.33, and bradycardia (50 beats / min) to approximately -0.67. This step unifies indicators with different dimensions to the same numerical range, eliminating the impact of dimensional differences on model training; at the same time, mapping is performed based on the clinical normal range, so that the degree of deviation from the normal range is directly reflected in the mapping results, enhancing the model's sensitivity to abnormal states.

[0038] Step S200: Extract temporal physiological features from physiological standardized data, extract clinical structured features from clinical laboratory standardized data, extract behavioral structured features from behavioral state standardized data, and extract image depth features from image standardized data.

[0039] Extracting temporal physiological features from physiologically standardized data can specifically include the following steps.

[0040] Step S211: Obtain the mean, variance, range, and slope of the physiological modality data over multiple sliding time windows to obtain the candidate feature set; For eight core physiological indicators—heart rate, blood oxygen saturation, body temperature, respiratory rate, systolic blood pressure, diastolic blood pressure, non-invasively estimated intracranial pressure (eICP), and cerebral perfusion pressure—four statistical measures were calculated within three sliding time windows: 1 hour, 4 hours, and 24 hours: mean, variance, range, and slope. The mean reflects the average level of the indicator within the time window. The variance reflects the degree of fluctuation of the indicator within the time window. A larger variance indicates more drastic fluctuations in the indicator within that time period, potentially suggesting physiological instability. The range reflects the maximum amplitude of change of the indicator within the time window. The range can capture extreme fluctuations in the indicator, such as a sharp increase in intracranial pressure. The slope reflects the trend of change of the indicator within the time window. It is obtained by linear regression of the data points within the window. A positive slope indicates an upward trend, and a negative slope indicates a downward trend. For example, a consistently positive and large eICP slope suggests progressively increasing intracranial pressure, a warning sign of intracranial hemorrhage or severe edema. Each window generates 4 statistics, and the 3 windows generate a total of 12 statistics. The 8 indicators generate a total of 96 candidate features.

[0041] Step S212: Calculate the mutual information retention value between candidate features and the prediction target, and select features that are statistically related to the prediction target from the candidate feature set based on the mutual information retention value; Mutual information (MI) is a measure in information theory used to gauge the degree of interdependence between two random variables. A higher mutual information value indicates a stronger correlation between the two variables. This step uses mutual information as an initial screening metric to calculate the correlation between each candidate feature and the predicted target (i.e., the true label Y, where Y=1 indicates the occurrence of complications, and Y=0 indicates the absence of complications).

[0042] The formula for calculating mutual information is: In this context, X represents a candidate feature, such as a statistic like "the mean of heart rate within a 1-hour window." The value space of X is divided into several discrete intervals based on the actual value range in the training data. In this embodiment, equal-frequency discretization is used to divide the continuous feature values ​​into 10 intervals. Y represents the true label, Y∈{0,1}. p(x,y) is the joint probability distribution of feature X with value x and label Y with value y, obtained by dividing the number of samples that meet the conditions in the training set by the total number of samples. p(x) is the marginal probability distribution of feature X with value x, obtained by dividing the number of samples in the training set where feature X falls within interval x by the total number of samples. p(y) is the marginal probability distribution of label Y with value y, obtained by dividing the number of samples in the training set with label y by the total number of samples.

[0043] In actual calculations, for each candidate feature f j (j=1,...,96), perform the following operations: (1) Feature fj The continuous values ​​are discretized into 10 equal-frequency intervals on the training set (each interval contains about 10% of the samples), resulting in the discretized feature values ​​xf.

[0044] (2) Statistical joint distribution: For each pair (x f ,y), where x f Given y ∈ {interval 1, interval 2, ..., interval 10} and y ∈ {0, 1}, calculate p(x f ,y)=count(xf,y) / Ntotal, where count(x,y) f ,y) represents the number of samples in the training set whose features fall within the interval xf and whose label is y, and Ntotal represents the total number of samples in the training set.

[0045] (3) Statistical marginal distribution: p(x f ) = count(xf) / Ntotal, where count(x) f ) represents the total number of samples whose features fall within the interval xf; p(y) = count(y) / Ntotal, where count(y) is the total number of samples with label y.

[0046] (4) Substitute into the mutual information formula to calculate I(f) j ;Y).

[0047] (5) For all 96 candidate features, according to the mutual information value I(f) j Sort Y from largest to smallest.

[0048] (6) Set the mutual information threshold θMI = 0.05 (this threshold is determined through cross-validation, so that the number of retained features is about 1 / 3 to 1 / 2 of the original number of features). Retain features with I(fj;Y)>θMI and proceed to the next round of screening. After the initial screening based on mutual information, 30-40 features are usually retained.

[0049] For example, in clinical data, the mutual information value between "the slope of eICP within a 24-hour window" and the label "intracranial hemorrhage" may be as high as 0.32, indicating a high correlation between the two; while the mutual information value between "the variance of body temperature within a 1-hour window" and "intracranial hemorrhage" may be only 0.01, indicating that the two are essentially unrelated. Initial screening using mutual information can quickly eliminate noisy features that are irrelevant or weakly correlated with the prediction target.

[0050] Step S213: Based on prior clinical knowledge, select temporal physiological features from features that are statistically relevant to the predicted target.

[0051] The initial mutual information screening retains features statistically relevant to the prediction target, but statistical relevance is not equivalent to clinical relevance. For example, a feature may have a coincidental statistical correlation with the label, but lack interpretability from a clinical pathophysiological perspective; or two features with similar mutual information may have vastly different clinical evidence, one of which is widely recognized as an important early warning indicator, while the other lacks clinical support. Therefore, this step introduces prior clinical knowledge to refine the features after the initial screening, ultimately identifying 16 high-contribution temporal physiological features.

[0052] The specific procedures for clinical preliminary screening are as follows: (1) Constructing a clinical prior knowledge base: Based on expert consensus in the field of neurosurgery and published clinical research findings, a knowledge base of physiological early warning indicators for postoperative complications of brain tumors was established. This knowledge base associates a set of known physiological indicator change patterns with each complication type (intracranial hemorrhage, epilepsy, infection, fall, deep vein thrombosis). For example:

[0053] Intracranial hemorrhage / increased intracranial pressure: progressive increase in eICP (24-hour slope > 2 mmHg / h), progressive decrease in CPP (24-hour slope < -1 mmHg / h), compensatory bradycardia (Cushing's response), and increased pulse pressure.

[0054] Infection / sepsis: Increased heart rate (1-hour average >90 beats / min), elevated body temperature (4-hour average >38.5℃), and increased respiratory rate (RR >20 breaths / min).

[0055] Epilepsy: Abnormal heart rate variability (significantly increased or decreased variance), changes in breathing patterns.

[0056] Deep vein thrombosis: There are no obvious specific physiological indicators; it relies more on clinical tests and behavioral characteristics.

[0057] Falls: There are no obvious specific physiological indicators; they depend more on behavioral characteristics.

[0058] (2) Feature-Clinical Concept Mapping: Map each candidate feature to the above clinical concepts. For example: the feature “24h ICPslope” is mapped to “progressive increase in intracranial pressure”; the feature “1h HRmean” is mapped to “average heart rate”; the feature “4h Tvariance” is mapped to “body temperature fluctuation”; and the feature “24h CPPslope” is mapped to “decreasing trend of cerebral perfusion pressure”.

[0059] (3) Clinical importance score: Each clinical concept was assigned a clinical importance weight (scored independently by 3 experts with the title of associate chief physician or above in neurosurgery, and the average score was taken, with a maximum score of 10 points). For example: "Progressive increase in intracranial pressure": 9.5 points; "Decrease trend in cerebral perfusion pressure": 9.0 points; "Compensatory bradycardia of heart rate": 8.5 points; "Elevated body temperature": 8.0 points; "Increased heart rate": 7.0 points; "Increased respiratory rate": 7.0 points; "Abnormal heart rate variability": 8.0 points.

[0060] (4) Comprehensive scoring and screening: For each candidate feature, its comprehensive screening score = α·Inorm + β·Cscore, where Inorm is the mutual information value normalized to the [0,1] interval; Cscore is the clinical importance score normalized to the [0,1] interval; α and β are weight coefficients. In this embodiment, α=0.4 and β=0.6, that is, the weight of clinical prior is higher than that of statistical correlation, so as to give priority to ensuring the clinical interpretability of the feature.

[0061] (5) Final selection: Sort by comprehensive score from high to low, and select the top 16 features as the final high-contribution time-series physiological features. If multiple features (such as the slopes of the 1h, 4h, and 24h windows) related to a key clinical concept (such as "progressive increase in intracranial pressure") are selected, the one with the clearest clinical significance (usually the 4h or 24h window) is retained first, and the rest are removed to ensure the diversity and redundancy of features are minimized.

[0062] As shown in Table 1, after the above two-stage screening, 16 high-contribution temporal physiological characteristics were finally obtained.

[0063] The aforementioned 16 characteristics are not used in isolation; they have an inherent physiological synergy. 24heICPslope reflects the long-term trend of intracranial pressure, 1heICPrange reflects short-term acute fluctuations, and 4heICPauc reflects the cumulative load. The combination of these three can comprehensively assess the intracranial pressure status: when the long-term trend is upward, short-term fluctuations are increasing, and the cumulative load exceeds the standard, the risk of intracranial hemorrhage or brain herniation is extremely high.

[0064] 24-hour CPP slope, 24-hour SBP mean, and 24-hour SBP variance are used to assess cerebral perfusion status. Decreased CPP and low and fluctuating SBP indicate a risk of cerebral ischemia.

