An intraoperative anomaly recognition method and device based on multi-modal data

Abstract

The application provides an intraoperative anomaly recognition method and device based on multi-modal data, comprising: acquiring real-time intraoperative multi-dimensional state information, and dynamically adjusting monitoring parameters of a current operation stage based on the intraoperative multi-dimensional state information, the monitoring parameters comprising an actual monitoring index set of the current operation stage, actual sampling frequencies of each monitoring index, and actual anomaly thresholds of each monitoring index; collecting monitoring data based on the dynamically adjusted monitoring parameters, and performing multi-level anomaly detection based on the monitoring data; in the case of determining that there is an intraoperative anomaly, outputting a graded alarm signal based on an anomaly type, a severity level, and a current operation context, which can realize intelligent allocation of monitoring resources, greatly reduce system load and artificial intervention demand on the premise of ensuring safety, significantly improve the accuracy and timeliness of anomaly recognition, and reduce the risk of missed reports and false reports.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for intraoperative anomaly identification based on multimodal data. Background Technology

[0002] Neurosurgical tumor resection is a high-risk procedure; intraoperative complications can cause serious harm or even endanger the patient's life. Currently, clinical intraoperative monitoring relies mainly on the surgeon's experience and limited monitoring equipment, which presents the following technical challenges:

[0003] (1) The monitoring plan lacks personalization: Current intraoperative monitoring systems use a uniform monitoring protocol and cannot be adjusted according to individual patient differences. Intraoperative risks vary significantly among different patients, but the monitoring system fails to reflect these differences.

[0004] (2) The monitoring strategy lacks dynamic adaptability: The risk level changes continuously during surgery, but the monitoring frequency and range of the existing monitoring system remain fixed. Visual monitoring fails when the field of vision is limited, but the system cannot automatically switch to non-visual monitoring mode. This static monitoring strategy results in insufficient monitoring during high-risk periods and excessive monitoring during low-risk periods.

[0005] (3) Insufficient accuracy in anomaly identification: Existing anomaly identification methods are mostly based on single-parameter threshold alarms, lacking multi-parameter correlation analysis and clinical contextual interpretation.

[0006] (4) The contradiction between monitoring burden and accuracy: To ensure accurate anomaly identification, it is often necessary to increase monitoring parameters and monitoring frequency, but this increases the cognitive burden on doctors and the processing load on the system. Doctors need to focus on multiple monitoring interfaces simultaneously, which can easily lead to distraction in complex surgical environments. Processing large amounts of monitoring data can also cause response delays, affecting real-time performance.

[0007] (5) Data latency affects real-time decision-making: Intraoperative key data inevitably experience delays, such as the processing time required for intraoperative images (ultrasound, MRI) and the sampling and analysis time required for laboratory tests (blood gas, coagulation). Existing systems can only wait during these data delays and cannot provide timely monitoring support, leading to delays in anomaly identification.

[0008] In summary, existing intraoperative static monitoring methods cannot simultaneously ensure the accuracy and timeliness of intraoperative abnormality identification under certain monitoring burdens. Summary of the Invention

[0009] This invention provides a method and apparatus for intraoperative anomaly identification based on multimodal data, which solves the problem that existing technologies cannot simultaneously ensure the accuracy and timeliness of intraoperative anomaly identification under certain monitoring burdens.

[0010] On one hand, the present invention provides a method for intraoperative anomaly identification based on multimodal data, used for intraoperative anomaly identification in neurosurgical tumor resection, including: Real-time acquisition of intraoperative multidimensional status information, and dynamic adjustment of monitoring parameters for the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes current surgical stage identification information, real-time physiological parameters of the patient, physician focus assessment information, and surgical field quality assessment information. Monitoring data is collected based on dynamically adjusted monitoring parameters, and multi-level anomaly detection is performed based on the monitoring data; If an intraoperative abnormality is confirmed, a graded alarm signal is output based on the type of abnormality, severity level, and current surgical context.

[0011] In one optional embodiment of this application, the method further includes, before the surgery begins: Based on patient individual characteristic data, tumor characteristic data, and physician skill data, a personalized monitoring blueprint is generated for different surgical stages. The monitoring blueprint includes the initial set of monitoring indicators for each surgical stage, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold for each monitoring indicator.

[0012] In one optional embodiment of this application, generating personalized monitoring blueprints for different surgical stages based on patient individual characteristic data, tumor characteristic data, and physician skill data specifically includes: Based on a pre-constructed multi-dimensional risk assessment model, the risk assessment result of the current surgery is determined. The input parameters of the multi-dimensional risk assessment model include patient baseline parameters, tumor parameters, and surgeon parameters. The patient baseline parameters include age, severity score of underlying disease, and coagulation function indicators. The tumor parameters include location risk classification, size, blood supply richness, and distance from important functional areas. The surgeon parameters include surgical proficiency index and recent complication rate. Based on the risk assessment results of the current surgery, allocate differentiated monitoring resource budgets for different stages of the current surgery; Based on the monitoring resource budget and inherent risks of different surgical stages, the initial set of monitoring indicators, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold of each monitoring indicator are determined for each surgical stage. The initial set of monitoring indicators for each surgical stage includes a subset of core monitoring indicators and a subset of auxiliary monitoring indicators, and the initial sampling frequency of each monitoring indicator is positively correlated with the inherent risk of the corresponding surgical stage.

[0013] In one optional embodiment of this application, the step of dynamically adjusting the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information specifically includes: The current risk level and the doctor's current level of focus are determined based on the intraoperative multidimensional status information. Based on the current risk level and the doctor's current focus at the current surgical stage, at least one of the following should be adjusted: the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator.

[0014] In one optional embodiment of this application, the method further includes, during the real-time acquisition of intraoperative multidimensional status information: Determine the target monitoring indicator whose test results are delayed beyond a preset tolerance time; The historical trend data of the target monitoring indicator, the associated monitoring data of the target monitoring indicator, the current surgical stage information, and the current operation type information are input into a pre-established short-term prediction model of the monitoring indicator to obtain the current estimated value of the target monitoring indicator. Based on the confidence level of the current estimated value of the target monitoring indicator, determine whether to trigger a pre-alarm.

[0015] In one optional embodiment of this application, the method further includes: The actual detection results of the target monitoring indicator are compared with the corresponding predicted values, and the short-term prediction model of the monitoring indicator is calibrated based on the comparison results.

[0016] In one optional embodiment of this application, the step of performing multi-level anomaly detection based on the monitoring data specifically includes: Based on the monitoring data, single-indicator real-time threshold comparison, multi-indicator correlation analysis, time window-based trend analysis, and machine learning model-based predictive detection are performed to determine the confidence score of intraoperative abnormalities. Based on the intraoperative abnormality confidence score, it is determined whether an intraoperative abnormality exists.

[0017] In one optional embodiment of this application, the formula for calculating the intraoperative abnormality confidence score is as follows: Score = α×S1+ β×S2+ γ×S3+ δ×S4; Wherein: S1 is the normalized score for the degree of anomaly of a single indicator, S2 is the consistency score for the correlation of multiple indicators, S3 is the score for the persistence of trends, and S4 is the confidence score for the prediction model; α, β, γ, and δ are dynamic weight coefficients that are adjusted according to the current surgical stage and risk level.

[0018] In one optional embodiment of this application, the step of outputting a graded alarm signal based on the abnormality type, severity level, and current surgical context specifically includes: Based on the type and severity of the abnormality, combined with the current stage of surgery, the patient's real-time condition, the doctor's focus, and the quality of the surgical field of vision, a comprehensive urgency score is calculated. The alarm level is determined based on the comprehensive urgency score, and the alarm sensory channel and information presentation density are adaptively selected based on the alarm level, the current surgical stage, and the doctor's focus, and a corresponding graded alarm signal is generated and output.

