Rectal cancer surgery risk assessment and early warning system based on multimodal data fusion

Abstract

The present invention relates to the technical field of assessment and early warning systems, specifically disclosing a rectal cancer surgical risk assessment and early warning system based on multimodal data fusion. The system comprises a multimodal data acquisition module, a data preprocessing module, a cross-modal fusion module, a dynamic risk assessment module, an intelligent early warning module, and a system self-checking module. By integrating multimodal data and dynamically assigning weights, the system achieves real-time updates of risks throughout the entire surgical cycle. Combined with graded early warning, autonomous detection and calibration, and risk tracing and countermeasure matching capabilities, the system comprehensively improves the comprehensiveness, timeliness, and accuracy of risk assessment, ensuring the long-term and reliable operation of the system, providing strong support for clinical decision-making, and effectively reducing surgical risks and the incidence of complications.

Description

Technical Field

[0001] The present invention relates to the technical field of assessment and early warning systems, and in particular to a rectal cancer surgery risk assessment and early warning system based on multimodal data fusion. Background Art

[0002] In the field of rectal cancer surgery, surgical risk assessment and early warning are of great significance for ensuring patient safety and optimizing surgical decision-making. However, existing technologies have the following significant shortcomings:

[0003] 1. Single data dimension and insufficient comprehensiveness of assessment: Traditional risk assessment methods often rely on a single type of data (such as clinical physiological indicators or imaging data), ignoring the synergistic effects of key information such as pathological characteristics and gene expression. For example, when determining the extent of tumor invasion based solely on CT images, it is difficult to consider the impact of KRAS gene mutation status on surgical difficulty, resulting in a one-sided risk assessment.

[0004] 2. Static assessments are unable to adapt to dynamic surgical changes: Existing systems often perform one-time risk assessments based on preoperative data, failing to consider the impact of real-time information such as the patient's physiological status (e.g., blood pressure fluctuations, blood oxygen levels), and tumor resection progress during surgery. If emergencies occur during surgery (e.g., massive bleeding), static assessment results cannot be updated in a timely manner, potentially delaying intervention.

[0005] 3. A crude early warning mechanism with low response accuracy: Traditional early warning systems often use a single threshold trigger (e.g., "alarm if risk exceeds threshold"), without differentiating between risk levels and urgency. For example, the same early warning method is used for different risks, such as "tumor adjacent to blood vessels" and "sudden drop in blood pressure during surgery." This makes it difficult for medical staff to quickly determine priority, impacting emergency response efficiency.

[0006] 4. Insufficient system reliability and a lack of self-testing and calibration mechanisms: Existing systems do not monitor the stability of data acquisition devices (such as sensors and imaging equipment) in real time. Device drift or data loss can lead to distorted assessment results. Furthermore, the lack of dynamic calibration of feature extraction and fusion models can lead to a decrease in accuracy over long-term use due to changes in data distribution. Summary of the Invention

[0007] In response to the deficiencies of the existing technology, the present invention provides a rectal cancer surgery risk assessment and early warning system based on multimodal data fusion, which solves the problems in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a rectal cancer surgical risk assessment and early warning system based on multimodal data fusion, comprising: a multimodal data acquisition module, a data preprocessing module, a cross-modal fusion module, a dynamic risk assessment module, an intelligent early warning module, and a system self-check module; the multimodal data acquisition module is configured to collect clinical data, imaging data, pathological data, and genetic data of the patient through a multi-source detection equipment group; the data preprocessing module is connected to the multimodal data acquisition module and performs noise reduction, standardization, and feature screening on the collected multimodal data; the cross-modal fusion module receives the multi-source data processed by the data preprocessing module, performs feature fusion through a deep learning network, and outputs a comprehensive feature vector; the dynamic risk assessment module constructs a real-time assessment model based on the comprehensive feature vector and outputs a surgical risk value and risk level; the intelligent early warning module is communicatively connected to the dynamic risk assessment module and triggers a multi-level early warning mechanism based on the risk value; and the system self-check module monitors and calibrates the operating status and data validity of the multimodal data acquisition module, the data preprocessing module, the cross-modal fusion module, the dynamic risk assessment module, and the intelligent early warning module in real time.

[0009] Preferably, the dynamic risk assessment module includes:

[0010] Preoperative assessment unit outputs basic risk value based on multimodal data within 72 hours before surgery , the calculation formula is:

[0011] ;

[0012] in, The feature weight is calculated by random forest Gini coefficient. The data used in the calculation of random forest Gini coefficient are derived from the physiological signals related to heart rate and blood pressure collected by the clinical parameter unit, the three-dimensional images of tumors obtained by the imaging detection unit, the tumor slice information collected by the pathological analysis unit, and the expression data of rectal cancer-related genes detected by the gene sequencing unit; and , is the risk contribution function;

[0013] Generate an initial risk benchmark by integrating the risk contributions of multi-dimensional features through weighted summation;

[0014] The intraoperative real-time unit updates the risk value every 5 minutes. The calculation formula is:

[0015] ;

[0016] in, is the attenuation coefficient, ranging from 0.01 to 0.05, which increases dynamically with the progress of the operation. The duration of the operation, is the real-time data weight, with an initial value of 0.3 and an increase of 0.1 every 30 minutes. The incremental risk of newly added data; through exponential decay, the influence of preoperative data is weakened, the weight of real-time intraoperative data is strengthened, and dynamic risk iteration is achieved;

[0017] The risk level classification unit includes the following steps: Step 1.1: Collect risk values ​​of 1,000 historical surgical data and use K-means clustering to generate 5 risk intervals; Step 1.2: Combined with clinical expert opinions, map the cluster centers to risk levels of 1-5, with level 1 being the lowest and level 5 being the highest; Step 1.3: When the risk values ​​updated twice in a row change across levels, trigger level calibration and recalculate the adjacent level thresholds.

[0018] Preferably, the multimodal data acquisition module includes:

[0019] The clinical parameter unit uses a multi-parameter monitor to synchronously collect physiological signals related to heart rate, blood pressure, and blood oxygen saturation, with a sampling frequency set to 200 Hz;

[0020] The imaging detection unit integrates 64-slice spiral CT and 3.0T MRI equipment to obtain three-dimensional tumor imaging data with a slice thickness controlled at 0.5-1mm;

[0021] Pathology analysis unit, which uses a digital pathology scanner to collect tumor tissue slice images with a resolution of no less than 0.2 μm / pixel;

[0022] The gene sequencing unit uses second-generation sequencing technology to detect the expression levels of 23 colorectal cancer-related genes related to KRAS and BRAF.