[0065] 4hHRmean, 4hTmean, 4hRRslope, and 6hTslope collectively reflect the state of infection or inflammation. When increased heart rate, elevated body temperature, rapid breathing, and a rising trend in body temperature occur simultaneously, the risk of infection is extremely high.

[0066] 24hHRvariance and 24hHRICPcross reflect the state of autonomic nervous system function. Abnormal heart rate variability accompanied by activation of the Cushing response increases the risk of seizures or intracranial pressure crisis.

[0067] 4-hour respiratory rate (RR) slope, 1-hour SpO2 min, and 1-hour RR range are used to assess respiratory function. The simultaneous presence of increased respiratory rate, decreased blood oxygenation, and unstable respiratory rhythm suggests respiratory dysfunction or risk of aspiration.

[0068] Step S220: Extract image depth features from image standardization data.

[0069] This step aims to extract deep semantic features from postoperative images (head CT or MRI) that reflect the postoperative pathological state of brain tumors. These features include key information such as the extent of edema, the presence and size of hemorrhages, the degree of ventricular compression, and midline shift.

[0070] This step uses the ResNet-18 network architecture proposed by He et al. at the 2016 CVPR conference (see He, K., Zhang, X., Ren, S., & Sun, J. Deep presidual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778) as the image feature extractor. This network contains 17 convolutional layers and 1 fully connected layer, mitigating the vanishing gradient problem through residual connections. This embodiment uses the official PyTorch implementation of ResNet-18 (torchvision 0.12.0) and initializes it using weights pre-trained on the ImageNet dataset.

[0071] A ResNet-18 residual network, pre-trained on the large medical image dataset ImageNet, is used as the feature extractor. This network contains 17 convolutional layers and 1 fully connected layer, mitigating the gradient vanishing problem through residual connections. First, for the input normalized image tensor I, the original high-dimensional feature vector is extracted using the ResNet18 network: The standardized image tensor I is 3×H×W in size, and is the original feature tensor after preprocessing (such as normalization and cropping). The original feature vector is 512×1 in size. Inputting into a fully connected layer yields the dimensionality-reduced image depth features. Among them, the image depth feature F img The feature vectors are dimensionality-reduced, with a size of 128×1; W fc The weight matrix is ​​128×512 in size; b fc The bias vector is 128×1 in size. When no new images are available, it is persistently stored in a time-decayed memory pool. The exponential moving average (EMA) is employed, using a decay coefficient to balance the weights of "historical memory" and "current input," thus achieving smooth feature updates and persistence. At time t, the memory vector... The update formula is: , It represents the memory state updated at time t, integrating historical information and current features, with a 128×1 specification. This is the persistent memory vector of the previous time step (t-1), with a size of 128×1. This represents the current image depth features, with a size of 128×1, and is the feature vector extracted by the neural network from the new image input at time t. When t=0 (initialization), it is usually set to... Or the zero vector. β is the decay coefficient, taken as 0.99 in this paper, a hyperparameter controlling the degree of memory retention. The closer β is to 1, the more historical memory is retained, and the smoother the update; the smaller β is, the more sensitive the model is to the current input.

[0072] Step S230: Extract clinical structured features from standardized clinical laboratory data. First, perform one-hot encoding on categorical variables: for a categorical variable with K classes, convert it into a K-dimensional binary vector, where only one dimension is 1 (corresponding to the current class), and the rest are 0. Second, construct interaction features to capture the interactions between indicators. Interaction features include product-based interaction features reflecting the compensatory state; and ratio-based interaction features reflecting the intensity of inflammation. The formula for calculating product-based interaction features is: , Interactive characteristics reflecting the coordinated changes in intracranial pressure and heart rate; The estimated intracranial pressure value is obtained after preprocessing (such as normalization) to normalize intracranial pressure. This is the preprocessed heart rate value. The formula for calculating the interaction feature based on the ratio is: , The interaction characteristic reflects the relative intensity of the inflammatory response; the higher the ratio, the greater the risk of infection. Normalized C-reactive protein is a sensitive indicator reflecting the level of inflammation in the body. ϵ is the normalized white blood cell count, an indicator reflecting the activity level of the immune system; ϵ is a smoothing constant, with a value of 1×10⁻.8 This is used to prevent numerical calculation errors (singularities) caused by a denominator of 0. The compensatory mechanism of cerebral perfusion pressure (CPP) was captured. Under physiological conditions, when intracranial pressure (ICP) increases, the body may maintain cerebral blood flow perfusion by regulating heart rate (HR). This product term can quantify this "pressure-frequency" coupling state and reflects the load on cerebral hemodynamics better than a single indicator. It is actually calculating the "level of inflammatory proteins per unit of white blood cell response".

[0073] If WBC is elevated but CRP is not, the low ratio may indicate non-infectious stress (such as trauma or drug response). If CRP is significantly elevated and WBC is also elevated (or WBC is not elevated due to immunosuppression), the ratio will be significantly larger, which usually indicates a severe bacterial infection or a strong cytokine storm.

[0074] Then, all the original features are concatenated with the constructed cross features to form a high-dimensional feature vector. Finally, LightGBM feature importance selection is used: a LightGBM model is trained, and an importance score is calculated based on the number of times each feature is used as a split node in the decision tree (or the information gain brought by the split), and the 64-dimensional features with the highest importance scores are retained.

[0075] One-hot encoding enables categorical variables to be handled by linear models; cross features integrate the physiological significance of multiple indicators; feature importance screening eliminates redundant and noisy features, reducing the risk of model overfitting.

[0076] Specifically, the steps for training the LightGBM model include: Training data: Positive samples are patients who developed complications within 30 days after surgery, and negative samples are patients who did not develop complications. Each sample contains 69 features. Model configuration: Number of decision trees T=100, learning rate η=0.05, maximum depth max_depth=5, minimum number of samples per leaf node min_child_samples=20, L1 regularization coefficient lambda_l1=0.0, L2 regularization coefficient lambda_l2=0.01. Early stopping strategy is adopted: training stops when the validation set AUC does not improve for 20 consecutive rounds.

[0077] Training process: Histogram discretization is performed on the 69-dimensional features, and the number of discrete buckets is set to 256; the gradient is calculated by selecting the top 50% of samples according to the gradient one-sided sampling strategy; decision trees are built one by one, and the optimal split is found by traversing all features and split points in each tree; the number of times each feature is selected as a split node is recorded as the feature importance score; the trained LightGBM model and the feature index of 64 retained features are obtained.

[0078] Step S240: Extract structured behavioral features from standardized behavioral data. Standardize and encode the following data: 24-hour activity steps, turning frequency, Montreal Cognitive Assessment Scale cognitive score (0-30), medication adherence rate (0-1), and fall risk score (0-100). The standardization formula is: xstd=(x-μtrain) / σtrain, where μtrain and σtrain are the mean and standard deviation on the training set. The processed structured behavioral feature vector has 5 dimensions.

[0079] Step S200 yielded 16-dimensional temporal physiological features, 128-dimensional image depth features, 64-dimensional clinical structured features, and 5-dimensional behavioral process features. These features directly reflect the unique pathophysiological patterns after brain tumor surgery, providing crucial information for the model's high-precision prediction. In particular, the application of non-invasive estimation of intracranial pressure features allows intracranial pressure information, previously requiring invasive monitoring, to be obtained non-invasively via TCD, expanding the scope of risk monitoring and enabling more patients to benefit from accurate early warning, thereby more effectively reducing unplanned readmission rates and the incidence of serious adverse events.

[0080] Step S300: Fuse temporal physiological features, image depth features, clinical structured features, and behavioral structured features to generate a fused feature vector.

[0081] Specifically, step S300 may include the following steps. Step S300 adopts a cross-modal multi-head attention mechanism to project four types of heterogeneous features onto a unified semantic space, dynamically evaluates the importance of each modality through learnable attention weights, and generates an adaptively weighted fusion feature vector.

[0082] Step S310: Project the temporal physiological features, image depth features, clinical structured features, and behavioral structured features into the same semantic space to obtain the temporal physiological projection feature vector, image depth projection feature vector, clinical structured projection feature vector, and behavioral structured projection feature vector.

[0083] This step aims to eliminate dimensional differences (heterogeneity) between different modalities by mapping the original features of different dimensions to a unified dimensional (128-dimensional) projection vector through a modality-specific fully connected layer.

[0084] For any mode m, its projection formula is p m =W m ×F m +b m Among them, F m W is the original characteristic vector. mis the projection weight matrix, a mode-specific learnable parameter matrix responsible for linearly transforming dm-dimensional features to 128 dimensions. m p is the projection bias vector, 128×1, a modality-specific bias term used to adjust the distribution center of features in a uniform space. m The projected feature vector is 128×1, and the output vector is the result of the projection. All modalities are consistent along this dimension, which facilitates subsequent concatenation or attention calculation. The temporal projection parameters are also included. Image projection parameters Clinical projection parameters and behavioral projection parameters All are uniformly initialized using Xavier, starting with zero.

[0085] Time-series physiological characteristics F temporal (16-dimensional) Image depth features F img (128 dimensions), clinical structured features F clinical (64 dimensions) and behavioral structured features F behavior (5-dimensional) features are projected onto a unified 128-dimensional space through independent fully connected layers. The core purpose of this step is "alignment." In the original data, image features (128-dimensional) contain far more information than behavioral features (5-dimensional). Through independent projection layers, the model can learn how to "expand" low-dimensional features (such as behavior and time series) to a high-dimensional space, while "compressing" or "refining" high-dimensional features (such as images and clinical data) into the same space, so that p temporal A certain dimension in p img One dimension in the equation is semantically comparable.