[0019] Secondly, the present invention also provides an intraoperative anomaly identification device based on multimodal data for intraoperative anomaly identification in neurosurgical tumor resection, comprising: The monitoring parameter dynamic adjustment module is used to acquire intraoperative multidimensional status information in real time and dynamically adjust the monitoring parameters of the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes the current surgical stage identification information, the patient's real-time physiological parameters, the doctor's focus assessment information, and the surgical field quality assessment information. An anomaly detection module is used to collect monitoring data based on dynamically adjusted monitoring parameters and to perform multi-level anomaly detection based on the monitoring data. The alarm module is used to output a graded alarm signal based on the type of abnormality, severity level, and current surgical context when an intraoperative abnormality is determined to exist.

[0020] This invention provides a method and apparatus for intraoperative anomaly identification based on multimodal data, used for intraoperative anomaly identification in neurosurgical tumor resection. It acquires multidimensional intraoperative status information in real time and dynamically adjusts monitoring parameters for the current surgical stage based on this information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each indicator, and the actual anomaly threshold for each indicator. The multidimensional status information includes current surgical stage identification information, real-time patient physiological parameters, physician focus assessment information, and surgical field quality assessment information. Monitoring data is collected based on the dynamically adjusted monitoring parameters, and multi-level anomaly detection is performed based on this data. When an intraoperative anomaly is determined, a graded alarm signal is output based on the anomaly type, severity level, and current surgical context. This enables intelligent allocation of monitoring resources, significantly reducing system load and manual intervention requirements while ensuring safety, and significantly improving the accuracy and timeliness of anomaly identification, reducing the risk of missed and false alarms. Attached Figure Description

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

[0022] Figure 1 A flowchart illustrating an intraoperative anomaly identification method based on multimodal data provided by the present invention; Figure 2 A schematic diagram illustrating the process of generating a personalized monitoring blueprint provided by this invention; Figure 3 A schematic diagram of the processing flow for detection result delay provided by the present invention; Figure 4 A structural block diagram of an intraoperative anomaly identification device based on multimodal data provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The specific implementation process of the intraoperative anomaly identification method and device based on multimodal data in the embodiments of the present invention will be described below.

[0025] Figure 1 This is a flowchart illustrating an intraoperative anomaly identification method based on multimodal data provided in an embodiment of the present invention, used for intraoperative anomaly identification in neurosurgical tumor resection, such as... Figure 1 As shown, the method may include: Step 101: Acquire intraoperative multidimensional status information in real time, and dynamically adjust the monitoring parameters of the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes the current surgical stage identification information, the patient's real-time physiological parameters, the doctor's focus assessment information, and the surgical field quality assessment information.

[0026] Specifically, based on the background technology, existing intraoperative monitoring solutions suffer from two core pain points: "lack of dynamic adaptability in monitoring strategies" and "the contradiction between monitoring burden and accuracy." To address these issues, this invention defines a closed loop from "perception" to "decision-making," aiming to transform the monitoring system from a static, passive data recorder into a dynamic, proactive intelligent sensing entity. More specifically, this invention, through research, has discovered that neurosurgical safety is a complex result of the dynamic interaction of four main entities: patient, physician, surgical process, and surgical environment. Based on this, the multidimensional state information in this invention provides real-time state monitoring of these four entities: The "current surgical stage identification information" defines the spatiotemporal coordinates of risk, providing a basic timeline and risk context for all monitoring, analysis, and decision-making. Surgical risks are not uniformly distributed but are strongly correlated with the surgical procedure. For example, the main risk in the craniotomy stage is massive bleeding due to damage to major blood vessels; the monitoring focus should be on sudden drops in blood pressure and compensatory increases in heart rate. In the tumor resection stage, the main risks are damage to important neurological functional areas and brain tissue traction / swelling; the monitoring focus should shift to neurophysiological signals (MEP / SEP) and intracranial pressure trends. Without "current surgical stage identification information," the system will lose its ability to judge macroscopic risk trends, cannot proactively allocate resources, and can only respond passively. Based on this, this embodiment of the invention divides the surgical stages into incision stage, craniotomy stage, dural opening stage, tumor exposure stage, lesion resection stage, hemostasis stage, and closure stage based on differences in surgical risk.

[0027] The incision stage is the initial step of the surgery. In this stage, a scalp incision is designed based on the tumor location. Using a scalpel or high-frequency electrocautery, the skin, subcutaneous tissue, and galea aponeurotica are cut layer by layer to form a skin flap, which is then flipped downwards to expose the skull. At this stage, the cranial cavity has not yet been entered, and the risk is relatively low, mainly due to hemodynamic fluctuations after anesthesia induction and stress responses to incision stimulation (transient increases in heart rate and blood pressure). Bleeding is mostly peripheral oozing and is relatively easy to control. The craniotomy stage involves drilling holes in the exposed skull using a cranial drill, then connecting the holes with a milling cutter to form a free bone flap. Finally, the bone flap is removed, exposing the underlying dura mater. The risk is significantly higher at this stage, mainly due to instrument slippage or excessive depth leading to accidental damage to the dura mater, venous sinuses, or brain tissue, causing severe bleeding. Bone fragments and thermal damage may also pose risks. The dura mater opening stage involves making a small, precise incision in the dura mater, then widening the incision along a predetermined route using microscissors or a blade, and flipping the dural flap to one side, ultimately fully exposing the arachnoid membrane on the surface of the brain tissue. The risks at this stage include accidental damage from delicate manipulation, potentially damaging draining veins on the cortical surface, leading to bleeding, or directly damaging the cerebral cortex. The tumor exposure stage primarily utilizes the brain's natural sulci or selects non-critical functional areas, using brain retractors, microsurgical suction devices, and bipolar electrocoagulation to gently separate brain tissue, gradually creating a surgical pathway to the tumor until its boundaries are clearly visible. The risks at this stage include traction and compression of normal brain tissue, potentially causing ischemia or neurological damage, and possibly damaging small blood vessels along the pathway. The lesion resection stage is the core and highest-risk stage of the surgery. This stage primarily uses instruments such as ultrasonic suction devices and tumor forceps to remove the tumor tissue in sections or en bloc, while simultaneously using bipolar electrocoagulation to treat tumor-feeding vessels and bleeding points. The risks peak at this stage, including: active massive bleeding, damage to important blood vessels, nerve conduction tracts, or functional areas surrounding the tumor, acute cerebral edema, and severe hemodynamic fluctuations. The hemostasis stage primarily involves systematic and thorough hemostasis of the surgical cavity after tumor resection, using bipolar electrocoagulation, hemostatic materials, irrigation, and other methods to treat all visible bleeding points until the surgical field is dry. The risk shifts from acute massive bleeding to occult, slow oozing. Incomplete hemostasis is a major cause of postoperative hematoma formation and subsequent reoperation. The closure stage is the final stage of the surgery, involving tight suturing of the dura mater, repositioning the bone flap and fixing it with connecting pieces, and finally, layered suturing of the scalp tissue. This stage carries the lowest overall risk, but there is still a risk of acute intracranial pressure increase (e.g., due to overly tight dural sutures or cerebral edema) or unexplained hypotension during closure.