[0023] Preferably, the standardization processing of the data preprocessing module includes:

[0024] The physiological signals collected by the clinical parameter unit are normalized using Z-score:

[0025] ;

[0026] in, is the original physiological parameter value, is the mean value of this parameter in the clinical database, is the reference standard deviation; by converting the data into a standard normal distribution, the dimensional differences of different physiological parameters are eliminated, and cross-parameter comparability is achieved;

[0027] Normalize the grayscale of the image data of the image detection unit:

[0028] ;

[0029] in, is the original pixel value of the image, The maximum and minimum values ​​of pixels in a single image, respectively; the image grayscale value is mapped to the range of 0-255 to unify the image brightness benchmark for different devices and different shooting conditions;

[0030] The feature screening steps specifically include: Step 2.1: Calculate the Pearson correlation coefficient between each feature and surgical risk ,reserve features; Step 2.2: Use recursive feature elimination method, take the out-of-bag error of random forest model as indicator, and iteratively remove the 5% features with the lowest contribution; Step 2.3: Further compress the feature dimension through L1 regularization and retain the feature subset with non-zero coefficients.

[0031] Preferably, the cross-modal fusion module adopts an attention mechanism fusion network, including:

[0032] Feature extraction layer, which uses a fully connected network to extract feature vectors from clinical data , ResNet50 is used to extract feature vectors from image data ;

[0033] Attention allocation layer, calculates the weight coefficient of clinical data modality:

[0034] ;

[0035] Similarly, calculate the weight coefficient of the image data modality , the weight coefficient of the pathological data modality , the weight coefficient of the genetic data modality , and satisfies ;

[0036] in, is the cosine similarity function , used to measure the eigenvector and weight matrix degree of matching; It is a 32×512 dimensional learnable matrix, which is optimized through training to highlight the high contribution modes. is the feature vector obtained after feature extraction of clinical data, is the feature vector obtained after feature extraction of image data, is the feature vector obtained after feature extraction of pathological data, is the feature vector obtained after feature extraction of gene data;

[0037] Fusion layer, output comprehensive feature vector , the formula is:

[0038] ;

[0039] The fusion network training steps include: Step 3.1: Initialize the network parameters and use Xavier to initialize the weight matrix W; Step 3.2: Train with the cross-entropy loss function; Step 3.3: Perform a modal importance assessment every 10 epochs and remove modal features whose contribution is continuously less than 5%; Step 3.4: When the validation set accuracy improves by less than 0.5% for 5 consecutive epochs, stop training and save the model parameters.

[0040] Preferably, the intelligent early warning module further includes:

[0041] The risk tracing unit identifies the main risk factors through SHAP value calculation:

[0042] ;

[0043] in, is a feature subset, Represents the total feature set Eliminate the After all the possible feature subsets, that is, Does not include Any combination of features, For feature subset The number of features contained in The weight coefficient of the value is used to balance the impact of feature subsets of different sizes on the results. is the total number of features, is the model output function; this formula quantifies the marginal contribution of a single feature to the risk assessment result. The top 3 features with the highest absolute values ​​are considered as the main risk factors. is the marginal contribution difference, indicating that Add the The change in the model output after the feature is added;

[0044] Countermeasure generation unit, intervention plan matching steps: Step 4.1: semantically match the main risk factors with the preset plan library; Step 4.2: calculate the expected effect score of the plan: ; Step 4.3: Push the three options with the highest scores and mark the contraindications of each option.

[0045] Preferably, the warning threshold setting and warning intensity calculation basis of the intelligent warning module are as follows:

[0046] When 3≤risk level<4, a level 1 warning is triggered. , the calculation formula is:

[0047] ;

[0048] When the risk level is ≥4, the second level warning is triggered, and the warning intensity is , the calculation formula is:

[0049] ;

[0050] The setting of warning intensity is based on the main risk factors identified by the risk tracing unit, and the risk factors are assigned The weight corresponding to the value is mapped to the range of 0-1 after weighted summation of the actual data value, which intuitively reflects the urgency of the risk;

[0051] The risk thresholds are level 3 and level 4 respectively. The highest risk value in history; the warning intensity ranges from 0 to 1, which determines the frequency of warning signals;

[0052] Response steps after the warning is triggered: Step 5.1: When the first-level warning is triggered, the risk factor analysis report is automatically pushed to the surgeon's terminal; Step 5.2: When the second-level warning is triggered, the operating room sound and light alarm is triggered simultaneously, and the backup medical team dispatch mechanism is activated; Step 5.3: The warning is released when the risk assessment value is <3 for three consecutive times and falls back to the low-risk range.

[0053] Preferably, the system self-check module performs the following operations:

[0054] Data validity test, calculate the integrity coefficient of each modal data, the calculation formula is:

[0055] ;

[0056] Among them, the integrity coefficient is calculated for the four modalities of clinical, imaging, pathology, and genetics respectively:

[0057] Clinical parameter modality: is the total data item of physiological signals related to heart rate and blood pressure, Valid data items without missing data or abnormal fluctuations;

[0058] Image data modality: is the total number of slices in CT / MRI, The number of clear slices without artifacts and with a thickness that meets the requirement of 0.5-1 mm;

[0059] Pathology data modality: is the total field of view of the tumor section, The number of effective fields with a resolution ≥ 0.2 μm / pixel and no staining bias;

[0060] Genetic data modality: is the total detection items of 23 target genes, Valid gene items with sequencing quality value Q30 ≥ 80%;

[0061] ; The integrity factor is used to quantify the data quality, The data is determined to be invalid;

[0062] Calibration Procedure: Step 6.1: When When the abnormal module is automatically located; Step 6.2: Perform baseline calibration on the abnormal module; Step 6.3: Collect 3 sets of verification data after calibration, if all meet the requirements , then normal operation is restored; otherwise, a hardware failure alarm is triggered.

[0063] Preferably, the gene expression level of the gene sequencing unit is normalized using:

[0064] ;

[0065] in, is the raw sequencing read, is the average reading of the internal reference gene; logarithmic transformation is used to compress the numerical range of highly expressed genes while retaining the differential information of lowly expressed genes;

[0066] The basis of this standardization method is to measure the gene expression of standard cell lines under three different experimental conditions, construct the gene correlation function, and select the stable interval with the function value fluctuation ≤5% as the benchmark; Average read counts compared to the reference gene Substituting into the formula, through The conversion compresses the numerical range of highly expressed genes while retaining the differential information of lowly expressed genes, making the gene expression levels of different samples comparable across experimental conditions and meeting the needs of collaborative analysis of genetic data and clinical and imaging-related data in multimodal fusion.