[0086] This step maps multimodal features from different sources and dimensions to the same semantic space, making subsequent attention calculations comparable.

[0087] Step S320: Calculate the attention scores corresponding to physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data respectively by querying the vector components.

[0088] For any modality m, its attention score a m The calculation formula is: , p represents the transpose of the modal projection eigenvectors, which is usually implicit when calculating the dot product, or can be regarded as the inner product of two row vectors. m q represents the modal projection feature vector of the m-th mode after projection through a fully connected layer, which is 1×128. i To query a vector component, retrieve the value of the vector in its i-th dimension. m,i For modal feature components, the modal feature vector p mThe value in the i-th dimension. m This reflects the degree of matching between the m-th modality and the query vector q; a higher score indicates a more important modality. The query vector component q... i The sample was initialized using a normal distribution with a mean of 0 and a standard deviation of 0.01.

[0089] If p m The closer the direction is to q, the smaller the angle between them, and the better the dot product result a. m The larger the value, the more relevant the information contained in the modality is to the "mission objective".

[0090] If the two are orthogonal or in opposite directions, and the dot product result is small or negative, it indicates that the information contained in the modality may be irrelevant or redundant for the current task.

[0091] This score reflects the importance of each modality feature for risk prediction in the current patient condition. This step uses a dot product attention mechanism to enable the model to dynamically evaluate the information value of different modalities based on the patient's specific situation. For example, for patients with significant changes in imaging, the attention score for the imaging modality will be higher.

[0092] Step S330: Map the attention score to the positive real number domain using an exponential function, and obtain the dynamic attention weights of the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality through normalization. The sum of the dynamic attention weights of the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality is 1.

[0093] This achieves the effect of "dynamically allocating attention," ensuring that the numerical range of the fused feature vector remains consistent with the projected features of each modality. This step transforms the original score into a probability distribution through a flexible maximization operation, allowing for a direct comparison of the importance of different modalities while guaranteeing numerical stability.

[0094] For any modality m, its dynamic attention weight w m The calculation formula is: Where M represents the set of all modes.

[0095] The weight vector calculated using this formula forms a discrete probability distribution. This means that the model is essentially "voting" to decide which mode is most important. For example, if w m =0.7 means that the model believes that 70% of the information in the current sample mainly comes from image data.

[0096] Step S340: Using the dynamic attention weights of the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality, the temporal physiological projection feature vector, the image depth projection feature vector, the clinical structured projection feature vector, and the behavioral structured projection feature vector are weighted and summed to obtain a fusion feature vector that includes the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality.

[0097] For the nth sample or time step in the batch data, its fused feature vector The calculation formula is: ,in, These are the attention weights for modality m in the nth sample. It is the projected feature vector of mode m in the nth sample.

[0098] If we consider matrix operations for the entire sample batch, the formula can be expressed as follows: , where ⊙ represents element-wise multiplication under the broadcast mechanism (i.e., multiplying the weight vector by each row of the feature matrix).

[0099] The formula is essentially a linear combination of vectors in a high-dimensional space. It is not a simple concatenation of features, but rather an intelligent "selection" and "mixing" of information from different modalities.

[0100] This fused feature vector, serving as the sole input to the risk prediction model, contains information from all modalities, with the contribution of each modality dynamically determined by the model based on the context. This step achieves soft selection of multimodal information through adaptive weighted fusion, offering greater expressive power and flexibility than simple concatenation or fixed-weight fusion.

[0101] Furthermore, it also includes a step of training the attention fusion parameters in steps S310-S340, using temporal projection parameters. Image projection parameters Clinical projection parameters and behavioral projection parameters The query vector q (1×128) is used for training.

[0102] Training data: Multimodal aligned data from the training set is used. Each sample contains temporal physiological features, image depth features, clinical structured features, and behavioral structured features, as well as corresponding complication labels.

[0103] Loss function: ; Lfusion: Loss function for feature fusion module (multi-label binary cross-entropy loss, corresponding to 5 independent postoperative risk classifications) k=1,2,…,5: Risk type index (corresponding to epilepsy, intracranial hemorrhage, infection, fall, deep vein thrombosis in sequence) y k : The true label of the kth type of risk (y k (∈{0,1}, where 0 represents no risk and 1 represents risk) P k : The probability score of the k-th risk prediction output by the model (P) k (∈(0,1), calculated by step S430) Binary Cross Entropy (BCE) summation form is suitable for training objectives involving independent prediction of multiple risks. But at this time, U and S in S410 k (P and γ in S420 are fixed, and T_k in S430 is temporarily fixed to 1).

[0104] Optimized configuration: The optimizer uses Adam, with an initial learning rate η = 1 × 10^{-3} and weight decay = 1 × 10^{-3}. -4 Batch size = 64. Cosine annealing learning rate decay is used. Early stopping strategy: stop when the validation set AUC does not improve for 10 consecutive rounds.

[0105] Step S300 uses adaptive weighted fusion to dynamically adjust the weights of each modality feature based on the patient's specific condition. For example, for patients with significant changes in imaging findings, the attention weight of the imaging modality is automatically increased; for patients with significant fluctuations in vital signs, the weight of temporal physiological features is automatically increased. This dynamic fusion strategy improves the model's AUC from 0.82 to 0.88, directly enhancing the accuracy of early warning and thus more effectively reducing unplanned readmission rates and the incidence of serious adverse events.

[0106] Step S400: Perform orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping on the fused feature vector in sequence to obtain the comprehensive risk score, risk level and risk type for five risks: epilepsy, intracranial hemorrhage, infection, fall and deep vein thrombosis.

[0107] This step introduces a non-standard mathematical processing logic to achieve decoupled assessment of the five risk types. Conventional methods typically use independent linear layers for multi-task prediction, but different tasks share feature representations, leading to feature interference between tasks. This step decouples the risk types through orthogonal basis projection, forcing the feature subspaces of different risk types to be orthogonal to each other, fundamentally avoiding feature interference.

[0108] Step S410: Construct an orthogonal basis matrix and a risk-specific selection matrix, project the fused feature vectors onto five mutually orthogonal 16-dimensional risk spaces, and output decoupled feature vectors corresponding to the five risk types respectively.

[0109] This step achieves linear independence separation of different risk characteristics, avoiding the coupling problem of "one feature predicting multiple risks," and laying the foundation for subsequent independent risk assessment. Specifically, this step includes:

[0110] The Gram-Schmidt orthogonalization process is used to generate the product that satisfies... Orthogonal basis matrix of the identity matrix Its 80 column vectors form an orthonormal basis, which are pairwise orthogonal and have a magnitude of 1. U is a fixed construct and does not participate in gradient descent training, ensuring deterministic orthogonality.

[0111] The five risk types are assigned to 16-dimensional subspaces to construct risk selection matrices. The basis vectors of the corresponding subspace are extracted from the orthogonal basis matrix using column index extraction rules. Columns (16×(k-1)+1) to (16×k) are extracted from U. S k Satisfy: When i ≠ j, S i T ×S j =0{16×16}, meaning the subspaces for different risk types are orthogonal to each other. k=1 corresponds to columns 1-16, representing the risk of epilepsy; k=2 corresponds to columns 17-32, representing the risk of intracranial hemorrhage; k=3 corresponds to columns 33-48, representing the risk of infection; k=4 corresponds to columns 49-72, representing the risk of falls; k=5 corresponds to columns 73-80, representing the risk of deep vein thrombosis.

[0112] The fused feature vector F fuse Projecting each feature vector onto the k-th risk subspace yields the decoupled feature vector. Among them, the adoption of Calculate the decoupling eigenvector. This is the projection matrix.

[0113] Among them, the risk selection matrix also satisfies orthogonality. This ensures that different risk subspaces are geometrically orthogonal and that the dot product of vectors within the subspace is zero. Each component of the decoupled feature vector represents the projection capability of the fused feature vector onto the basis vectors of the corresponding risk subspace, and different decoupled feature vectors are linearly independent, thus achieving complete decoupling of risk features.

[0114] Step S420: The decoupled feature vectors are processed by calculating projection energy, normalization, and adaptive power sharpening to obtain the sharpening energy of each risk.

[0115] This step amplifies strong risk signals, suppresses noise, and adaptively enhances risk contrast based on energy distribution, making the energy proportions of different risks more distinct. Specifically, this step includes:

[0116] Calculate the original energy for each risk subspace, highlighting the contribution of the large-amplitude component.

[0117] The original energy calculation formula is: , where p=3.

[0118] Calculate the normalized ratio of the original energy to the total energy for each risk subspace to eliminate the energy dimension differences between different risks and achieve comparability. ,in, The normalized ratio of the original energy to the total energy for each risk subspace is ε = 1 × 10⁻⁶. −8 .

[0119] The adaptive power exponent is adjusted based on the maximum energy percentage to adapt to different energy distribution scenarios. The formula for calculating the adaptive power exponent is as follows: When the maximum energy percentage When it approaches 1, To amplify advantages; when the maximum energy percentage When it approaches 0.2, To maintain linearity.

[0120] The original energy of each risk subspace is amplified through a power-law transformation to either amplify the proportion difference of advantageous risks or maintain a uniform distribution, resulting in the sharpening energy of each risk space, thus improving risk differentiation. Sharpening Energy , This power-law transformation can adaptively adjust the contrast of risk signals according to the current energy distribution. When a certain risk is significantly dominant, the contrast is widened further; when all risks are insignificant, the original distribution is maintained.

[0121] Step S430: Based on the sharpening energy, original energy, and historical prediction standard deviation as inputs, the sharpening energy is mapped to a risk score through calibration using the total energy modulation factor and adaptive temperature coefficient.