[0028] "Real-time physiological parameters of the patient" are the core object and ultimate goal of monitoring, serving as direct and objective physiological indicators reflecting the patient's vital signs. They are the ultimate target of all technological services and the most fundamental basis for judging abnormalities. No matter how intelligent the system is, its final output is based on the analysis of physiological parameters such as heart rate, blood pressure, blood oxygen, and cerebral oxygen. These parameters are the "target variables" for adjusting other dimensions of information. For example, adjusting the monitoring frequency (based on the surgical stage and the doctor's condition) is precisely to more accurately capture changes in these parameters.

[0029] "Physician focus assessment information" is a brain-computer interface used to optimize human-computer interaction, assessing the cognitive load and processing capacity of the surgical subject. Alarm effectiveness = alarm accuracy × alarm reachability. Even the most accurate alarm is ineffective if it is ignored due to the doctor's intense focus on the delicate procedure or if the response is delayed due to fatigue. For example, if the doctor is suturing a blood vessel under a microscope, the system should use a non-invasive alarm (such as a flashing red halo at the edge of the microscope's field of view) to avoid startling the doctor and causing hand tremors.

[0030] "Surgical field of view quality assessment information" is used to evaluate the availability and quality of the primary information acquisition channels (vision) during surgery. Because neurosurgery is a procedure that relies heavily on vision, when the core visual channels are disrupted, not only do surgeons face difficulties in making decisions, but computer vision-based automated monitoring (such as bleeding detection) also fails.

[0031] The information from these four dimensions collectively constitutes a minimal complete set, capable of comprehensively and in real-time characterizing the risks, subjects, operatives, and environment during neurosurgery. This design transforms the intraoperative anomaly identification device based on multimodal data of this invention from a passive "data monitor" into an "intelligent surgical partner" capable of understanding the surgical context, anticipating changes in risk, and collaborating with the surgeon. This is the fundamental technological foundation that distinguishes this invention from all existing technologies, enabling it to simultaneously achieve the triple goals of "reducing workload, improving accuracy, and ensuring timeliness."

[0032] The specific methods for obtaining intraoperative multidimensional status information are as follows: The acquisition of identification information for the current surgical stage includes the following steps: 1. The intraoperative anomaly identification device receives data streams from multiple devices in the operating room in parallel, forming the input features of the identification model. Input features include: surgical microscope video stream, providing the most direct visual scene; instrument status bus signals, which acquire the real-time status of key instruments (including the activation status, output power / energy value of high-frequency electrosurgical unit / bipolar electrocoagulation device; the activation status, amplitude level, and suction pressure of ultrasonic aspirator; the rotation speed and working status of cranial drill / milling cutter; and the flow status of the washing / suction system) through an operating room equipment integration network (such as ORiN, IEEE 11073 SDC protocol) or direct hardware interface; coordinates from the neuronavigation system, acquiring the real-time position of the surgical instrument tip in the preoperative three-dimensional image space, and its relative distance to the preset tumor and key structures (such as the motor cortex and major blood vessels) models; and surgical timing and preset procedures, acquiring the surgical time elapsed from the surgical anesthesia information management system, as well as a rough timeline of stages based on the preoperative plan.

[0033] 2. Extract representative features from the raw data to distinguish different surgical stages: Obtain tissue color and texture histograms from the video (to differentiate skin, skull, dura mater, brain tissue, tumor, and hemorrhage), instrument visibility (detecting whether instruments of specific shapes, such as milling cutter heads and microscissors, appear in the field of view and their proportion), and surgical field depth and field of view (estimated using binocular microscope parallax or focal information); statistically analyze the cumulative activation duration, total energy release, and start / stop frequency of each instrument within a specific time window (e.g., 30 seconds) from the instrument signals. Calculate the average distance, minimum distance, and rate of change between the instrument and the tumor boundary from the navigation data.

[0034] 3. Identification Model and Inference: A lightweight temporal classification model is employed, using the previously extracted multimodal feature sequences as input. Specifically, multimodal features are extracted in real-time within a sliding time window (e.g., 10-15 seconds). The feature vectors are input into the pre-trained temporal classification model, which outputs the probability distribution of the current window belonging to each preset surgical stage (e.g., outputting a discrete "stage label" and its corresponding confidence score). The Viterbi algorithm or a simple sliding window voting method can be used to smooth the output of continuous windows, avoiding frequent jumps, and finally determining the most likely surgical stage label.

[0035] Temporal classification models can choose Hidden Markov Models (HMMs) or Bidirectional Long Short-Term Memory (LSTM) networks that can model temporal dependencies. The training process can utilize a large amount of historical surgical multimodal data (with stage labels manually annotated by experts) for supervised learning, enabling the model to learn multimodal feature patterns from different stages.

[0036] The acquisition of real-time physiological parameters of patients includes the following steps: 1. Obtain data such as electrocardiogram, heart rate, invasive / non-invasive arterial blood pressure, blood oxygen saturation, end-tidal carbon dioxide, respiratory rate, and body temperature from the anesthesia workstation / monitor; 2. Obtain information such as brain oxygen saturation, invasive intracranial pressure, and neurophysiological monitoring (MEP / SEP) from specialized monitoring equipment (accessed via a dedicated interface or analog acquisition card).

[0037] The above data is acquired using standard medical device data communication protocols, such as HL7 (for numerical and alarm information) or IEEE 11073 (for waveform and continuous data streams). For older equipment, serial communication or data acquisition via its analog output interface in conjunction with a data acquisition card can be used. After acquisition, operations such as timestamp synchronization, filtering and noise reduction, outlier removal (using simple rules or algorithms to remove transient outliers caused by motion artifacts or connection interruptions), resampling, and alignment can be performed. Finally, a set of physiological parameter time series with uniform timestamps and preliminary cleaning is output, including waveform and numerical data.

[0038] Obtaining physician focus assessment information includes the following steps: 1. Acquire multimodal sensing data, including eye-tracking data (acquired via a miniature eye tracker integrated into the surgical microscope eyepiece or a lightweight eye-tracking glasses worn by the surgeon; raw data includes fixation point coordinates, pupil diameter, and blink frequency), and surgical instrument kinematic data (acquired via a miniature inertial measurement unit integrated into the handle of the main surgical instruments or an existing optical / electromagnetic surgical instrument tracking system in the operating room; including the instrument's three-dimensional spatial position, velocity, acceleration, and angular velocity). It may also include physiological data acquired from the surgeon's wearable devices, such as the low-frequency / high-frequency component ratio of heart rate variability and skin conductance activity.

[0039] 2. Attention Scale Calculation: Indicators reflecting attention scale are extracted from raw sensor data, including fixation stability (calculating the percentage of time a continuous fixation point spends within a key surgical area and the average fixation duration; at high attention scales, the fixation point remains stable in the key area), saccade patterns (calculating saccade rate and average saccade amplitude; frequent, large-scale disordered saccades may indicate distraction or search difficulties), and pupil diameter changes (pupil dilation is associated with increased cognitive load) from instrument motion data; and motion smoothness (calculating the "jerk" of the motion trajectory, i.e., the derivative of acceleration; a smaller value indicates smoother, more automated, and more skilled motion, indirectly reflecting efficient operation under high attention scales) and standstill / hesitation time (identifying pauses and their frequency where motion speed is below a threshold; unnecessary or excessively long pauses may indicate decision-making difficulties or distraction) from eye-tracking data.

[0040] 3. Attention Score Fusion: Multiple features mentioned above (such as gaze stability and motion fluency) are input into a regression model (e.g., support vector regression, random forest regression). This model has been trained in a laboratory or simulated surgical environment by having doctors perform standard tasks while simultaneously using subjective scales or task performance as "true" attention labels. The model output is a continuous value between 0 and 1, representing a real-time attention score. A value close to 1 indicates high focus, while a value close to 0 indicates severe distraction or fatigue.