[0067] Preferably, the risk contribution function Using the Sigmoid function:

[0068] ;

[0069] in, For the The contribution function of each feature to the surgical risk is Mapped to the [0,1] interval, it indicates the contribution ratio of the feature to the overall risk. is the slope parameter, which reflects the sensitivity of the feature to the risk, and its value ranges from 0.5 to 2. is the threshold parameter; the Sigmoid function maps the eigenvalue to the range of 0-1, quantifying the contribution of a single feature to the overall risk.

[0070] The present invention effectively overcomes the limitations of existing technologies through the collaborative design of multimodal data fusion, dynamic risk assessment, and intelligent early warning mechanisms, and has the following beneficial effects:

[0071] 1. The system uses a multimodal data acquisition module to simultaneously acquire clinical physiological signals (heart rate, blood pressure, etc.), imaging data (CT / MRI tumor characteristics), pathological section information, and the expression levels of rectal cancer-related genes, breaking through the limitations of traditional single-data evaluation. Combined with the dynamic weight allocation of different data types by the cross-modal fusion network, it can comprehensively consider multi-dimensional factors such as tumor aggressiveness, patient physiological reserve, and genetic susceptibility, significantly improving the ability to identify complex risks (such as occult vascular invasion and high-recurrence risk subtypes).

[0072] 2. The dynamic risk assessment module realizes real-time risk updates from preoperative to intraoperative: preoperatively, a basic risk benchmark is generated based on multimodal data within 72 hours. During surgery, the risk value is adjusted every 5 minutes based on the latest monitoring data, and the influence weight of real-time data is gradually strengthened as the operation progresses. This dynamic iterative mechanism can timely capture the impact of intraoperative emergencies (such as bleeding and fluctuations in vital signs) on risk, solving the problem that traditional static assessments cannot adapt to dynamic changes in surgery.

[0073] 3. The intelligent early warning module triggers differentiated responses based on risk levels: a risk factor analysis report is pushed during a level one warning to assist doctors in adjusting surgical strategies; a level two warning simultaneously activates audio and visual alarms and dispatches backup teams to ensure rapid handling of emergencies; by quantifying the intensity of warnings and matching corresponding response measures, medical staff can accurately prioritize risks and shorten emergency decision-making time.

[0074] 4. The system's self-check module automatically identifies abnormal equipment or invalid data and triggers the calibration process by monitoring the operating status and data integrity of each module in real time. After calibration, the validity is confirmed through multiple sets of verification data to ensure the quality of data collection and the stability of evaluation results, reduce the risk of misjudgment due to equipment drift or data loss, and ensure the long-term reliable operation of the system.

[0075] 5. Through the risk tracing unit, key risk factors (such as specific gene mutations and abnormal tumor differentiation) are located, and combined with the countermeasure generation unit to match the preset intervention plan library, high-success rate and high-convenience response measures are quickly pushed, while marking contraindications; this function provides clear decision-making references for medical staff during surgery, helps reduce complications and optimize surgical pathways.

[0076] In summary, the present invention integrates multimodal data and dynamically assigns weights to achieve real-time updates of risks throughout the entire surgical cycle. Combined with graded warning, autonomous detection and calibration, and risk tracing and countermeasure matching functions, it comprehensively improves the comprehensiveness, timeliness, and accuracy of risk assessment, ensures the long-term reliable operation of the system, provides strong support for clinical decision-making, and effectively reduces surgical risks and the incidence of complications. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a principle block diagram of the rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to the present invention;

[0078] Figure 2 This is a principle block diagram of the multimodal data acquisition module in the rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to the present invention;

[0079] Figure 3 This is a principle block diagram of the cross-modal fusion module in the rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to the present invention;

[0080] Figure 4 This is a principle block diagram of the dynamic risk assessment module in the rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to the present invention;

[0081] Figure 5 This is a principle block diagram of the intelligent warning module in the rectal cancer surgery risk assessment and warning system based on multimodal data fusion described in the present invention. DETAILED DESCRIPTION

[0082] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0083] like Figure 1-Figure 5As shown, the present invention provides a technical solution: a rectal cancer surgery risk assessment and early warning system based on multimodal data fusion, comprising: a multimodal data acquisition module, a data preprocessing module, a cross-modal fusion module, a dynamic risk assessment module, an intelligent early warning module and a system self-check module; the multimodal data acquisition module is configured to collect the patient's clinical data, imaging data, pathological data and genetic data through a multi-source detection equipment group; the data preprocessing module is connected to the multimodal data acquisition module and performs noise reduction, standardization and feature screening on the collected multimodal data; the cross-modal fusion module receives the multi-source data processed by the data preprocessing module, performs feature fusion through a deep learning network, and outputs a comprehensive feature vector; the dynamic risk assessment module constructs a real-time assessment model based on the comprehensive feature vector and outputs a surgical risk value and risk level; the intelligent early warning module is communicatively connected to the dynamic risk assessment module and triggers a multi-level early warning mechanism according to the risk value; the system self-check module monitors and calibrates the operating status and data validity of the multimodal data acquisition module, the data preprocessing module, the cross-modal fusion module, the dynamic risk assessment module and the intelligent early warning module in real time.

[0084] More specifically, Figure 4 As shown, the dynamic risk assessment module includes: a preoperative assessment unit, an intraoperative real-time unit and a risk level classification unit;

[0085] The preoperative assessment unit calculates the basic risk value based on the patient's multimodal fusion feature vector F within 72 hours before surgery. , as the initial benchmark for surgical risk, is calculated as:

[0086] ;

[0087] in, is the feature dimension of the comprehensive feature vector F (1024 dimensions); For the The weight of each feature is calculated using a random forest model: the random forest is configured with 1000 decision trees, with a maximum depth of 20 for each tree and a minimum number of samples for leaf nodes of 5. The feature importance is calculated based on the Gini coefficient. The data used in the calculation of the random forest Gini coefficient come from the physiological signals related to heart rate and blood pressure collected by the clinical parameter unit, the three-dimensional images of the tumor obtained by the imaging detection unit, the tumor slice information collected by the pathological analysis unit, and the expression data of rectal cancer-related genes detected by the gene sequencing unit; (such as the tumor infiltration depth feature =0.12, KRAS gene mutation characteristics =0.08); is the risk contribution function, and the eigenvalue is mapped to the 0-1 interval to represent the risk contribution ratio.

[0088] Implementation process: Input multimodal fusion features within 72 hours before surgery (including clinical, imaging, pathological, and genetic characteristics); calculated dimension by dimension , sum to get the basic risk value (Value range 0-100); Example: If a patient's tumor diameter characteristic , corresponding weight , then the risk contribution of this feature is The weighted sum of all features is obtained point.