[0122] This step allows the risk score to simultaneously reflect both "feature strength" and "model prediction confidence." The output is smoother when uncertainty is high and more explicit when confidence is high. Specifically, step S430 includes:

[0123] The sum of the original energies of the five risk subspaces is calculated to reflect the projected intensity of the fusion feature across all risk subspaces.

[0124] The tanh function is used to map the sum of the original energies of the five risk subspaces into confidence factors. The reference energy threshold E0 = 2.0, the output τ ∈ (0,1), the larger the original total energy Etotal, the closer τ is to 1, and the higher the confidence level.

[0125] The adaptive temperature coefficient is adjusted based on the prediction standard deviation to adapt to uncertain scenarios. Adaptive temperature coefficient: T k =1+λ×σ k , where σ k The historical forecast standard deviation is given, with an uncertainty amplification factor λ = 0.5, and the forecast standard deviation σ is given. k The larger the value of Tk, the higher the uncertainty; the larger the value of Tk, the smoother the sigmoid function.

[0126] The sharpening energy is converted into a risk score using a composite formula, which integrates confidence level and uncertainty calibration. The composite formula is as follows: When the confidence factor τ is low, the output is close to 0.5 (unable to determine), and when τ is high, it is determined by the sharpening energy E. k_sharp and Tk Dominant score. Risk score. A higher numerical value indicates a higher probability of the risk occurring. The (1-τ)×0.5 term provides a prior probability of 0.5 (indicating "cannot be determined") when the total energy is low, while the τ term makes the output dominated by the calibrated E_k_sharp when the total energy is high. After this step, five risk scores are output: P_epilepsy, P_intracranial hemorrhage, P_infection, P_fall, and P_deep vein thrombosis.

[0127] Furthermore, it also includes the temperature coefficient T. k The steps for conducting training.

[0128] Learnable parameter: Temperature coefficient baseline value T_base (initially 1.0), but T k =1+λ×σ k , where σ k The standard deviation of historical predictions is used for statistical calculation and is not included in gradient updates.

[0129] Calibration method: Platt scaling is used, and the optimal T is learned on the validation set by minimizing the negative log-likelihood loss. base : Calibrate the loss function: Platt scaling calibration probability formula: ; Where Lcal: negative log-likelihood loss function in the probability calibration phase (optimization objective: minimize this loss to learn the optimal temperature coefficient); Nval: total number of validation set samples; i: validation set sample index (i=1,2,…,Nval); yi: true label of the i-th sample (binary classification label, y i ∈{0,1}); :The predicted probability zi after Platt scaling calibration: the original output logit, that is, the unnormalized original output before calibration; Tbase: the base temperature scaling coefficient to be learned, the core parameter of Platt scaling, and the optimal value is obtained through optimization on the validation set.

[0130] Optimized configuration: Use the L-BFGS optimizer (quasi-Newton method), set the upper limit of the number of iterations to 100 times, and do not set the batch size (use the entire validation set).

[0131] Step S440: Output the comprehensive risk score, risk level and risk type by extracting the maximum value, quantifying and converting, and threshold judgment of the risk score.

[0132] This step converts the abstract risk score into an intuitive evaluation result, providing a clear risk prompt for clinical decision-making. Specifically, step S440 includes:

[0133] Screen out the maximum value among the 5 risk scores to determine the dominant risk.

[0134] Quantify the maximum risk score into a percentage integer score to obtain the comprehensive score and improve clinical readability.

[0135] Determine the low, medium, and high three risk levels according to the threshold range of the comprehensive score. When the comprehensive score Stotal ≤ 30, it is a low risk; 30 < Stotal ≤ 60, it is a medium risk; Stotal > 60, it is a high risk.

[0136] Output the result based on the risk type corresponding to the maximum score. If the maximum score is less than 0.5, it is determined as "no clear risk type". Among them, the risk type corresponding to the maximum score is determined by Determine.

[0137] Furthermore, the method also includes jointly fine-tuning all the learnable parameters of steps S310 - S430 to achieve overall optimization.

[0138] Loss function: The total loss function consists of three parts.

[0139] (1) Risk prediction loss Among them, the risk category weight w k is calculated according to the sample proportion of each type of risk in the training set: w k = N total / (N class × N k ). After calculation: w_infection = 2.64, w_intracranial hemorrhage = 3.79, w_epilepsy = 6.57, w_fall = 9.24, w_deep vein thrombosis = 6.57.

[0140] (2) Decoupling regularization loss Unlike conventional schemes that constrain features to be orthogonal, this scheme directly constrains the correlation between risk scores: ; where ρ(P i , P j ρ(Pj) is the Pearson linear correlation coefficient between the probability Pi of the i-th risk prediction and the probability Pj of the j-th risk prediction, representing the degree of linear correlation between the two types of risk predictions. i , P j ) = Cov(P i ,P j ) / (σ i ·σ j ); This loss term minimizes the linear correlation between different risk scores, encouraging the model to make independent judgments on different risk types.

[0141] (3) L2 regularization loss Where θ is a single trainable parameter of the model, including all learnable parameters such as network weights and biases; Θ is the set of all trainable parameters.

[0142] Total loss function: ; Where Ltotal is the total loss function of the model, the final optimization objective of model training, which is to minimize this value; Lpred is the risk prediction weighted loss (main task loss); λdecouple is the risk decoupling loss weight coefficient, a hyperparameter that controls the strength of the decoupling constraint, λdecouple=0.01; Ldecouple is the risk decoupling constraint loss; λreg is the L2 regularization loss weight coefficient, a hyperparameter that controls the regularization strength and suppresses overfitting, λreg=1×10 -4 The value is determined through grid search; Lreg is the L2 regularization loss. Optimized configuration: The optimizer uses Adam, with an initial learning rate η = 1 × 10⁻⁶. -5 A smaller learning rate is used to avoid corrupting already trained parameters; weight decay is 1 × 10⁻⁶. -4 Batch size = 32. Cosine annealing learning rate decay is used. Early stopping strategy: stop when the validation set AUC does not improve for 15 consecutive rounds. Gradient pruning threshold is set to 1.0.

[0143] A training process is as follows: Algorithm: End-to-end joint fine-tuning Input: Training set D_train, validation set D_val, pre-training parameters Θ_init Output: Optimal model parameters Θ* Initialization: Θ ← Θ_init, η ← 1×10⁻ 5 , best_AUC ← 0, patience_counter ← 0 for epoch = 1 to E_max (E_max=100): for batch in D_train: # Forward Propagation F_fuse ← Steps S100-S300 (batch) Z_k ← S_k^T × F_fuse E_k_raw ← (Σ|z_{k,i}|³)^{1 / 3} E_k_norm ← E_k_raw / (ΣE_j_raw + ε) γ ← 0.5 + 0.5×(1 - max(E_k_norm)) E_k_sharp ← (E_k_norm)^{γ} P_k ← (1-τ)×0.5 + τ×sigmoid((E_k_sharp-0.5) / T_k) # Loss Calculation L_pred ← -Σ[w_k×y_k×log(P_k) + (1-y_k)×log(1-P_k)] L_decouple ← Σ_{i≠j} |ρ(P_i, P_j)| L_total ← L_pred + 0.01×L_decouple + 1e-4×L_reg # Backpropagation ∇Θ ← ∂L_total / ∂Θ ∇Θ ← clip(∇Θ, -1.0, 1.0) # Gradient clipping # Parameter Update Θ ← Adam(Θ, ∇Θ, η, β1=0.9, β2=0.999) # verify AUC_val ← Evaluate the model's AUC on D_val η ← Cosine Annealing Update if AUC_val>best_AUC: best_AUC ← AUC_val Θ* ← Θ patience_counter ← 0 else: patience_counter ← patience_counter + 1 if patience_counter ≥ 15: break Output Θ* Step S500: Based on the marginal contribution of temporal physiological characteristics, clinical structured characteristics, behavioral structured characteristics, and image depth characteristics to the prediction results, retrieve authoritative medical knowledge bases to obtain evidence-based support and dynamically generate interpretable evidence.

[0144] This step, based on temporal physiological characteristics, clinical structured characteristics, behavioral structured characteristics, and imaging depth characteristics, and combined with the prediction results of step S440, dynamically generates evidence-based explanations through feature marginal contribution calculation, risk feature vector construction, authoritative medical knowledge base retrieval, and natural language fusion. Its core function is to overcome the limitations of conventional decision tree rule extraction by using a knowledge retrieval enhancement mechanism to ensure that the model's prediction results possess both "technical interpretability" and "medical authority," providing clear evidence support for clinical decision-making. This step specifically includes:

[0145] Step S510: Using all temporal physiological features, clinical structured features, behavioral structured features, imaging depth features, and risk type outputs as input, calculate the Shapley value for each feature using the KernelSHAP algorithm to filter out key features and their contribution. This step quantifies the marginal impact of individual features on the prediction results, identifies the core factors driving risk prediction, and provides "technical causal evidence" for subsequent interpretation. Further, step S510 includes:

[0146] The temporal physiological features, clinical structured features, behavioral structured features, and image depth features are integrated into a complete feature set N (N is the total number of all input features, and a single feature is denoted as i).

[0147] according to Calculate the value of ϕi by iterating through all subsets of the feature set and calculating the marginal contribution of each feature to the prediction result. S is any subset that does not contain feature i, |S| is the number of features in subset S, |N| is the total number of features, f(S) is the prediction result of the model using only subset S, f(S∪{i}) is the prediction result of the model after adding feature i, and [f(S∪{i})−f(S)] is the marginal contribution of feature i.