[0041] Obtaining information on the quality of the surgical field includes the following steps: 1. Quality assessment algorithm: For each frame or every few frames in the surgical microscope's main camera video stream, the following analysis is performed in parallel: A. Assessment of Bleeding and Bloodstains: Color Space Conversion: Convert the image from RGB to HSV color space; Threshold Segmentation: Set threshold ranges for H (hue), S (saturation), and V (brightness) to extract all pixel regions in the image that are red (fresh blood) and dark red / black (old blood clots); Morphological Processing: Perform opening and closing operations on the segmented regions to remove noise and connect adjacent regions; Quantization Calculation: Calculate the percentage of the total number of pixels in the bleeding area to the total number of pixels in the entire surgical field (which needs to be predefined through image segmentation), as the "bleeding index".

[0042] B. Sharpness / Haze / Occlusion Assessment: Local Sharpness Calculation: The image is divided into multiple small patches (e.g., 16x16 pixels). For each patch, the variance of its Laplacian operator or Brenner gradient is calculated. These values ​​reflect the richness of image edges and details; higher values ​​indicate greater sharpness. Global Score: The average sharpness value of all patches is calculated, along with the percentage of patches below the sharpness threshold. A low average and a high percentage of low-sharpness patches indicate an overall blurry image (possibly due to lens fog, improper focus, or excessive tissue fluid / irrigation fluid coverage).

[0043] C. Visibility assessment of key structures: Using a lightweight target detection model (such as a lightweight version of YOLO), the visibility of key anatomical structures (such as target blood vessels, tumor boundaries, and nerves) in the current frame is detected in real time, as well as the proportion and outline clarity of the visible parts.

[0044] 2. Comprehensive Score Generation: The above-mentioned evaluation indicators (bleed index, average sharpness, proportion of low-resolution areas, and visibility of key structures) are input into a weighted summation model to generate a comprehensive "visual field quality score." The formula for calculating the visual field quality score Q is: Q = w1× (1-H) + w2× C + w3× (1-O) + w4× V; Wherein: w1, w2, w3, and w4 are the weights of H, C, O, and V corresponding to the current surgical stage, respectively. The values ​​of w1, w2, w3, and w4 can be preset for different surgical stages based on expert experience. H, C, O, and V are normalized values; Bleeding index H: 0 indicates no bleeding, 1 indicates severe bleeding; Clarity index C: 0 indicates completely blurred, 1 indicates very clear; Occlusion index O: 0 indicates no occlusion, 1 indicates complete occlusion; Visibility of key structures V: 0 indicates invisible, 1 indicates fully visible.

[0045] Based on the above, embodiments of the present invention can dynamically adjust the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information. Specifically, the dynamic adjustment of the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information includes: The current risk level and the doctor's current level of focus are determined based on the intraoperative multidimensional status information. Based on the current risk level and the doctor's current focus at the current surgical stage, at least one of the following should be adjusted: the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator.

[0046] This invention dynamically adjusts monitoring parameters based on two core factors: the current risk level at the current surgical stage and the surgeon's current level of focus. The fundamental reason is that these two factors together constitute a "risk-human factor" collaborative decision-making model, which can accurately resolve the core contradiction in intraoperative monitoring: how to focus on the right information at the right time and in the right way. Surgical risks are not uniformly distributed during surgery but fluctuate dramatically with the progress and nature of the procedure. For example, the risk of massive bleeding and nerve damage is much higher during the "lesion resection" stage than during the "incision" or "closure" stages. The computing power, bandwidth, and the attention of medical staff are all limited and scarce resources. A static, average allocation would lead to insufficient resources during high-risk periods and wasted resources during low-risk periods.

[0047] Therefore, positively correlating monitoring intensity with real-time risk levels is an inevitable choice for optimizing resources. When the system determines that the current risk level has increased, it means that patient safety is facing a greater and more urgent threat. At this time, it is imperative to immediately: expand the monitoring scope (increase the set of monitoring indicators): include potentially related risk indicators in the monitoring, such as simultaneously monitoring central venous pressure and hemoglobin trends when the bleeding risk is high; increase monitoring density (increase sampling frequency): more quickly capture instantaneous deterioration of vital signs, gaining time for early intervention; and tighten the safety margin (lower the abnormal threshold): capture abnormal signs with higher sensitivity, erring on the side of adding some controllable early warnings rather than missing a real crisis.

[0048] Adjustments based on risk levels enable the system to precisely focus its "monitoring firepower" on the most dangerous operational aspects, thereby significantly improving safety redundancy during high-risk periods without increasing the overall system load.

[0049] Furthermore, the effectiveness of an alarm depends on the accuracy of human-computer interaction. A perfect anomaly detection algorithm is worthless if its alarms are not effectively received and processed by doctors. The ultimate goal of an alarm is to facilitate appropriate human intervention. However, during high-intensity surgeries lasting several hours, doctors' attention, information processing capabilities, and decision-making abilities can fluctuate due to fatigue, stress, distraction, or technical difficulties.

[0050] Based on this, embodiments of the present invention adaptively adjust the information presentation strategy according to the doctor's real-time status to ensure that key information can "penetrate" the doctor's cognitive load and be effectively perceived. Specifically, when the doctor's concentration is high, they are usually performing delicate operations or making critical decisions. Non-invasive, low-interference alarm methods (such as visual flashing within a microscope's field of view) should be used to avoid startling the doctor and causing operational errors. At this time, for stable or low-risk abnormalities that have been identified, non-critical alarms can even be temporarily suppressed. When the doctor's concentration is low (such as when fatigued, distracted, or dealing with complex problems), assistance and prompts should be proactively enhanced. For example, critical alarms can be upgraded to sound or even tactile prompts; alarm information can be simplified to highlight core decision points; or key information can be simultaneously pushed to the display screen of the assistant doctor or anesthesiologist to establish redundant reminders.

[0051] This invention introduces a physician focus assessment, which can reduce interference when the physician is at their peak ability and enhance assistance when they need it, thereby optimizing the human-machine efficiency of the entire surgical team and reducing medical errors caused by information overload or transmission failure.

[0052] More specifically, determining the risk level is a multi-dimensional information fusion quantitative assessment process aimed at calculating the comprehensive probability of a patient experiencing a serious adverse event in the current surgical context in real time. To ensure the accuracy of the risk level assessment, this embodiment of the invention derives the current risk level for the current surgical stage by weighted fusion of the risk contribution values ​​of the aforementioned four key dimensions, specifically including the following steps: Step 1: Quantify the inherent risks of the surgical phase (R_phase).

[0053] Preset baseline values ​​for each stage: A baseline risk value between 0 and 1 is pre-set for each of the seven surgical stages. This value is determined based on historical surgical complication statistics. In preferred embodiments of the present invention, the baseline risk values ​​for each surgical stage are: R_incision = 0.10; R_craniotomy = 0.35; R_dural opening = 0.40; R_tumor exposure = 0.60; R_lesion resection = 0.90; R_hemostasis = 0.65; R_closure = 0.20.

[0054] Direct assignment: The current surgical stage identified in real time by the intraoperative abnormality identification device of the present invention is the risk value R_phase of this dimension.

[0055] Step 2: Calculate the patient's real-time physiological risk (R_physio).

[0056] 1. Data preparation: Obtain time-series data of the patient's systolic blood pressure (SBP), heart rate (HR), blood oxygen saturation (SpO2), and end-tidal carbon dioxide (ETCO2) within the past 30-second time window (sampling frequency ≥1Hz).