[0089] The intraoperative real-time unit receives new multimodal data (such as intraoperative physiological signals and surgical process images) every 5 minutes and updates the risk value. , the calculation formula is:

[0090] ;

[0091] in, The real-time risk value at t minutes into the operation; is the attenuation coefficient, with an initial value of 0.01, which increases dynamically as the operation progresses: 0-60 minutes after the operation =0.01, 61-120 minutes =0.03, more than 120 minutes = 0.05, reducing preoperative data by exponential decay the impact of; is the duration of the operation (minutes), starting from the moment of skin incision; The weight of real-time data is 0.3 initially and increases by 0.1 every 30 minutes (upper limit 0.9), which strengthens the influence of intraoperative data. The incremental risk of the newly added data is calculated using real-time intraoperative features (such as bleeding volume and blood pressure fluctuations): the same random forest weights as before the operation are used. , but the features are replaced with real-time data from the past 5 minutes, and the output is an incremental value from 0 to 100.

[0092] Implementation process: Trigger an update every 5 minutes: Collect intraoperative data within the past 5 minutes (such as when the amount of bleeding is >500ml, Increase 15 points); calculate the attenuation term and real-time items , and we get ; Example: When the operation is 60 minutes in, ,but .

[0093] The risk classification unit maps the risk value to levels 1-5 (level 1 is the lowest and level 5 is the highest) and dynamically calibrates the threshold. The steps are as follows:

[0094] Step 1.1: Collect the complete risk value series of 1000 rectal cancer surgeries (including preoperative and intraoperative ); Use K-means clustering algorithm: set the number of clusters to 5, the initial cluster centers to [20, 40, 60, 80, 100], and stop when the center change is <0.1, generating 5 risk intervals (such as [0-25), [25-45), [45-65), [65-85), [85-100]).

[0095] Step 1.2: Combine the opinions of 5 gastrointestinal surgery experts and map the cluster centers to risk levels:

[0096] Level 1 (low risk): cluster center 20, corresponding risk value <25;

[0097] Level 2 (medium-low risk): cluster center 40, corresponding to 25≤risk value<45;

[0098] Level 3 (medium risk): cluster center 60, corresponding to 45≤risk value<65;

[0099] Level 4 (medium-high risk): cluster center 80, corresponding to 65≤risk value<85;

[0100] Level 5 (high risk): Cluster center 100, corresponding risk value ≥ 85.

[0101] Step 1.3: Dynamic calibration When the risk value of two consecutive updates changes across levels (e.g. , jumping from level 2 to level 4), triggering threshold calibration; recalculating adjacent level thresholds: using the risk value distribution of the latest 100 surgical data, taking the median value of adjacent cluster centers as the new threshold (e.g., the original level 2-3 threshold of 45 is adjusted to 48 after calibration).

[0102] More specifically, Figure 2 As shown, the multimodal data acquisition module includes: a clinical parameter unit, an imaging detection unit, a pathological analysis unit, and a gene sequencing unit; the clinical parameter unit uses a multi-parameter monitor to synchronously collect physiological signals related to heart rate, blood pressure, and blood oxygen saturation, with a sampling frequency set to 200Hz; the imaging detection unit integrates a 64-row spiral CT and a 3.0T MRI device to obtain three-dimensional tumor imaging data, with a layer thickness controlled at 0.5-1mm; the pathological analysis unit uses a digital pathology scanner to collect tumor tissue slice images with a resolution of not less than 0.2μm / pixel; the gene sequencing unit uses second-generation sequencing technology to detect the expression levels of 23 colorectal cancer-related genes related to KRAS and BRAF;

[0103] The gene expression level of the gene sequencing unit is normalized by:

[0104] ;

[0105] in, The original sequencing reads, that is, the number of gene fragments detected by the second-generation sequencing technology, directly reflect the original expression level of the gene. The average reading of internal reference genes, such as GAPDH and ACTB (which are stably expressed in cells and not affected by experimental conditions), is used as the internal reference genes. By calculating the average sequencing reading of multiple internal reference genes, batch differences in sample preparation and sequencing are eliminated. Logarithmic transformation is used to compress the numerical range of highly expressed genes while retaining the differential information of lowly expressed genes.

[0106] The basis of this standardization method is: by measuring the gene expression of the standard cell line under three different experimental conditions (such as different sequencing batches, reagent batches), constructing the gene correlation function (with the expression of 23 target genes as parameters), and selecting the stable interval with the function value fluctuation ≤5% as the benchmark; Average reads compared to reference genes (GAPDH, ACTB, etc.) Substituting this into the formula, the log2 transformation is used to compress the numerical range of highly expressed genes (such as genes with expression levels >10,000), while retaining the differential information of lowly expressed genes (such as genes with expression levels <100). This makes the gene expression levels of different samples comparable across experimental conditions, meeting the needs of collaborative analysis of genetic data and clinical and imaging-related data in multimodal fusion.

[0107] First pass Calculate the expression ratio of the target gene to the reference gene to eliminate the absolute reading difference caused by the total sample volume and sequencing depth; add 1 and take the logarithm with base 2 ( ), compresses the numerical range of highly expressed genes through logarithmic transformation (avoiding excessive influence of highly expressed genes on subsequent analysis), while retaining subtle differences in lowly expressed genes (ensuring that expression changes of low-abundance genes can be captured). Through the coordinated collection and standardized processing of clinical parameter units, imaging detection units, pathological analysis units and gene sequencing units, the multimodal data acquisition module can comprehensively obtain multi-dimensional information reflecting the patient's physiological status, tumor characteristics and molecular levels, providing a high-quality data foundation for subsequent cross-modal fusion and risk assessment.

[0108] More specifically, the standardization processing of the data preprocessing module includes:

[0109] The physiological signals collected by the clinical parameter unit are normalized using Z-score:

[0110] ;

[0111] in, is the original physiological parameter value (such as a heart rate of 85 beats / min), is the mean value of this parameter in the clinical database (heart rate =70 beats / min, systolic blood pressure , blood oxygen saturation ), is the reference standard deviation (heart rate =5 beats / min, systolic blood pressure =10mmHg, blood oxygen saturation ); Calculate the normalized value in real time for each frame of original physiological signal (one data point every 5ms at 200Hz sampling rate); for example, the original heart rate of 85 beats / min is normalized to , mapped to the deviation value under the standard normal distribution, eliminating the dimensional difference between heart rate (beats / minute) and blood pressure (mmHg), and achieving direct comparison across parameters; by converting the data to the standard normal distribution, eliminating the dimensional difference between different physiological parameters, and achieving comparability across parameters;

[0112] Normalize the grayscale of the image data of the image detection unit:

[0113] ;

[0114] in, is the original pixel value of the image (CT value range is -1024~4096HU, MRI signal intensity range is 0~65535), They are the maximum and minimum values ​​of pixels in a single image (such as , ); Map the image grayscale value to the range of 0-255 to unify the image brightness benchmark for different devices and different shooting conditions;

[0115] DICOM format image data is processed independently on a single slice basis; for example: original pixel values The CT pixels are normalized to , mapped to the grayscale range of 0-255, unifying the brightness deviation caused by different devices (such as different brands of CT) and different scanning conditions (such as tube voltage differences).