[0148] Sort the absolute values ​​of ϕi for all features, select the top 5 features, and retain their names, original values, and corresponding contributions.

[0149] In one example, the list of the top 5 key features by ϕi value is formatted as ["name":f1,"value":v1,"contribution":ϕ1,"name":f2,"value":v2,"contribution":ϕ2,...,"name":f5,"value":v5,"contribution":ϕ5] In this context, each "name" represents a feature name (e.g., "mean heart rate" or "intracranial hemorrhage imaging feature A"), "value" represents the original numerical value of the feature, and "contribution" represents a Shapley value (a positive value indicates that the feature promotes increased risk, while a negative value indicates that it inhibits risk).

[0150] This step quantifies the "contribution" of each feature to the prediction result, clarifying the core driving factors of risk prediction; it serves as the core input for constructing the risk feature vector in step S520, and also provides the technical basis for explaining "why the risk is determined" in step S540.

[0151] Step S520: Based on the key features, risk type output results, and abnormal indicators in the original features, construct a structured risk feature vector.

[0152] This step uses the Top-5 key features from step S510, the prediction results from step S440, and abnormal indicators from the original features as input to construct a structured risk feature vector Rfeatures. It integrates the three types of information—prediction results, key features, and abnormal signals—to form a standardized retrieval input, providing a unified and structured query basis for subsequent searches of authoritative medical knowledge bases. Specifically, step S520 includes:

[0153] The key features, risk type output results, and abnormal indicators in the original features are assembled according to a preset structure to ensure that the information in each field is complete.

[0154] Standardize the assembled text information (e.g., unify the expression of medical terminology) to avoid missed detections due to inconsistent terminology during retrieval.

[0155] The standardized, assembled text information is transformed into structured risk feature vectors (Rfeatures), ensuring that the fields are clear and machine-readable. For example, All fields are required. If the list of abnormal indicators has no data, it will be filled with an empty list. The terminology standardization rules adopt the "Medical Terminology Standard" (e.g., "deep vein thrombosis" is unified as "deep vein thrombosis formation") to ensure that the terminology in the search query is consistent with that in the knowledge base and to improve the accuracy of the search.

[0156] Step S530: Based on the risk feature vector, select highly relevant knowledge items from the preset authoritative medical knowledge base using a hybrid retrieval strategy.

[0157] This step links the model's "technical characteristics" with "medical evidence-based support," providing authoritative clinical guidelines, expert consensus, or literature support for subsequent interpretations and enhancing the credibility of the interpretations. Specifically, step S530 includes:

[0158] Based on the "risk_type" and "top_features" fields of the risk feature vector, a natural language query statement Q is generated, with the formula representing risk factors, such as "risk factors for intracranial hemorrhage, mean heart rate, intracranial high-density shadow, peak blood pressure". The formula is... .

[0159] A hybrid strategy of "BM25 text matching + vector cosine similarity" is employed to calculate the relevance score between query Q and each knowledge record D in the knowledge base. The formula for calculating the relevance score is as follows: BM25(Q,D) is the text matching score (range 0-1) between query Q and knowledge record D, with a weight of 0.4 (emphasizing keyword matching); CosineSimilarity(E(Q),E(D)) is the cosine similarity (range 0-1) between the vector representation of Q E(Q) and the vector representation of D E(D), with a weight of 0.6 (emphasizing semantic association); the technical significance of weight allocation is to balance "keyword exact matching" and "semantic similarity matching", thereby improving retrieval recall and accuracy.

[0160] Sort by relevance score and return the top five knowledge items with the highest relevance scores from highest to lowest.

[0161] The top five knowledge items in terms of relevance score are then weighted twice using a combination of "evidence level weight + time decay factor," and the highest-scoring knowledge items are ultimately selected. The final selection score is calculated using the following formula: The weighting of evidence levels (Wevidence) is as follows: Class A = 1.0, Class B = 0.8, Class C = 0.6. Technically, this prioritizes high-quality evidence. The time decay factor (Wtime) is the publication year of the current year (with a decay cap of 0.3). Technically, this prioritizes recently published literature to ensure the timeliness of the evidence. Ultimately, the top 2-3 points are retained, ensuring both sufficiency of evidence and avoiding information overload.

[0162] Step S540: Generate a natural language explanation by fusing the retrieved relevant knowledge entries, feature contribution, and risk feature vectors using templates.

[0163] This step organically combines "technical feature contributions" with "medical evidence-based support," producing easily understandable, logically coherent, and authoritative explanatory text to help medical professionals understand "why the model made this judgment" and "what the medical basis for this conclusion is." Based on this, this step specifically includes:

[0164] Risk conclusion information, key driving features, and medical evidence are extracted from relevant knowledge entries, feature contribution, and risk feature vectors. Risk conclusions include type, level, and score; key driving features include name and contribution direction; and medical evidence includes risk factor correspondence, evidence level, and source.

[0165] By matching key driving features with "risk factors" in knowledge entries, a "feature-medical basis" association is established. For example, the feature "intracranial high-density shadow" corresponds to the knowledge entry "risk factors for intracranial hemorrhage: intracranial high-density shadow".

[0166] Based on a pre-defined natural language template, the aforementioned related information is integrated into an explanatory text that includes conclusions, explanations of feature contributions, supporting medical evidence, and intervention recommendations.

[0167] The explanatory text was grammatically corrected and its fluency adjusted to avoid machine-like expression and ensure readability for medical staff.

[0168] Step S600: Generate a rehabilitation plan based on risk type, risk level, risk score, and interpretable evidence through knowledge graph matching and multi-hop reasoning.

[0169] This step takes risk type, risk level, risk score, explainable evidence, and key features as input. Through a specially constructed knowledge graph of brain tumor post-operative rehabilitation, combined with multi-hop causal reasoning, it discovers the causal chain of risk-intervention, generates personalized rehabilitation plans, and triggers tiered early warnings. This step overcomes the limitations of conventional template matching. Through the structured associations and reasoning capabilities of the knowledge graph, the rehabilitation plan possesses the characteristics of "causal traceability, personalized measures, and implementable execution," while tiered early warnings ensure the timeliness of high-risk interventions. Specifically, step S600 includes:

[0170] Step S610: Using authoritative medical data, including clinical guidelines, expert consensus, and rehabilitation pathway literature, as input, construct a rehabilitation knowledge graph that includes multiple types of nodes such as risk, indicators, thresholds, and interventions, as well as causal / correlation relationships.

[0171] This deficiency provides a structured medical knowledge carrier for subsequent multi-hop reasoning, transforming scattered rehabilitation knowledge into a computable "node-relationship" form to support the causal chain mining of risks and interventions. Specifically, this step includes:

[0172] Define risk nodes, indicator nodes, threshold nodes, intervention nodes, monitoring nodes, departmental nodes, and contraindication nodes to form a node set V, and specify a concrete instance for each node. Specifically, the risk node Vrisk corresponds to five risk categories (epilepsy, intracranial hemorrhage, infection, falls, deep vein thrombosis) and other common postoperative complications of brain tumors; the indicator node Vindicator corresponds to specific indicators of four input features (such as "C-reactive protein (CRP)", "procalcitonin (PCT)", "heart rate", "body temperature slope", etc.); the threshold node Vthreshold represents the clinically abnormal threshold for each indicator (such as "CRP > 50 mg / L", "PCT > 0.5 ng / mL", "body temperature slope > 0.5°C / hour"); and the intervention node Vthreshold represents the clinically abnormal threshold for each indicator (such as "CRP > 50 mg / L", "PCT > 0.5 ng / mL", "body temperature slope > 0.5°C / hour"). Vintervention (prevention node): Rehabilitation treatment and management measures (e.g., "antibiotic treatment", "suspension of rehabilitation training", "intravenous infusion of ceftriaxone"); Vmonitor (monitoring node): Indicator monitoring measures (e.g., "body temperature measurement every 2 hours", "CRP retest every 6 hours"); Vdept (department node): Clinical department implementing the intervention / monitoring (e.g., "ICU", "Nursing Department", "Laboratory Department", "Rehabilitation Department"); Vcontraindication (contraindication node): Contraindications to the intervention measures (e.g., "history of penicillin allergy", "contraindication to anticoagulants").

[0173] Based on medical logic, the relationships between nodes are analyzed, defining relationships such as indicator-risk, indicator-threshold, threshold-risk, intervention-prevention-risk, intervention-treatment-risk, monitoring-indicator, intervention-department, intervention-contraindication, and monitoring-intervention, with a clear semantic definition for each type of relationship. Specifically, "indicators" (indicator-risk relationship) indicates that an abnormal indicator can indicate the occurrence of risk, such as "procalcitonin" to "infection risk"; "exceeds-threshold" (indicator-threshold relationship) indicates that the indicator value exceeds a preset abnormal threshold, such as "C-reactive protein" to "CRP>50"; "triggers" (threshold-risk relationship) indicates that an indicator exceeding a threshold triggers risk (such as "CRP>50" to "infection risk"); "prevents" (intervention-risk relationship) indicates that intervention measures can prevent the occurrence of risk (such as "antibiotic treatment" to "infection risk"); and "treats" (intervention-risk relationship) indicates that intervention measures can treat existing risks. Risks that occur (e.g., "hemostatic drugs" to "risk of intracranial hemorrhage"); monitors (monitoring to indicators) indicate that the monitoring measures are for specific indicators (e.g., "measure body temperature every 2 hours" to "body temperature"); execute-by (intervention to department) indicates that the intervention measures are performed by a specific department (e.g., "cefotaxime intravenous infusion" to "ICU"); contraindicated-with (intervention to contraindications) indicates that the intervention measures conflict with a contraindication (e.g., "cefotaxime" to "history of penicillin allergy"); follows (monitoring to intervention) indicates that the monitoring measures need to be implemented in conjunction with the intervention (e.g., "recheck CRP" to "antibiotic treatment").