[0057] 2. Calculate the three subfractions: Blood pressure variability fraction (S_bp): Calculate the standard deviation (σ_sbp) and mean (μ_sbp) of systolic blood pressure within this window. Calculate the coefficient of variation: CV_sbp = σ_sbp / μ_sbp. Normalize to the 0-1 range: S_bp = min(CV_sbp / 0.15, 1.0).

[0058] Inappropriate Response Score (S_response): Rule-based judgment: Determines whether a response is inappropriate based on clinical rules. Rule A (Hemorrhage Compensation): If the anomaly detection device simultaneously detects a visual field hemorrhage index > threshold and the heart rate does not increase by more than 10% from baseline, an inappropriate flag is triggered. Rule B (Brainstem Stimulation): If, during a specific operational phase (e.g., posterior fossa manipulation), blood pressure and heart rate do not exhibit the expected transient fluctuations, an inappropriate flag is triggered. Scoring: Within the current time window, S_response increases by 0.5 for each rule triggered, up to a maximum of 1.0.

[0059] Parameter Deviation Score (S_deviation): Sets an individualized ideal range for each parameter (e.g., SBP: target value ±20%; SpO2: >95%). Calculates the average value of each parameter within the current window and its percentage deviation from the ideal range: Deviation % = |Average Value - Target Value| / (Range Width / 2), with a maximum value of 100%. S_deviation = the maximum deviation percentage of the four parameters (0-1).

[0060] 3. Weighted fusion: R_physio=W1 S_bp+W2 S_response+W3 S_deviation. Where W1=0.4, W2=0.3, W3=0.3, and the sum is 1.

[0061] Step 3: Calculate the surgical field risk (R_visual).

[0062] 1. Direct conversion: Obtain the visual field quality score Q calculated in real time as mentioned above.

[0063] 2. Linear transformation: R_visual = 1.0 - Q. This value directly reflects the risk of visual impairment caused by bleeding, fogging, or obstruction.

[0064] Step 4: Calculate trend risk (R_trend).

[0065] 1. Select trend indicators: Select 1-2 core indicators that are most relevant to the current surgical stage for trend analysis. For example, in the lesion resection stage, the core indicators are "systolic blood pressure (SBP)" and "visual field hemorrhage index (H)".

[0066] 2. Calculate the trend slope: For each selected metric, take the data points from the past 60 seconds.

[0067] Use linear least squares to fit a straight line and calculate its slope k (unit: index value / minute). For example, k_sbp = -8 mmHg / minute means that the blood pressure drops by 8 mmHg per minute.

[0068] 3. Normalized to a risk score: A slope threshold (k_threshold) for "clinically significant deterioration" is preset for each indicator. For example, k_threshold for SBP = -10 mmHg / min, and k_threshold for bleeding index H = +0.2 / min.

[0069] The trend score for each indicator is min(|k| / |k_threshold|, 1.0). If the slope is benign (e.g., rising blood pressure), the score is 0.

[0070] 4. Comprehensive calculation: R_trend = the average value of the trend scores of the selected indicators.

[0071] Step 5: Perform integrated calculation of the overall risk level (R_total) 1. Weighted Summation: The risk values ​​of the four dimensions mentioned above are combined according to preset weights. The calculation formula is as follows: R_total = W_phase R_phase + W_physio R_physio + W_visual R_visual +W_trend R_trend; Example weights: W_phase=0.4, W_physio=0.3, W_visual=0.2, W_trend=0.1.

[0072] 2. Limiting and Output: Ensure that the value of R_total is between 0.0 and 1.0. This value represents the overall risk level at the current moment.

[0073] 3. Classification by level: Low risk: R_total < 0.3; Medium risk: 0.3 ≤ R_total < 0.6; High risk: 0.6 ≤ R_total < 0.8; Very high risk: R_total ≥ 0.8.

[0074] The methods for determining a doctor's focus have been described in detail above, and will not be repeated here in the embodiments of the present invention.

[0075] After determining the current risk level and the doctor's current level of focus at the current surgical stage, this embodiment of the invention can adjust at least one of the following based on the current risk level and the doctor's current level of focus at the current surgical stage: the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. Specifically: The adjustment decision in this embodiment of the invention is executed by an engine that includes conditional rules and quantitative formulas. Let the current overall risk level be R, and the doctor's focus level be A.

[0076] The first type of adjustment: adjustments to the set of monitoring indicators.

[0077] Rule engine driven: The intraoperative anomaly identification device in this embodiment of the invention maintains an "indicator-scene" mapping rule base.

[0078] Example rule 1: IF R≥0.6 AND the current stage is “lesion resection” THEN activate the “jugular bulb oxygen saturation” item in the monitoring indicator set.

[0079] Example Rule 2: IF A<0.4 THEN indicates that core blood pressure monitoring simultaneously uses both "invasive arterial pressure" and "non-invasive arterial pressure" data sources for cross-validation.

[0080] At the operational level, the intraoperative anomaly identification device of this invention can dynamically load or unload specific monitoring indicator modules from the overall indicator library according to matching rules.

[0081] The second type of adjustment: adjustment of the sampling frequency.

[0082] Quantification formula adjustment: Each index i has its own base sampling frequency F_base(i).

[0083] Calculate the risk adjustment factor: F_risk(i) = 1.0 + Sens_i R. Where Sens_i is the "risk sensitivity coefficient" of the indicator, with a higher coefficient (e.g., 1.5) for highly important indicators (such as blood pressure) and a lower coefficient (e.g., 0.3) for less important indicators.

[0084] Calculate the attention compensation factor: F_attention(i) = 1.0 + Comp_i (1 - A). Where Comp_i is the “focus compensation coefficient” of the indicator. For alarm indicators that require a fast response (such as heart rate), the coefficient is positive (e.g., 0.8) to increase the sampling rate when the doctor is distracted; for back-end analysis indicators, the coefficient can be 0 or negative to save resources.

[0085] Calculate the new frequency: F_new(i) = F_base(i) F_risk(i) F_attention(i).

[0086] Application upper and lower limits: F_final(i) = max( F_min(i), min( F_max(i), F_new(i) ) ). Where F_min(i) and F_max(i) are predefined by the physical characteristics of the equipment and clinical needs.

[0087] The third type of adjustment: adjustment of abnormal thresholds.

[0088] Quantification formula adjustment: Each threshold (such as the lower limit of systolic blood pressure Th_base) has its base value.

[0089] Determine the direction and magnitude of the adjustment: Base offset (from risk): Offset_risk = T_sens R. T_sens is the "threshold risk sensitivity". For the lower limit of blood pressure, T_sens is positive (e.g., 0.1, meaning that for every 1 increase in risk, the threshold increases by 10% of the base value) to make the threshold more sensitive.

[0090] Human-based adaptation offset (derived from attention): Offset_attention = T_comp (1 - A). T_comp is the "threshold human factor adaptation coefficient", the sign and size of which are set according to the characteristics of the indicator and the strategy to avoid alarm fatigue.

[0091] Calculate the new threshold: Th_new = Th_base + (Offset_risk + Offset_attention) Th_base. For the upper limit threshold, the offset is usually negative to lower the threshold.

[0092] Clinical limits are applied: Th_final = max(Th_abs_min, min(Th_abs_max, Th_new)). Th_abs_min and Th_abs_max refer to the "absolute minimum threshold" and "absolute maximum threshold," respectively. For example, the lower limit of systolic blood pressure should not be lower than 70 mmHg, regardless of any adjustments.