[0116] The feature screening steps specifically include: Step 2.1: Pearson correlation coefficient filtering: Calculate the Pearson correlation coefficient between each feature and surgical risk outcome (such as whether postoperative complications occur or not) ;reserve characteristics (such as tumor diameter and bleeding risk) =0.42, retaining this feature; while the patient's age and the risk of anastomotic fistula =0.21, and are removed. Step 2.2: Recursive Feature Elimination (RFE) optimization: Using the random forest model as the base classifier, set up 100 decision trees and calculate the out-of-bag error (OOBerror) of each feature; remove the 5% features with the highest OOB error (i.e., the features with the lowest contribution to risk prediction) at each iteration, and repeat the iteration until the number of remaining features is 50% of the initial value. Step 2.3: L1 Regularization (Lasso) Compression: Construct a Lasso regression model for the remaining features and set the regularization parameter Features with non-zero regression coefficients (such as tumor infiltration depth, KRAS gene mutation status, and other features with significantly non-zero coefficients) are retained, and the final feature dimension is compressed to 30%-40% of the initial value to reduce the interference of redundant information on subsequent fusion models.

[0117] More specifically, Figure 3 As shown, the cross-modal fusion module adopts an attention mechanism fusion network, including: feature extraction layer, attention allocation layer and fusion layer;

[0118] The feature extraction layer uses an adapted network structure to extract high-dimensional feature vectors from pre-processed data of four categories: clinical, imaging, pathology, and genetics:

[0119] Clinical data feature extraction: using a 3-layer fully connected network to extract feature vectors (dimension 512);

[0120] Network structure: input layer (number of nodes = number of clinical features after preprocessing, such as 20 dimensions) → hidden layer 1 (256 nodes, ReLU activation) → hidden layer 2 (512 nodes, ReLU activation) → output layer (512 nodes, no activation); map standardized physiological features such as heart rate and blood pressure into 512-dimensional abstract feature vectors , retaining key information of clinical status;

[0121] Image data feature extraction: Use pre-trained ResNet50 network to extract feature vectors (dimension 1024).

[0122] Network configuration: Remove the last three fully connected layers of ResNet50, retaining the fourth to last layer (output dimension 2048), and add a new fully connected layer (1024 nodes, ReLU activation).

[0123] Fine-tuning strategy: 5000 rectal cancer imaging data were used to fine-tune the network parameters (learning rate 0.001, ), so that the network focuses on image features such as tumor morphology and infiltration range, and outputs a 1024-dimensional vector .

[0124] Pathology data feature extraction: 2-layer CNN is used to extract feature vectors (dimension 512);

[0125] Network structure: Input layer (224×224×3 pathological slice image) → Convolutional layer 1 (32 3×3 convolution kernels, ReLU, maximum pooling) → Convolutional layer 2 (64 3×3 convolution kernels, ReLU, maximum pooling) → Fully connected layer (512 nodes, output vector ).

[0126] Gene data feature extraction: using a 2-layer fully connected network to extract feature vectors (dimension 512);

[0127] Network structure: input layer (number of nodes = number of normalized gene features, such as 23 dimensions) → hidden layer (256 nodes, ReLU) → output layer (512 nodes, output vector ).

[0128] The attention allocation layer calculates the weight coefficient of the clinical data modality:

[0129] ;

[0130] Similarly, calculate the weight coefficient of the image data modality , the weight coefficient of the pathological data modality , the weight coefficient of the genetic data modality , and satisfies ;

[0131] Features: Similarity of high-contribution modalities (such as pathological features of late-stage tumors) The larger the value, the higher the corresponding weight coefficient (e.g. can rise to 0.4-0.5);

[0132] in, is the cosine similarity function , used to measure the eigenvector and weight matrix degree of matching; It is a 32×512 dimensional learnable matrix, which is optimized through training to highlight the high contribution modes. is the feature vector obtained after feature extraction of clinical data, is the feature vector obtained after feature extraction of image data, is the feature vector obtained after feature extraction of pathological data, is the feature vector obtained after feature extraction of gene data;

[0133] The fusion layer outputs a comprehensive feature vector , the formula is:

[0134] ;

[0135] It is the feature vector (dimension 512) after feature extraction of clinical data, which contains abstract features of physiological signals such as heart rate and blood pressure; It is the feature vector (dimension 1024) extracted from the image data by ResNet50, which contains image features such as tumor morphology and location; It is the feature vector (dimension 512) after feature extraction of pathological data, including pathological features such as tumor differentiation degree; It is the feature vector (dimension 512) after feature extraction of gene data, including the expression characteristics of genes such as KRAS and BRAF.

[0136] The fusion network training steps include: Step 3.1: Initialize the network parameters and use Xavier to initialize the weight matrix W; Step 3.2: Train with the cross-entropy loss function; Step 3.3: Perform a modal importance assessment every 10 epochs and remove modal features whose contribution is continuously less than 5%; Step 3.4: When the validation set accuracy improves by less than 0.5% for 5 consecutive epochs, stop training and save the model parameters.

[0137] More specifically, Figure 5 As shown, the intelligent early warning module also includes: a risk tracing unit and a countermeasure generation unit;

[0138] The risk tracing unit identifies the main risk factors through SHAP value calculation:

[0139] ;

[0140] in, is a feature subset, Represents the total feature set Eliminate the After all the possible feature subsets, that is, Does not include Any combination of features, For feature subset The number of features contained in yes The weight coefficient of the value is used to balance the impact of feature subsets of different sizes on the results. is the total number of features, is the model output function; this formula quantifies the marginal contribution of a single feature to the risk assessment result. The top 3 features with the highest absolute values ​​are considered as the main risk factors. is the marginal contribution difference, indicating that Add the The amount of change in the model output after the feature is added.

[0141] Determination of major risk factors: Calculate the SHAP values ​​of all 1024 features and select the top three features with the highest absolute values ​​as major risk factors. For example, if the SHAP value of "tumor invasion of mesenteric vessels" is 15.2 (the largest contribution), the SHAP value of "intraoperative systolic blood pressure <90 mmHg" is 12.8, and the SHAP value of "BRAFV600E mutation" is 9.5, then these three features are marked as the current major risk factors.