[0174] Node instances and relation instances are assembled into a structured knowledge graph G=(V,E,R) in the form of triples "Vi−R−Vj", ensuring the accuracy and completeness of the relations. Vi and Vj are two types of node instances, and R is a relation instance. For example, "C-reactive protein" (indicator node) is connected to "CRP>50" (threshold node) through the relation "exceeds_threshold", and "CRP>50" is connected to "infection risk" (risk node) through the relation "triggers".

[0175] Step S620: Map the risk type, key features, feature values ​​and contribution to the corresponding nodes in the knowledge graph, calculate the matching confidence and construct an anomaly indicator set.

[0176] This step establishes the connection between the "risk features output by the model" and the "structured nodes of the knowledge graph," providing a clear starting point and confidence level for subsequent multi-hop inference. Specifically, this step includes:

[0177] The output risk type results are directly matched to the risk nodes in the knowledge graph; Traverse the names of key features and perform terminological standardization matching with the indicator nodes of the knowledge graph; for example, “C-reactive protein” → “C-reactive protein (CRP)” node. If a match fails, mark it as an “unmatched indicator” (which will not be used for subsequent inference).

[0178] The successfully matched indicator value is compared with the preset indicator threshold at the corresponding threshold node to calculate the threshold matching degree. For example, CRP=95mg / L → “CRP>50mg / L” node; when the indicator value exceeds the upper limit of the threshold / belows the lower limit of the threshold, the threshold matching degree confthreshold=1.0, completely abnormal; when the indicator value is within the “threshold boundary ±10%” range, the threshold matching degree=0.5 (critically abnormal); when the indicator value is far from the threshold range, the threshold matching degree=0.0 (no abnormality). For each successfully matched "indicator-threshold" pair, the matching confidence is determined based on the threshold matching degree. The calculation formula is: matching confidence confmatch = ϕ × confthreshold, where ϕ reflects the weight of the indicator's impact on risk, and the threshold matching degree reflects the degree of indicator anomaly. ϕ originates from step S510, ranging from [-1, 1], and is used in the calculation after taking its absolute value, as only the positive contribution of abnormal indicators is considered. This step comprehensively considers the "weight of impact" and "degree of anomaly" of the indicator on risk, quantifying the reliability of the node mapping.

[0179] The indicator nodes, threshold nodes, and matching confidence scores are integrated to form a set of abnormal indicators.

[0180] Step S630: Using the mapping node, the set of abnormal indicators, and the evidence level of the evidence-based basis as input, the knowledge graph is traversed in a forward and backward multi-hop manner to calculate the risk confirmation score and the comprehensive score of the causal chain, and to select multiple high-confidence causal chains.

[0181] This step involves uncovering the complete causal logic of "abnormal indicators → risk occurrence → intervention / monitoring," ensuring that subsequent rehabilitation measures are directly related to the causes of the risk and avoiding unfounded recommendations. This step specifically includes:

[0182] Starting from the risk node, traverse forward along the relationship from intervention to risk prevention, from intervention to risk, and from monitoring to intervention to discover all reachable intervention nodes and monitoring nodes, and record the reasoning path.

[0183] Starting from the abnormal indicators in the abnormal indicator set, we traverse backwards along the relationships from indicators to risks and from indicators to thresholds to verify the reachability from monitoring nodes to risk nodes. Representing reachable nodes is 1, and representing unreachable nodes is 0.

[0184] Based on the reverse traversal results, a risk confirmation score is calculated. This score quantifies the degree to which anomaly indicators confirm risk. Risk Confirmation Score Formula is_reachable(Vindicato,Vrisk) is an indicator function. It is 1 if there is a valid path from the indicator node to the risk node, and 0 otherwise. confrisk∈[0,1], which is technically a quantification of the comprehensive confirmation of the current risk by all abnormal indicators. The higher the score, the clearer the cause of the risk.

[0185] By integrating intervention and monitoring nodes in the forward traversal and indicator-threshold-risk path in the reverse traversal, a complete causal chain or an extended chain containing monitoring nodes can be constructed.

[0186] C k =[V indicatori →V thresholdi →V risk →V interventionj ]. C k For a causal chain, V indicatori V is the i-th abnormal indicator. thresholdi For the i-th threshold, V risk V is a risk node. interventionj Let j be the j-th intervention node.

[0187] Calculate the overall score for each causal chain, and select the top-ranked causal chains based on their overall scores from highest to lowest.

[0188] The comprehensive scoring formula for causal chains is: scorechain∈[0,1]; weight conf risk Take 0.4, prioritizing causal chains with clearly defined risk causes; d is the inference path length, the number of nodes minus 1, such as V indicator →V threshold →V risk With d=2, a weight (1 / d) of 0.3 is used to prioritize chains with short paths and direct causal relationships (avoiding the uncertainty of complex paths); a weight of evidence_level of 0.3 is used to prioritize causal chains with authoritative medical evidence (Level A > Level B > Level C). This step integrates three dimensions—clarity of cause, simplicity of path, and authority of evidence—to select the optimal causal chain. The top-3 causal chains are selected by sorting the scorechain from highest to lowest. Technically, this ensures the targeted and sufficient nature of rehabilitation measures.

[0189] Step S640: Extract intervention and monitoring measures from the selected causal chains, taboo nodes and departmental nodes of the knowledge graph, prioritize them and check for taboos and conflicts, and generate a structured rehabilitation plan.

[0190] This step translates the abstract causal chain into actionable clinical measures, ensuring the timeliness, safety, and operability of the protocol through prioritization and contraindication checks. Specifically, this step includes:

[0191] All intervention and monitoring nodes are extracted from the screened causal chains, and duplicate measures are removed to form an initial set of measures. Duplicate measures are, for example, the same intervention appearing in multiple chains, and are retained only once.

[0192] A comprehensive scoring system based on causal chains prioritizes measures and clarifies implementation timelines.

[0193] The priority determination rule for the comprehensive score based on causal chains is as follows: High priority: Overall score ≥ 0.7, execution time limit "within 1 hour", indicating core measures that need to be implemented urgently; Medium priority: 0.4 ≤ overall score < 0.7, execution time limit "within 4 hours", indicating important measures that need to be implemented promptly; Low priority: Overall score <0.4, execution time limit "within 24 hours", indicating routine measures that can be implemented at an optional time; the execution priority is allocated according to the reliability of the causal chain to ensure that high-confidence measures are implemented first.

[0194] Each intervention node in the initial set of measures is traversed, and the corresponding contraindication node is queried through the intervention-to-contraindication relationship. This is matched with the patient's current state. If a conflict exists, the priority of the measure is lowered or it is removed directly, and a contraindication reminder is displayed. Lowering the priority of the measure, for example, means reducing it from high priority to medium priority.

[0195] By querying the departmental relationships of each intervention and monitoring measure, the implementing entity can be clearly identified.

[0196] The measures, priorities, implementing departments, implementation time limits, and contraindications are integrated into a standardized rehabilitation plan.

[0197] The following is an example of a rehabilitation plan: ===Rehabilitation Plan (Generated Based on Knowledge Graph Reasoning)=== [Risk Summary] - Risk type: Infection - Risk level: High risk (Level 2) Overall risk score: 85 points - Risk confirmation score: 0.82 [Causal Reasoning Chain] 1. C-reactive protein (95 mg / L) exceeding the threshold (>50) → indicates infection risk → triggers antibiotic treatment 2. Procalcitonin (1.2 ng / mL) exceeds the threshold (>0.5) → confirms infection risk → triggers enhanced surveillance. 3. A 6-hour temperature slope (0.8°C / hour) exceeding the threshold (>0.5) indicates infection progression and triggers a pause in rehabilitation training. Intervention measures

Taboos and Reminders

[0198] Based on the same inventive concept, this application also provides a virtual device for obtaining postoperative rehabilitation plans for brain tumor patients. In the embodiments of this application, the virtual device can be a software system deployed on a computer. For example, the virtual device can be a deep learning system built on the PyTorch framework, a machine learning system built on LightGBM, or a knowledge graph system built on Neo4j. Furthermore, the virtual device in the embodiments of this application can be deployed on devices with computing capabilities, such as servers, cloud computing platforms, and edge computing nodes.

[0199] In addition, in this embodiment of the application, physical resources can be understood as hardware resources, such as GPU resources, CPU resources, memory resources and storage resources.

[0200] Figure 2 Figure 3 is a schematic diagram of a virtual device for obtaining postoperative rehabilitation plans for brain tumor patients, provided in an embodiment of this application. This virtual device can be deployed on a computer or server to automatically identify risks and generate rehabilitation plans from multimodal data of postoperative brain tumor patients. Figure 3 As shown, the virtual device includes: a data acquisition and standardization unit 100, a feature extraction unit 200, a feature fusion unit 300, a risk identification unit 400, an interpretable evidence generation unit 500, a rehabilitation plan generation unit 600, and a training unit 700. Wherein:

[0201] The data acquisition and standardization unit 100 is used to acquire physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data of patients after brain tumor surgery, and to standardize the acquired data to obtain physiological standardized data, clinical laboratory standardized data, behavioral state standardized data, and imaging standardized data.

[0202] The feature extraction unit 200 is used to extract temporal physiological features from physiological standardized data, clinical structured features from clinical laboratory standardized data, behavioral structured features from behavioral state standardized data, and image depth features from image standardized data.