[0093] Building upon the above, and addressing the issue of "lack of personalization in monitoring schemes" in the background technology, this invention introduces personalization at the starting point of dynamic adjustment. This ensures that the initial state of the intraoperative anomaly identification device matches the unique risk spectrum of the patient-disease-doctor, rather than starting from a general default value. Based on this, this invention identifies high-risk patients and high-risk stages before surgery begins, achieving proactive and precise deployment of monitoring resources. Furthermore, the experience of senior surgeons (such as "the tumor is close to a major blood vessel in this location; extra care must be taken with blood pressure during dissection") can be converted into executable initial values ​​for monitoring parameters. Specifically, before surgery begins, the method further includes: Based on patient individual characteristic data, tumor characteristic data, and physician skill data, a personalized monitoring blueprint is generated for different surgical stages. The monitoring blueprint includes the initial set of monitoring indicators for each surgical stage, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold for each monitoring indicator.

[0094] More specifically, Figure 2 This is a schematic diagram illustrating the process of generating personalized monitoring blueprints provided by the present invention, such as... Figure 2 As shown, the generation of personalized monitoring blueprints for different surgical stages based on patient individual characteristic data, tumor characteristic data, and physician skill data specifically includes: Step 201: Based on a pre-constructed multi-dimensional risk assessment model, determine the risk assessment result of the current surgery; wherein, the input parameters of the multi-dimensional risk assessment model include patient baseline parameters, tumor parameters, and surgeon parameters. The patient baseline parameters include: age, severity score of underlying disease, and coagulation function indicators. The tumor parameters include: location risk classification, size, blood supply richness, and distance from important functional areas. The surgeon parameters include: surgical proficiency index and recent complication rate.

[0095] Specifically, the intraoperative anomaly identification device of this invention can automatically extract patient baseline parameters, tumor parameters, and surgeon parameters from the Hospital Information System (HIS), Laboratory Information System (LIS), Picture Archiving System (PACS), hospital surgical database, and hospital quality monitoring database. After obtaining the above parameters, the following steps are performed: 1. Data quantification and preprocessing. This includes: Patient baseline parameter quantification: Age is directly input using the patient's actual age (years); the severity of underlying diseases is scored using a modified Charlesson Comorbidity Index or the American Association of Anesthesiologists (AASA) Physical Status Classification. Specifically, the Charlesson Index is calculated by adding weighted scores corresponding to different diseases based on the patient's medical history (e.g., 1 point for myocardial infarction, 1 point for congestive heart failure, 1 point for diabetes, etc.), and the total score is the final score. ASA classification: determined by the anesthesiologist's preoperative assessment, with grades I to V quantified as 1 to 5 points respectively; coagulation function indicators are comprehensively scored using the International Normalized Ratio (INR), platelet count, and fibrinogen level. Specifically, the normal range is set as follows: INR 0.8-1.2, platelet count >150×10⁻⁶. 9 / L, fibrinogen 2-4 g / L, calculate deviation: index deviation = |measured value - median of normal value| / (normal range width / 2), comprehensive score: take the maximum value or weighted average of the deviation of the three indicators as the coagulation function score, the larger the value, the worse the function.

[0096] Tumor parameter quantification: Location risk grading is based on preoperative MRI, labeled by neurosurgeons or using pre-trained deep learning models, referring to anatomical atlases (such as Brodmann partitions). Specific procedures: Brain regions are divided into non-functional areas (1 point), adjacent functional areas (2 points), functional areas (3 points), and core areas such as the brainstem / thalamus (4 points); size is measured on axial MRI images, determining the maximum tumor diameter (cm); blood supply richness is based on preoperative MRI dynamic contrast-enhanced sequences or perfusion-weighted imaging. Specific procedures: The enhancement rate of the tumor parenchyma or relative cerebral blood flow is calculated and normalized to a score of 0-1 (e.g., no enhancement = 0, significant enhancement = 1); distance to important functional areas is measured in neuronavigation software, determining the shortest Euclidean distance (mm) from the tumor edge to preset key functional cortices such as the motor cortex and language area. Specific procedures: Distance >20mm = 1 point, 10-20mm = 2 points, 5-10mm = 3 points, <5mm = 4 points.

[0097] Surgical surgeon parameter quantification: Surgical proficiency index is based on hospital surgical database statistics. Specific procedures: Calculate the total number of similar tumor resection surgeries (same location, similar size) independently performed by the surgeon in the past three years. Percentile classification by case count: Top 25% = 3 points (proficient), middle 50% = 2 points (competent), bottom 25% = 1 point (novice); Recent complication rate is based on hospital quality monitoring data. Specific procedures: Calculate the percentage of cases with serious intraoperative or postoperative complications (such as massive hemorrhage, new-onset permanent neurological dysfunction) among all neurosurgical surgeries performed by the surgeon in the past year. Complication rate <2% = 1 point (low), 2%-5% = 2 points (medium), >5% = 3 points (high).

[0098] 2. Risk assessment calculation. This includes: Input the current surgical procedure's quantitative parameter vector X=[age, underlying disease score, coagulation score, location risk, tumor size, blood supply score, distance score, proficiency index, complication rate] into the trained multidimensional risk assessment model.

[0099] The model outputs a continuous overall surgical risk score R_overall (e.g., ranging from 0 to 10). This score comprehensively reflects the inherent risk level of the surgery determined by factors from the patient, the tumor, and the physician.

[0100] For the selection and training of the risk assessment model: a multiple linear regression model was adopted. Training data: the above-mentioned quantitative parameters from a large number of historical surgical cases were collected as features, and the "comprehensive risk score" (0-10 points) assessed by experts based on postoperative outcomes (such as whether serious complications occurred, surgical difficulty, and intraoperative stability) was used as a label. Training process: the model was trained using standard methods such as the least squares method to obtain the weight coefficient of each feature.

[0101] Based on this, the embodiments of the present invention can determine the accurate risk assessment results of the current surgery.

[0102] Step 202: Based on the risk assessment results of the current surgery, allocate differentiated monitoring resource budgets for different surgical stages of the current surgery.

[0103] Specifically, after obtaining the risk assessment results for the current surgery, differentiated monitoring resource budgets can be allocated to different surgical stages based on these results. The specific allocation method is as follows: 1. Determine the total monitoring resource budget: Set a total resource budget base B_base (e.g., 100 units) to represent the basic monitoring resource requirements for a routine, low-risk surgery.

[0104] Calculate the total budget: B_total = B_base × (1 + k × R_overall). Here, k is a scaling factor (e.g., 0.1), such that for every 1-point increase in the overall risk score R_overall, the total budget increases by 10%. This ensures that high-risk surgeries receive more monitoring resources.

[0105] 2. Allocate budgets for each stage of the surgery: Define the inherent risk coefficient for each stage: Based on the analysis of a large amount of surgical data, a risk coefficient C_phase is preset for each stage, with a sum of 1. For example: incision (0.05), craniotomy (0.15), dural opening (0.10), tumor exposure (0.20), lesion resection (0.30), hemostasis (0.15), closure (0.05).

[0106] The base budget for the calculation phase is: B_phase_base = B_total × C_phase.

[0107] Adjustments based on physician proficiency: Introduce a physician proficiency adjustment factor λ. For physicians with a high proficiency index (e.g., 3 points), λ = 0.9, indicating strong risk control capabilities, allowing for a slight reduction in budget; for physicians with low proficiency (e.g., 1 point), λ = 1.1, indicating a need to increase budget to strengthen monitoring.

[0108] Final phase budget: B_phase_final = B_phase_base × λ. The budget for each phase represents the resources available for enabling auxiliary monitoring metrics and increasing sampling frequency in that phase.

[0109] Based on this, the embodiments of the present invention can ensure the rationality of the monitoring resource budget allocated to different surgical stages before surgery, and avoid monitoring pressure, i.e., monitoring resource waste.