[0142] The countermeasure generation unit matches the preset intervention plan library based on the main risk factors and outputs high-priority countermeasures. The steps are as follows:

[0143] Step 4.1. Semantic matching:

[0144] Preset Protocol Library: Contains over 100 clinically validated surgical interventions for rectal cancer, categorized by risk type (e.g., "bleeding risk," "tumor residual risk," "anastomotic leakage risk," etc.). Each protocol includes fields such as "applicable risk factors," "operational procedures," "historical success rate," "implementation ease," and "contraindications."

[0145] The matching method uses a double matching of "keywords and cosine similarity": extract keywords of major risk factors (such as "tumor invasion of blood vessels" to extract "vascular invasion" and "tumor infiltration"); calculate the cosine similarity between the keywords and the "applicable risk factors" field in the solution library (threshold ≥ 0.7) to screen out relevant solutions (such as "vascular invasion" matches "vascular ligation" and "laparoscopic vascular separation").

[0146] Step 4.2. Calculation of expected effect scores: Calculate the expected effect scores for the matching solutions: The success rate of this protocol in the past 500 cases with similar risks (e.g., success rate of "vascular ligation" = 92%); scored by three surgical experts (1-5 points, 1 point is the most complex, 5 points is the most convenient), and the average is taken (e.g., convenience of "laparoscopic vascular separation" = 4.2); For example, if the historical success rate of a protocol is 90% and the convenience is 4.0, then .

[0147] Step 4.3. Plan push and contraindication marking: Press Sort in descending order and push the three highest-scoring options to the surgeon's terminal (e.g., operating room display, doctor's tablet). Simultaneously mark the contraindications for each option: exclude inappropriate measures based on risk factor characteristics (e.g., if the primary risk factor is "massive intraoperative bleeding," mark "anticoagulant use" as a contraindication; if the risk factor is "intestinal obstruction," mark "excessive intestinal traction" as a contraindication).

[0148] Push content example:

[0149] Plan A: Laparoscopic vascular clipping , Contraindications: blood vessel diameter>5mm;

[0150] Plan B: Ultrasonic scalpel precise separation Contraindications: severe adhesions to adjacent organs; Plan C: conversion to laparotomy for hemostasis , Contraindications: Patients with hemodynamic instability.

[0151] More specifically, the risk contribution function Using the Sigmoid function:

[0152] ;

[0153] in, For the The contribution function of each feature to the surgical risk is Mapped to the [0,1] interval, it indicates the contribution ratio of the feature to the overall risk. is the slope parameter, which reflects the sensitivity of the feature to the risk, and its value ranges from 0.5 to 2. is the threshold parameter (feature critical value, such as the tumor diameter threshold of 5 cm); the Sigmoid function maps the feature value to the range of 0-1, quantifying the contribution of a single feature to the overall risk.

[0154] enter : No. The standardized value of each feature (such as tumor diameter 5 cm, KRAS gene expression level standardized value 1.2) comes from the comprehensive feature vector after cross-modal fusion .

[0155] Output : Feature contribution ratio, ranging from [0,1], indicating the contribution of this feature to the overall surgical risk (e.g. indicates that this characteristic contributes 80% of the risk weight).

[0156] Slope parameter : Controls the steepness of the function curve, reflecting the sensitivity of the feature to the risk; the value range is 0.5-2, among which: high sensitivity features (such as intraoperative bleeding volume) are taken (The curve is steep, and a small change in the characteristic value will lead to a significant fluctuation in the contribution ratio); low-sensitivity features (such as patient age) are taken (The curve is flat, and the change of eigenvalue has a mild effect on the contribution ratio).

[0157] Threshold parameter : Characteristic critical value, that is, the turning point of risk contribution ratio ( hour, =0.5); set according to clinical guidelines and historical data, for example:

[0158] Tumor diameter characteristics: (The contribution ratio is <0.5 when the diameter is ≤5cm and quickly rises to >0.5 when the diameter is >5cm);

[0159] KRAS gene mutation characteristics: (When the standardized expression level is ≤0.8, the contribution ratio is <0.5, and when it is >0.8, it indicates a significantly increased mutation risk).

[0160] Through the "S-shaped" curve of the Sigmoid function, the eigenvalue The nonlinear mapping to the [0,1] interval is consistent with the characteristics of medical risk: "slow rise at low thresholds, rapid transition near critical values, and saturation at high thresholds."

[0161] when (Eigenvalues ​​are well below the critical value): (The contribution is extremely low. For example, when the tumor diameter is 2 cm, the contribution to the bleeding risk is close to 0); when (Eigenvalues ​​are close to critical values): Rapidly jump from about 0.5 (for example, when the tumor diameter is close to 5 cm, the contribution ratio suddenly increases from 0.4 to 0.6); (Eigenvalues ​​are much higher than the critical value): (Contribution is saturated, for example, when the tumor diameter is 8 cm, the contribution to risk reaches the upper limit).

[0162] More specifically, the warning threshold setting and warning intensity calculation basis of the intelligent warning module are as follows:

[0163] When 3≤risk level<4 (corresponding to the transition from medium risk to medium-high risk, such as the risk value is in the range of 45-65), a level 1 warning is triggered, and the warning intensity is , the calculation formula is:

[0164] ;

[0165] When the risk level is ≥4 (corresponding to medium-high risk and above, such as risk value ≥65), the second-level warning is triggered, and the warning intensity is , the calculation formula is:

[0166] ;

[0167] The setting of warning intensity is based on the main risk factors identified by the risk tracing unit, and the risk factors are assigned The weight corresponding to the value is mapped to the range of 0-1 after weighted summation of the actual data value, which intuitively reflects the urgency of the risk;

[0168] The risk thresholds for levels 3 and 4 were determined by ROC curve analysis to make the sensitivity ≥ 90%, respectively. The highest risk value in history (95% percentile); the warning intensity is 0-1, which determines the frequency of the warning signal (such as the first-level warning prompts every 30 seconds, and the second-level warning continues with sound and light alarms).

[0169] The value range is [0.6, 0.9], near Linear increase, such as =45 hours =0.6, =65 hours =0.9, the corresponding alarm frequency gradually increases from once every 30 seconds to once every 10 seconds.

[0170] The value range is [0.9, 1.0], near Linear increase, such as =65 hours =0.9, =90 hours =1.0, the corresponding alarm frequency is continuous sound and light alarm (no interval).