[0203] The feature fusion unit 300 is used to fuse temporal physiological features, image depth features, clinical structured features, and behavioral structured features to generate a fused feature vector.

[0204] The risk identification unit 400 is used to sequentially perform orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping on the fused feature vector to obtain risk scores for five risks: epilepsy, intracranial hemorrhage, infection, fall, and deep vein thrombosis, and generate risk level and risk type based on the risk scores.

[0205] The interpretable evidence generation unit 500 is used to retrieve evidence-based evidence from an authoritative medical knowledge base based on the marginal contribution of temporal physiological characteristics, clinical structured characteristics, behavioral structured characteristics, and image depth characteristics to the prediction results, and dynamically generate interpretable evidence.

[0206] The rehabilitation plan generation unit 600 is used to generate rehabilitation plans based on risk type, risk level, risk score, and interpretable evidence through knowledge graph matching and multi-hop reasoning.

[0207] Training unit 700 is used to train, validate and optimize the learnable parameters in feature extraction unit 200, feature fusion unit 300 and risk identification unit 400.

[0208] like Figure 2 As shown, the data acquisition and standardization unit 100 includes: a timestamp alignment module 150, a missing value filling module 160, an outlier handling module 170, and a normalization module 180. Wherein:

[0209] The timestamp alignment module 150 is used to convert all multimodal raw data into relative time coordinates with T0 as the origin, using the end time of the operation as T0, and uniformly resample to a frequency of once per minute, and output time-aligned multimodal data.

[0210] The missing value imputation module 160 is used to employ different imputation strategies for different types of data with missing patterns: linear interpolation is used for time-series physiological data; mean values ​​of the same disease and postoperative days are used for clinical laboratory data, with an is_imputed_clin flag added; forward valid values ​​are used for behavioral data; and forward imputation of the most recent image is used for imaging data, with a time decay weight added, outputting complete multimodal data.

[0211] The outlier processing module 170 is used to process outliers using a two-level filtering mechanism: the first level is clinical range filtering, which sets reasonable value ranges for each indicator based on prior clinical knowledge; the second level is statistical filtering based on the three-standard-deviation principle; for values ​​marked as outliers, local median is used for correction, and clean multimodal data is output.

[0212] Normalization module 180 is used to linearly map data to the [-1,1] interval using the clinical normal range defined by medical prior knowledge.

[0213] like Figure 2 As shown, the feature extraction unit 200 includes: a temporal physiological feature extraction module 210, an image depth feature extraction module 220, a clinical structured feature extraction module 230, and a behavioral structured feature extraction module 240. Wherein:

[0214] The temporal physiological feature extraction module 210 is used to obtain the calculated mean, variance, range and slope of physiological modality data in multiple sliding time windows to obtain a candidate feature set; calculate the mutual information retention value between the candidate features and the prediction target; select features that are statistically related to the prediction target from the candidate feature set based on the mutual information retention value; and select temporal physiological features from the features that are statistically related to the prediction target based on clinical prior knowledge.

[0215] The image depth feature extraction module 220 includes: a ResNet-18 feature extraction subunit 221, a fully connected dimensionality reduction subunit 222, and a time-decayed memory pool subunit 223. The ResNet-18 feature extraction subunit 221 inputs a standardized image tensor into a pre-trained ResNet-18 network, outputting a 512-dimensional original feature vector. The fully connected dimensionality reduction subunit 222 performs a linear transformation on the 512-dimensional original feature vector to reduce its dimensionality to 128 dimensions. The time-decayed memory pool subunit 223 updates Mt if a new image is input at time t and image depth features are extracted; otherwise, it maintains Mt. t = M t-1 Add time decay weights to output the memory vector M. t ∈ℝ 128 .

[0216] The clinical structured feature extraction module 230 is used to perform one-hot encoding on categorical variables; construct product interaction features and ratio interaction features; concatenate all original features with interaction features; use LightGBM to filter feature importance, retain the 64-dimensional features with the highest importance scores, and output clinical structured features.

[0217] The behavioral structured feature extraction module 240 is used to perform Z-score standardization on five behavioral indicators to obtain five-dimensional behavioral features F_behavior ∈ ℝ5 .

[0218] The feature fusion unit 300 includes: a feature projection module 310, an attention score calculation module 320, a weight normalization module 330, and a weighted summation module 340. The feature projection module 310 projects temporal physiological features, image depth features, clinical structured features, and behavioral structured features onto the same semantic space, obtaining temporal physiological projection feature vectors, image depth projection feature vectors, clinical structured projection feature vectors, and behavioral structured projection feature vectors. The attention score calculation module 320 calculates the attention scores corresponding to physiological modality data, clinical laboratory modality data, behavioral state modality data, and image modality data respectively by querying the vector components. The weight normalization module 330 maps the attention scores to the positive real number domain using an exponential function and obtains the dynamic attention weights for the physiological modality, clinical laboratory modality, behavioral state modality, and image modality through normalization. The sum of the dynamic attention weights for the physiological modality, clinical laboratory modality, behavioral state modality, and image modality is 1. The weighted summation module 340 is used to perform weighted summation on the temporal physiological projection feature vector, image depth projection feature vector, clinical structured projection feature vector, and behavioral structured projection feature vector using the dynamic attention weights of the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality, to obtain a fusion feature vector that includes the physiological modality, the clinical testing modality, the behavioral state modality, and the imaging modality.

[0219] like Figure 2 As shown, the risk identification unit 400 includes: an orthogonal basis projection decoupling module 410, a sharpening module 420, a risk score acquisition module 430, and a risk output module 440. Wherein:

[0220] The orthogonal basis projection decoupling module 410 is used to construct orthogonal basis matrices and risk-specific selection matrices, project the fused feature vectors onto five mutually orthogonal 16-dimensional risk spaces, and output decoupled feature vectors corresponding to the five risk types respectively.

[0221] The sharpening module 420 is used to calculate the projection energy, normalize, and perform adaptive power sharpening on the decoupled feature vectors to obtain the sharpening energy for each risk.

[0222] The risk score acquisition module 430 is used to map the sharpening energy to a risk score based on the sharpening energy, the original energy, and the historical prediction standard deviation as inputs, through calibration with the total energy modulation factor and the adaptive temperature coefficient.

[0223] The risk output module 440 is used to extract the maximum value, quantify and convert the risk score, and determine the threshold to output the comprehensive risk score, risk level and risk type.

[0224] Please see Figure 2 The interpretable basis generation unit 500 includes: a SHAP contribution calculation module 510, a risk feature vector construction module 520, a knowledge base retrieval module 530, and an interpretive text generation module 540. The SHAP contribution calculation module 510 uses all temporal physiological features, clinical structured features, behavioral structured features, image depth features, and risk type output results as input, and calculates the Shapley value of each feature using the KernelSHAP algorithm to filter out key features and their contributions. The risk feature vector construction module 520 constructs a structured risk feature vector based on key features, risk type output results, and abnormal indicators in the original features. The knowledge base retrieval module 530 uses the risk feature vector to filter highly relevant knowledge entries from a pre-set authoritative medical knowledge base using a hybrid retrieval strategy. The interpretive text generation module 540 uses template fusion to generate a natural language interpretation by combining the retrieved relevant knowledge entries, feature contributions, and risk feature vectors.

[0225] like Figure 2 As shown, the rehabilitation plan generation unit 600 includes: a knowledge graph construction module 610, a node matching module 620, a multi-hop reasoning module 630, and a plan generation module 640. Wherein:

[0226] The knowledge graph construction module 610 is used to construct a rehabilitation knowledge graph containing multiple types of nodes such as risk, indicators, thresholds, and interventions, as well as causal / correlation relationships, by taking authoritative medical data, including clinical guidelines, expert consensus, and rehabilitation pathway literature, as input.

[0227] The node matching module 620 is used to map risk type, key features, feature values ​​and contribution to corresponding nodes in the knowledge graph, calculate matching confidence and construct an anomaly indicator set.

[0228] The multi-hop reasoning module 630 is used to take mapping nodes, anomaly indicator sets, and evidence levels of evidence as inputs, and to calculate risk confirmation scores and comprehensive causal chain scores by traversing the knowledge graph in both forward and reverse directions, and to filter multiple high-confidence causal chains.

[0229] The solution generation module 640 is used to extract intervention and monitoring measures from the screened causal chains, taboo nodes and departmental nodes of the knowledge graph, prioritize them and check for taboos and conflicts, and generate a structured rehabilitation solution.

[0230] like Figure 3As shown, the training unit 700 includes: a LightGBM training module 710, a ResNet fine-tuning module 720, an attention fusion training module 730, a temperature coefficient calibration module 740, and an end-to-end joint fine-tuning module 750. Wherein:

[0231] LightGBM training module 710 is used to independently train the LightGBM model in clinical structured feature extraction module 230. It uses clinical laboratory data and behavioral data from the training set, totaling 1,478 samples. An early stopping strategy is employed: training stops when the validation set AUC does not improve for 20 consecutive rounds.

[0232] The ResNet fine-tuning module 720 is used to fine-tune the ResNet-18 and dimensionality reduction layers in the image depth feature extraction module 220. The first two residual blocks (Layer 1 and Layer 2) of ResNet-18 are frozen, and only Layer 3, Layer 4, and the newly added fully connected layers W_fc and b_fc are fine-tuned. The optimizer uses Adam with an initial learning rate η = 1 × 10^{-5} and an early stopping strategy: the learning stops if the validation set AUC does not improve for 10 consecutive rounds.