[0110] Step 203: Based on the monitoring resource budget and inherent risks of different surgical stages, determine the initial set of monitoring indicators, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold of each monitoring indicator for each surgical stage; wherein, the initial set of monitoring indicators for each surgical stage includes a subset of core monitoring indicators and a subset of auxiliary monitoring indicators, and the initial sampling frequency of each monitoring indicator is positively correlated with the inherent risk of the corresponding surgical stage.

[0111] Specifically, the calculation method for determining the initial monitoring indicator set for each stage is as follows: 1. Establish a basic database: Define the core indicator set: identify five basic items that must be monitored throughout the entire process and do not consume dynamic budget: electrocardiogram, invasive arterial blood pressure, pulse oxygen saturation, end-tidal carbon dioxide, and body temperature. Their resource cost is recorded as 0.

[0112] Define the auxiliary indicator library and costs: List the optional specific monitoring items, and have the systems engineer and clinical experts jointly assign an integer resource cost to each item (representing its computational and attentional load). Examples are as follows: Central venous pressure: cost 5; Motor evoked potentials: cost 15; Somatosensory evoked potentials: cost 10; Brain oxygen saturation: cost 8; Intracranial pressure: cost 12; Intraoperative ultrasound: cost 20.

[0113] Definition Phase - Indicator Correlation Matrix: This matrix consists of scores given back-to-back by at least three senior neurosurgeons, averaged to form a 7-row (phase) × N-column (auxiliary indicator) table. Each cell value ranges from 0 to 1, representing the importance of the indicator at that phase. For example, the correlation score for "motor evoked potentials" can reach 0.9 in the "lesion resection" phase and 0.1 in the "incision" phase.

[0114] 2. Execute the "cost-benefit optimization selection" algorithm (calculated independently for each stage): Input: The allocation budget B for this stage (integer, such as 45 units).

[0115] Step 1: Calculate the benefit-cost ratio. For this stage, calculate the benefit-cost ratio V_i for each indicator i in the auxiliary indicator library: V_i = Relevance score Rel_i of the indicator in this stage / Resource cost Cost_i of the indicator.

[0116] Step 2: Sort. Sort all auxiliary indicators in descending order of V_i value from highest to lowest.

[0117] Step 3: Iterative selection.

[0118] 1) Start from the top of the sort list and examine the first metric.

[0119] 2) If the Cost_i of this indicator is less than or equal to the current remaining budget B_remaining, then add this indicator to the auxiliary monitoring subset of this stage and update the remaining budget: B_remaining = B_remaining - Cost_i.

[0120] 3) If the Cost_i of this metric is greater than the current remaining budget, then skip this metric.

[0121] 4) Move to the next metric in the list and repeat steps 2)-3) until all metrics have been checked or the budget is exhausted.

[0122] Output: Initial set of monitoring indicators for this stage = core indicator set ∪ selected subset of auxiliary indicators.

[0123] The specific calculation method for determining the initial sampling frequency of each monitoring indicator is as follows: 1. Set the base frequency: Set a base sampling frequency F_base(i) (unit: Hertz Hz) for each monitoring indicator i under calm, low-risk conditions. For example: invasive arterial blood pressure: 100 Hz; electrocardiogram: 500 Hz; pulse oximetry: 1 Hz; exercise evoked potential: 0.2 Hz (i.e., once every 5 seconds); body temperature: 0.05 Hz (i.e., once every 20 seconds).

[0124] 2. Define the risk sensitivity coefficient: Assign a risk sensitivity coefficient k_i to each indicator i. This coefficient represents the degree to which the sampling frequency of that indicator should be amplified as the risk increases. It is determined by clinical importance; for example: invasive arterial blood pressure k_i = 1.0 (high sensitivity); electrocardiogram k_i = 0.8; pulse oximetry k_i = 0.5; exercise evoked potentials k_i = 0.6; body temperature k_i = 0.2 (low sensitivity). 3. Calculate the initial frequency: For a specific surgical phase, its inherent risk coefficient is C_phase (a decimal between 0 and 1, such as 0.30 for the lesion resection phase). Then, the initial sampling frequency F_initial(i, phase) of index i in this phase is calculated using the following formula: F_initial(i, phase) = F_base(i) × [1 + (k_i × C_phase)]; For example, during the lesion resection phase (C_phase=0.30), the initial frequency of invasive arterial blood pressure is calculated: F_initial(IBP, cut-off) = 100 Hz × [1 + (1.0 × 0.30)] = 100 Hz × 1.30 = 130 Hz.

[0125] Boundary constraints: The calculated F_initial must be limited to the range of [F_min(i), F_max(i)] allowed by the physical capabilities of the monitoring device.

[0126] The specific calculation method for determining the initial anomaly thresholds for each monitoring indicator is as follows: 1. Obtain individualized patient baseline: After anesthesia induction is completed and the patient's vital signs are stable, record the average values ​​of key physiological parameters over 5 minutes as individual baseline values. For example: Systolic blood pressure baseline (SBP_base): the average arterial systolic blood pressure during the stable period; Diastolic blood pressure baseline (DBP_base): the average arterial diastolic blood pressure during the stable period; Heart rate baseline (HR_base): the average heart rate during the stable period; Blood oxygen saturation baseline (SpO2_base): the average blood oxygen saturation during the stable period; End-tidal carbon dioxide baseline (ETCO2_base): the average ETCO2 during the stable period.

[0127] 2. Calculate the baseline threshold: Based on clinical guidelines and individual patient baselines, calculate the preliminary alarm threshold: Lower limit of systolic blood pressure: Th_sbp_low_base = max(90 mmHg, 0.8 × SBP_base) (take the larger value between 90 mmHg and 80% of the baseline, taking into account both general standards and individualization).

[0128] Upper limit of heart rate: Th_hr_high_base = min(130 bpm, 1.4 × HR_base) (take the smaller value between 130 bpm and 140% of the baseline).

[0129] Lower limit of blood oxygen saturation: Th_spo2_low_base = max(92%, SpO2_base - 3%) (take the larger value between 92% and 3% lower than the baseline).

[0130] Lower limit of carbon dioxide at end-tidal: Th_etco2_low_base = max(30 mmHg, 0.85 × ETCO2_base).

[0131] Upper limit of carbon dioxide at end-tidal: Th_etco2_high_base = min(45 mmHg, 1.15 × ETCO2_base).

[0132] 3. Application Stage Inherent Risk Adjustment: For each surgical stage, the inherent risk coefficient (R_phase) of that stage is used for threshold adjustment. R_phase is a decimal between 0 and 1, determined by clinical statistics (e.g., incision 0.05, craniotomy 0.15, dural opening 0.10, tumor exposure 0.20, lesion resection 0.30, hemostasis 0.15, closure 0.05).

[0133] Calculate the phase risk adjustment factor: α_phase = R_phase × k_threshold Where k_threshold is the threshold adjustment sensitivity coefficient, which is generally taken as 0.1-0.2. For example, if k_threshold=0.15, then α_phase=0.045 in the lesion resection stage (R_phase=0.30).

[0134] Adjust the threshold (adjusting in different directions depending on the threshold type): Safety lower limit threshold (such as lower limit of blood pressure, lower limit of blood oxygen): Th_low_adj = Th_low_base×(1 +α_phase) (Raise the lower limit to make the alarm more sensitive).

[0135] Safety upper limit threshold (such as upper limit of heart rate, upper limit of ETCO2): Th_high_adj = Th_high_base × (1 -α_phase) (lower the upper limit to make the alarm more sensitive).