[0171] Response steps after the warning is triggered: Step 5.1: When the first-level warning is triggered, the risk factor analysis report is automatically pushed to the surgeon's terminal; Step 5.2: When the second-level warning is triggered, the operating room sound and light alarm is triggered simultaneously, and the backup medical team dispatch mechanism is activated; Step 5.3: The warning is released when the risk assessment value is <3 for three consecutive times and falls back to the low-risk range.

[0172] More specifically, the system self-check module performs the following operations:

[0173] Data validity test, calculate the integrity coefficient of each modal data, the calculation formula is:

[0174] ;

[0175] in,

[0176] Clinical parameter modality: is the total data items of physiological signals related to heart rate and blood pressure (such as 120,000 data points for 5 minutes at 200 Hz sampling), Valid data items with no missing data or abnormal fluctuations (such as sudden rise or fall in blood pressure beyond the clinically reasonable range);

[0177] Image data modality: is the total number of slices in CT / MRI, The number of clear slices without artifacts and with a thickness that meets the requirement of 0.5-1 mm;

[0178] Pathology data modality: is the total field of view of the tumor section, The number of effective fields with a resolution ≥ 0.2 μm / pixel and no staining bias;

[0179] Genetic data modality: is the total detection items of 23 target genes, Valid gene items with sequencing quality value Q30 ≥ 80%;

[0180] , when data collection is not updated for more than 30 minutes, ; The integrity factor is used to quantify the data quality, The data is determined to be invalid;

[0181] The calibration process is as follows:

[0182] Step 6.1. Abnormal module positioning, calculate the four types of modal data (clinical, imaging, pathology, and gene) separately ,position The exception module:

[0183] Abnormal clinical parameter unit: such as missing rate of heart rate data > 20% ,Combined with the sensor connection status detection, mark “physiological signal sensor failure”;

[0184] Abnormal imaging detection unit: if the CT image slice thickness deviation is >0.2mm, it will be marked as "device parameter drift";

[0185] Pathology / gene unit abnormality: If the slice image resolution is <0.2μm / pixel or the sequencing data Q30 is <80%, it will be marked as "data acquisition quality does not meet the standard".

[0186] Step 6.2. Perform targeted calibration for abnormal modules:

[0187] Clinical parameter unit: Connect the multi-parameter monitor to a standard analog signal source (outputting known heart rate and blood pressure signals), and adjust the sensor gain so that the deviation between the measured value and the standard value is ≤2%;

[0188] Image detection unit: Perform water phantom calibration on CT equipment, adjust tube voltage and tube current parameters to ensure image noise value ≤ 5HU; perform uniform field calibration on MRI equipment to ensure signal uniformity ≥ 90%;

[0189] Pathology analysis unit: calibrate the focus and lighting intensity of the digital pathology scanner to ensure that the structural similarity between the scanned image and the standard slice is ≥95%;

[0190] Gene sequencing unit: Verify the sensitivity of the fluorescence channel of the sequencer by testing with standards (gene fragments of known concentration) to ensure that the detection error is ≤5%.

[0191] Step 6.3. Calibration Verification and Fault Alarm:

[0192] After calibration, collect 3 sets of verification data for the abnormal module (the amount of data for each set = 10% of the normal collection amount) and recalculate :If 3 sets of data , determine that the calibration is valid, the system resumes normal operation, and records the calibration time and parameter adjustment values.

[0193] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Rectal cancer surgery risk assessment and early warning system based on multimodal data fusion, characterized by: include: a multimodal data acquisition module configured to collect clinical data, imaging data, pathological data, and genetic data of a patient through a multi-source detection device group; A data preprocessing module, connected to the multimodal data acquisition module, performs noise reduction, standardization and feature screening on the collected multimodal data; The cross-modal fusion module receives multi-source data processed by the data preprocessing module, performs feature fusion through the deep learning network, and outputs a comprehensive feature vector; A dynamic risk assessment module, which constructs a real-time assessment model based on the comprehensive feature vector and outputs a surgical risk value and risk level; An intelligent early warning module, communicating with the dynamic risk assessment module, triggering a multi-level early warning mechanism based on risk values; System self-check module, which performs real-time monitoring and calibration of the operating status and data validity of the multimodal data acquisition module, data preprocessing module, cross-modal fusion module, dynamic risk assessment module, and intelligent early warning module; The dynamic risk assessment module includes: Preoperative assessment unit outputs basic risk value based on multimodal data within 72 hours before surgery , the calculation formula is: ； in, The feature weight is calculated by random forest Gini coefficient. The data used in the calculation of random forest Gini coefficient are derived from the physiological signals related to heart rate and blood pressure collected by the clinical parameter unit, the three-dimensional images of tumors obtained by the imaging detection unit, the tumor slice information collected by the pathological analysis unit, and the expression data of rectal cancer-related genes detected by the gene sequencing unit; and , is the risk contribution function; Generate an initial risk benchmark by integrating the risk contributions of multi-dimensional features through weighted summation; The intraoperative real-time unit updates the risk value every 5 minutes. The calculation formula is: ； in, is the attenuation coefficient, ranging from 0.01 to 0.05, which increases dynamically with the progress of the operation. The duration of the operation, is the real-time data weight, with an initial value of 0.3 and an increase of 0.1 every 30 minutes. The incremental risk of newly added data; through exponential decay, the influence of preoperative data is weakened, the weight of real-time intraoperative data is strengthened, and dynamic risk iteration is achieved; Risk level classification unit, the classification steps include: Step 1.1: Collect risk values ​​from 1,000 historical surgical data and generate five risk intervals using K-means clustering. Step 1.2: Based on clinical expert opinions, cluster centers were mapped to risk levels of 1-5, with level 1 being the lowest and level 5 being the highest. Step 1.3: When the risk values ​​of two consecutive updates change across levels, level calibration is triggered and the adjacent level thresholds are recalculated; The standardization processing of the data preprocessing module includes: The physiological signals collected by the clinical parameter unit are normalized using Z-score: ； in, is the original physiological parameter value, is the mean value of this parameter in the clinical database, is the reference standard deviation; by converting the data into a standard normal distribution, the dimensional differences of different physiological parameters are eliminated, and cross-parameter comparability is achieved; Normalize the grayscale of the image data of the image detection unit: ； in, is the original pixel value of the image, The maximum and minimum values ​​of pixels in a single image, respectively; the image grayscale value is mapped to the range of 0-255 to unify the image brightness benchmark for different devices and different shooting conditions; The feature screening steps specifically include: Step 2.1: Calculate the Pearson correlation coefficient between each characteristic and surgical risk ,reserve characteristics; Step 2.2: Use recursive feature elimination to iteratively remove the 5% features with the lowest contribution, using the out-of-bag error of the random forest model as the indicator; Step 2.3: Further compress the feature dimension through L1 regularization and retain the feature subset with non-zero coefficients.

2. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 1 is characterized in that: The multimodal data acquisition module includes: The clinical parameter unit uses a multi-parameter monitor to synchronously collect physiological signals related to heart rate, blood pressure, and blood oxygen saturation, with a sampling frequency set to 200 Hz; The imaging detection unit integrates 64-slice spiral CT and 3.0T MRI equipment to obtain three-dimensional tumor imaging data with a slice thickness controlled at 0.5-1mm; Pathology analysis unit, which uses a digital pathology scanner to collect tumor tissue slice images with a resolution of no less than 0.2 μm / pixel; The gene sequencing unit uses second-generation sequencing technology to detect the expression levels of 23 colorectal cancer-related genes related to KRAS and BRAF.

3. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 2, characterized in that: The cross-modal fusion module adopts an attention mechanism to fuse the network, including: Feature extraction layer, which uses a fully connected network to extract feature vectors from clinical data , ResNet50 is used to extract feature vectors from image data ; Attention allocation layer, calculates the weight coefficient of clinical data modality: ； Similarly, calculate the weight coefficient of the image data modality , the weight coefficient of the pathological data modality , the weight coefficient of the genetic data modality , and satisfies ; in, is the cosine similarity function , used to measure the eigenvector and weight matrix degree of matching; It is a 32×512 dimensional learnable matrix, which is optimized through training to highlight the high contribution modes. is the feature vector obtained after feature extraction of clinical data, is the feature vector obtained after feature extraction of image data, is the feature vector obtained after feature extraction of pathological data, is the feature vector obtained after feature extraction of gene data; Fusion layer, output comprehensive feature vector , the formula is: ； The fusion network training steps include: Step 3.1: Initialize the network parameters and use Xavier to initialize the weight matrix W; Step 3.2: Train with cross entropy loss function; Step 3.3: Perform a modal importance assessment every 10 epochs and remove modal features whose contribution is consistently less than 5%; Step 3.4: When the validation set accuracy improves by less than 0.5% for five consecutive epochs, stop training and save the model parameters.

4. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 3 is characterized in that: The intelligent early warning module also includes: The risk tracing unit identifies the main risk factors through SHAP value calculation: ； in, is a feature subset, Represents the total feature set Eliminate the After all the possible feature subsets, that is, Does not include Any combination of features, is a feature subset The number of features contained in The weight coefficient of the value is used to balance the impact of feature subsets of different sizes on the results. is the total number of features, is the model output function; this formula quantifies the marginal contribution of a single feature to the risk assessment result. The top 3 features with the highest absolute values ​​are considered as the main risk factors. is the marginal contribution difference, indicating that Add the The change in the model output after the feature is added; Countermeasure generation unit, intervention plan matching steps: Step 4.1: Semantically match the main risk factors with the preset solution library; Step 4.2: Calculate the expected effect score of the program: ； Step 4.3: Push the three options with the highest scores and mark the contraindications for each option.

5. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 4 is characterized in that: The warning threshold setting and warning intensity calculation basis of the intelligent warning module are as follows: When 3≤risk level<4, a level 1 warning is triggered. , the calculation formula is: ； When the risk level is ≥4, the second level warning is triggered, and the warning intensity is , the calculation formula is: ； The setting of warning intensity is based on the main risk factors identified by the risk tracing unit, and the risk factors are assigned The weight corresponding to the value is mapped to the range of 0-1 after weighted summation of the actual data value, which intuitively reflects the urgency of the risk; The risk thresholds are level 3 and level 4 respectively. The highest risk value in history; the warning intensity ranges from 0 to 1, which determines the frequency of warning signals; Response steps after the warning is triggered: Step 5.1: When a level 1 warning is issued, the risk factor analysis report is automatically sent to the surgeon’s terminal; Step 5.2: When the second-level warning is triggered, the operating room sound and light alarms are triggered simultaneously, and the backup medical team dispatch mechanism is activated; Step 5.3: Conditions for lifting the warning: the risk assessment value is <3 for three consecutive times and falls back to the low-risk range.

6. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 5, characterized in that: The system self-check module performs the following operations: Data validity test, calculate the integrity coefficient of each modal data, the calculation formula is: ； Among them, the integrity coefficient is calculated for the four modalities of clinical, imaging, pathology, and genetics respectively: Clinical parameter modality: is the total data item of physiological signals related to heart rate and blood pressure, Valid data items without missing data or abnormal fluctuations; Image data modality: is the total number of slices in CT / MRI, The number of clear slices without artifacts and with a thickness that meets the requirement of 0.5-1 mm; Pathology data modality: is the total field of view of the tumor section, The number of effective fields with a resolution ≥ 0.2 μm / pixel and no staining bias; Genetic data modality: is the total detection items of 23 target genes, Valid gene items with sequencing quality value Q30 ≥ 80%; is the time decay factor. When data collection is not updated for more than 30 minutes, ; The integrity factor is used to quantify the data quality, The data is determined to be invalid; Calibration process: Step 6.1: When Automatically locate abnormal modules; Step 6.2: Perform baseline calibration on the abnormal module; Step 6.3: Collect 3 sets of verification data after calibration. If all of them meet the requirements , then normal operation is restored; otherwise, a hardware failure alarm is triggered.

7. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 6, characterized in that: The gene expression level of the gene sequencing unit is normalized by: ； in, is the raw sequencing read, is the average reading of the internal reference gene; logarithmic transformation is used to compress the numerical range of highly expressed genes while retaining the differential information of lowly expressed genes; The basis of this standardization method is to measure the gene expression of standard cell lines under three different experimental conditions, construct the gene correlation function, and select the stable interval with the function value fluctuation ≤5% as the benchmark; Average read counts compared to the reference gene Substituting into the formula, through The conversion compresses the numerical range of highly expressed genes while retaining the differential information of lowly expressed genes, making the gene expression levels of different samples comparable across experimental conditions and meeting the needs of collaborative analysis of genetic data and clinical and imaging-related data in multimodal fusion.

8. The rectal cancer surgery risk assessment and early warning system based on multimodal data fusion according to claim 7, characterized in that: The risk contribution function Using Sigmoid function: ； in, For the The contribution function of each feature to the surgical risk is Mapped to the [0,1] interval, it indicates the contribution ratio of the feature to the overall risk. is the slope parameter, which reflects the sensitivity of the feature to the risk, and its value ranges from 0.5 to 2. is the threshold parameter; the Sigmoid function maps the eigenvalue to the range of 0-1, quantifying the contribution of a single feature to the overall risk.

Images