[0233] Attention fusion training module 730 is used to train the learnable parameters in feature fusion unit 300: W_temporal, b_temporal, W_img, b_img, W_clinical, b_clinical, and query vector q. The optimizer uses Adam with an initial learning rate η = 1 × 10⁻⁶. -3 Early stopping strategy: Stop when the validation set AUC does not improve for 10 consecutive rounds.

[0234] Temperature coefficient calibration module 740 is used to calibrate the temperature coefficient in risk identification unit 400. Platt scaling is employed to learn the optimal T_base on the validation set by minimizing the negative log-likelihood loss.

[0235] The end-to-end joint fine-tuning module 750 is used for joint fine-tuning of all learnable parameters.

[0236] In this embodiment, the virtual device deployed on the second computer is pre-installed with a program for risk identification and rehabilitation plan generation. When a rehabilitation plan needs to be generated, the virtual device starts the program and allocates the physical resources of the second computer to the running rehabilitation plan generation program. This ensures that the program is only started when a rehabilitation plan needs to be generated, thereby saving the resources of the second computer.

[0237] The virtual device in this embodiment uses an orthogonal basis projection decoupling module to force the feature subspaces of different risk types to be orthogonal to each other, fundamentally avoiding the feature interference problem in multi-task learning. The energy normalization module uses the L3 norm to calculate the projected energy, making it more sensitive to large-amplitude risk signals. An adaptive temperature coefficient is introduced through the risk output module, enabling the model to express its confidence level in its own predictions.

[0238] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for obtaining a postoperative rehabilitation program for brain tumor patients, characterized in that, include: We obtained physiological modality data, clinical laboratory modality data, behavioral status modality data, and imaging modality data of patients after brain tumor surgery, and standardized the obtained data to obtain physiological standardized data, clinical laboratory standardized data, behavioral status standardized data, and imaging standardized data. Extract time-series physiological features from physiologically standardized data, extract clinical structured features from clinical laboratory standardized data, extract behavioral structured features from behavioral state standardized data, and extract image depth features from image standardized data. The temporal physiological features, image depth features, clinical structured features, and behavioral structured features are fused to generate a fused feature vector. The fused feature vectors are sequentially subjected to orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping to obtain the comprehensive risk score, risk level and risk type for five risks: epilepsy, intracranial hemorrhage, infection, fall and deep vein thrombosis. Based on the marginal contribution of temporal physiological characteristics, clinical structured characteristics, behavioral structured characteristics, and image depth characteristics to the prediction results, authoritative medical knowledge bases are searched to obtain evidence-based support, and interpretable evidence is dynamically generated. A rehabilitation plan is generated based on the risk type, risk level, risk score, and interpretable evidence.

2. The method of obtaining according to claim 1, characterized in that, Extracting temporal physiological features from physiologically standardized data, including: The mean, variance, range, and slope of physiological modality data over multiple sliding time windows are obtained to obtain a candidate feature set; Calculate the mutual information retention value between candidate features and the prediction target, and select features that are statistically relevant to the prediction target from the candidate feature set based on the mutual information retention value; Based on prior clinical knowledge, temporal physiological features are selected from those statistically relevant to the predicted target.

3. The method of obtaining according to claim 1, characterized in that, The temporal physiological features, image depth features, clinical structured features, and behavioral structured features are fused to generate a fused feature vector, including: By projecting temporal physiological features, image depth features, clinical structured features, and behavioral structured features into the same semantic space, we obtain temporal physiological projection feature vectors, image depth projection feature vectors, clinical structured projection feature vectors, and behavioral structured projection feature vectors. By querying the vector components, attention scores are calculated for physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data, respectively. The attention scores are mapped to the positive real number domain using an exponential function, and the dynamic attention weights of the physiological modality, clinical laboratory modality, behavioral state modality, and imaging modality are obtained by normalization. The sum of the dynamic attention weights of the physiological modality, clinical laboratory modality, behavioral state modality, and imaging modality is 1. By utilizing the dynamic attention weights of the physiological modality, the clinical laboratory modality, the behavioral state modality, and the imaging modality, the temporal physiological projection feature vector, the imaging depth projection feature vector, the clinical structured projection feature vector, and the behavioral structured projection feature vector are weighted and summed to obtain a fusion feature vector that includes the physiological modality, the clinical laboratory modality, the behavioral state modality, and the imaging modality.

4. The method of obtaining according to claim 1, characterized in that, The fused feature vectors are sequentially subjected to orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping to obtain comprehensive risk scores, risk levels, and risk types for five risks: epilepsy, intracranial hemorrhage, infection, falls, and deep vein thrombosis. Construct orthogonal basis matrices and risk-specific selection matrices, project the fused feature vectors onto five mutually orthogonal 16-dimensional risk spaces, and output decoupled feature vectors corresponding to the five risk types respectively; The decoupled feature vectors are processed by calculating projection energy, normalization, and adaptive power sharpening to obtain the sharpening energy for each risk. Based on the sharpening energy, raw energy, and historical prediction standard deviation as inputs, the sharpening energy is mapped to a risk score through total energy modulation factor and adaptive temperature coefficient calibration. The risk score is extracted by maximum value extraction, quantification transformation and threshold judgment, and outputs a comprehensive risk score, risk level and risk type.

5. The method of obtaining according to claim 4, characterized in that, Construct orthogonal basis matrices and risk-specific selection matrices, project the fused feature vectors onto five mutually orthogonal 16-dimensional risk spaces, and output decoupled feature vectors corresponding to the five risk types, including: The Gram-Schmidt orthogonalization process is used to generate the product that satisfies... Orthogonal basis matrix of the identity matrix Its 80 column vectors form an orthonormal basis, which are pairwise orthogonal and have a magnitude of 1; The five risk types are assigned to 16-dimensional subspaces to construct risk selection matrices. The basis vectors of the corresponding subspace are extracted from the orthogonal basis matrix using column index extraction rules. k=1 corresponds to columns 1-16, representing the risk of epilepsy; k=2 corresponds to columns 17-32, representing the risk of intracranial hemorrhage; k=3 corresponds to columns 33-48, representing the risk of infection; k=4 corresponds to columns 49-72, representing the risk of falls; k=5 corresponds to columns 73-80, representing the risk of deep vein thrombosis. The fused feature vectors are projected onto the k-th risk subspace to obtain the decoupled feature vectors.

6. The method of obtaining according to claim 4, characterized in that, The decoupled feature vectors are processed by calculating projection energy, normalization, and adaptive power sharpening to obtain the sharpening energy for each risk; including: Calculate the original energy for each risk subspace, highlighting the contribution of the large-amplitude component; Calculate the normalized ratio of the original energy to the total energy for each risk subspace; The adaptive power exponent is adjusted based on the maximum energy percentage. The original energy of each risk subspace is amplified by power-law transformation to amplify the proportion difference of advantageous risks or to maintain a uniform distribution, thus obtaining the sharpened energy of each risk space.

7. The method of obtaining according to claim 4, characterized in that, Based on sharpening energy, raw energy, and historical prediction standard deviation as inputs, the sharpening energy is mapped to a risk score through total energy modulation factor and adaptive temperature coefficient calibration; including: Calculate the sum of the original energies of the five risk subspaces to reflect the projection intensity of the fusion feature across all risk subspaces; The tanh function is used to map the original energy sum of the five risk subspaces into confidence factors; Adjust the adaptive temperature coefficient based on the predicted standard deviation; The sharpening energy is converted into a risk score using a composite formula.

8. The method of obtaining according to claim 4, characterized in that, The risk score is processed through maximum value extraction, quantification transformation, and threshold determination to output a comprehensive risk score, risk level, and risk type; including: The maximum value among the five risk scores is used to determine the dominant risk. The maximum risk score is quantified into a percentage-based integer score to obtain the comprehensive score; Based on the threshold range of the comprehensive score, three risk levels are determined: low, medium, and high. Output results based on the risk type corresponding to the highest score.

9. A virtual device for obtaining postoperative rehabilitation plans for brain tumor patients, characterized in that, include: The data acquisition and standardization unit is used to acquire physiological modality data, clinical laboratory modality data, behavioral state modality data, and imaging modality data of patients after brain tumor surgery, and to standardize the acquired data to obtain physiological standardized data, clinical laboratory standardized data, behavioral state standardized data, and imaging standardized data. The feature extraction unit is used to extract temporal physiological features from physiologically standardized data, clinical structured features from clinical laboratory standardized data, behavioral structured features from behavioral state standardized data, and image depth features from image standardized data. The feature fusion unit is used to fuse temporal physiological features, image depth features, clinical structured features, and behavioral structured features to generate a fused feature vector. The risk identification unit sequentially performs orthogonal basis projection decoupling, energy normalization and power sharpening, and adaptive temperature coefficient mapping on the fused feature vector to obtain risk scores for five risks: epilepsy, intracranial hemorrhage, infection, fall, and deep vein thrombosis. Based on the risk scores, it generates risk level and risk type. The interpretable evidence generation unit is used to retrieve evidence-based evidence from authoritative medical knowledge bases based on the marginal contribution of temporal physiological characteristics, clinical structured characteristics, behavioral structured characteristics, and image depth characteristics to the prediction results, and dynamically generate interpretable evidence. The rehabilitation plan generation unit is used to generate rehabilitation plans based on risk type, risk level, risk score, and interpretable evidence through knowledge graph matching and multi-hop reasoning.

10. The virtual device according to claim 9, characterized in that, Also includes: The training unit trains, validates, and optimizes the learnable parameters in the feature extraction unit, feature fusion unit, and risk identification unit.

Images