[0136] Two-way threshold (e.g., ETCO2 range): Adjusts both the upper and lower limits simultaneously to narrow the range.

[0137] For example: During the lesion resection phase (R_phase=0.30, α_phase=0.045), the patient's baseline systolic blood pressure was 120 mmHg. Lower limit of basal systolic blood pressure: Th_sbp_low_base = max(90, 0.8×120) = max(90, 96) = 96mmHg; Adjusted lower limit of systolic blood pressure: Th_sbp_low_adj = 96×(1 + 0.045) = 96×1.045≈100mmHg.

[0138] 4. Application of Physician Proficiency Compensation Adjustment: A secondary fine-tuning is performed based on the surgeon's proficiency index to balance safety and interference. Obtain the doctor's proficiency index: such as 3 points (expert), 2 points (proficient), 1 point (novice).

[0139] Calculate the proficiency adjustment factor: Expert (3 points): β_skill = 0.98 (slightly relax the threshold to reduce unnecessary interference with experts); Proficiency (2 points): β_skill = 1.00 (no adjustment); Beginner (1 point): β_skill = 1.05 (Tighten the threshold appropriately to provide more protective warnings); Application adjustment: Th_tuned = Th_adj × β_skill.

[0140] 5. Apply absolute clinical safety limits: To ensure physiological safety, all thresholds must be limited to medically recognized absolute safety ranges. Systolic blood pressure lower limit: absolute minimum 70 mmHg, absolute maximum 120 mmHg; Heart rate limit: Absolute maximum 150 bpm; Lower limit of blood oxygen saturation: absolute minimum 88%; End-tidal carbon dioxide range: absolute lower limit 25 mmHg, absolute upper limit 55 mmHg; Final threshold calculation formula: Th_final = max( Th_abs_min, min( Th_abs_max, Th_tuned ) ).

[0141] Where: Th_abs_min: the absolute minimum value of the threshold for this type; Th_abs_max: the absolute maximum value of the threshold for this type; the min() function ensures that the threshold is not higher than the absolute maximum value; the max() function ensures that the threshold is not lower than the absolute minimum value.

[0142] Through the above steps, this embodiment of the invention generates personalized initial abnormality thresholds for each monitoring indicator at each surgical stage. These thresholds take into account both individual patient differences and the inherent risk levels of different surgical stages, while also being moderately optimized based on physician experience, ultimately ensuring the most suitable alarm sensitivity within an absolutely clinically safe range.

[0143] Furthermore, Figure 3 This is a schematic diagram of the processing flow for the detection result delay provided by the present invention, as shown below. Figure 3 As shown, in the process of acquiring multidimensional intraoperative status information in real time, the method further includes: Step 1011: Determine the target monitoring indicator whose detection result delay exceeds the preset tolerance time; Step 1012: Input the historical trend data of the target monitoring indicator, the associated monitoring data of the target monitoring indicator, the current surgical stage information, and the current operation type information into the pre-established short-term prediction model of the monitoring indicator to obtain the current estimated value of the target monitoring indicator; Step 1013: Based on the confidence level of the current estimated value of the target monitoring indicator, determine whether to trigger a pre-alarm.

[0144] Specifically, during neurosurgical procedures, some critical monitoring data (such as blood gas analysis, coagulation function, and certain drug concentrations) have an inherent delay of 30 seconds to several minutes. Existing systems can only display old data or blank data during this waiting period, creating a monitoring blind spot. To address this issue, this invention provides intelligent predictive values ​​during the delay period, achieving continuous monitoring. Through predictive models, early warnings can be issued 1-3 minutes before the actual data arrives, giving the clinical team valuable response time and truly achieving preventative monitoring.

[0145] More specifically, this embodiment of the invention pre-establishes a database of delayed indicators, specifically maintaining a list of delayed indicators that need to be monitored, including: blood gas analysis indicators: arterial partial pressure of oxygen (PaO2), arterial partial pressure of carbon dioxide (PaCO2), pH value, and base excess (BE); coagulation function indicators: activated clotting time (ACT), prothrombin time (PT), and activated partial thromboplastin time (APTT); biochemical indicators: blood glucose, electrolytes (potassium, sodium, calcium), and hemoglobin; and drug concentration indicators: blood concentration of anesthetic drugs (such as propofol and remifentanil). Simultaneously, based on clinical urgency and the testing process, specific thresholds (i.e., preset tolerance times) are set for each indicator: blood gas analysis (routine): 180 seconds; blood gas analysis (urgent): 90 seconds; coagulation function (ACT): 120 seconds; blood glucose: 300 seconds; electrolytes: 600 seconds.

[0146] Based on this, embodiments of the present invention can determine the target monitoring indicator whose detection result delay exceeds a preset tolerance time. After determining the target monitoring indicator, embodiments of the present invention can input the historical trend data of the target monitoring indicator, the associated monitoring data of the target monitoring indicator, the current surgical stage information, and the current operation type information into a pre-established short-term prediction model for the monitoring indicator to obtain the current estimated value of the target monitoring indicator. Specifically, the extraction of historical trend data can be achieved in the following way: For the target monitoring indicator X, extract the most recent N valid data points (e.g., N=6, with a time span of the past 15-30 minutes), and calculate the following features: 1. The moving average of the most recent 3 points; 2. The first derivative (rate of change): ΔX / Δt; 3. The second derivative (acceleration): Δ(ΔX / Δt) / Δt; 4. The trend direction: the slope of the fitted line of the most recent 3 points.

[0147] The selection of associated monitoring data can be based on public medical knowledge bases. Taking PaO2 as an example, its associated monitoring data includes: Real-time physiological parameters: pulse oxygen saturation (SpO2): correlation coefficient 0.85; end-tidal carbon dioxide (ETCO2): correlation coefficient -0.70; peak airway pressure (Ppeak): correlation coefficient -0.60.

[0148] Ventilator settings: Inhaled oxygen concentration (FiO2): directly affects PaO2; Positive end-expiratory pressure (PEEP): affects oxygenation; Tidal volume (Vt): affects ventilation.

[0149] Patient status: Mean arterial pressure (MAP): reflects perfusion; Heart rate (HR): reflects metabolic demand.

[0150] The methods for acquiring information on the current surgical stage have been described in detail above and will not be repeated here. The current operation type information is identified from the surgical instrument usage signals: 1. Electrocoagulation hemostasis: may cause tissue hypoxia; 2. Tumor traction: may affect local blood flow; 3. Irrigation and suction: may alter the local environment. The seven surgical stages are uniquely encoded and used as input to the short-term prediction model for monitoring indicators. The current operation type information is encoded as an operation type vector and used as input to the short-term prediction model for monitoring indicators.

[0151] The preferred short-term prediction model for monitoring indicators is the Lightweight Gradient Boosting Decision Tree (LightGBM) model. For model training, complete monitoring data from the past 300 neurosurgical procedures can be collected. The timestamps of the multi-source data are aligned at a 1Hz frequency, and the actual delay detection values ​​are used as the prediction targets (i.e., labels). The data is then divided into training, validation, and test sets in a 7:2:1 ratio for model training and validation.

[0152] Once the target monitoring indicator is marked, the intraoperative abnormality identification device performs the following steps: 1. Immediately extract the complete feature vector at the current time point; 2. Call the corresponding pre-trained model (i.e., the short-term prediction model for monitoring indicators; each delayed indicator has its own independent model). 3. Model output: Predicted value (e.g., PaO2 = 95 mmHg); 4. Prediction result caching: Store the predicted values ​​in a circular buffer for subsequent analysis.

[0153] Based on this, embodiments of the present invention further determine the confidence level of the current estimated value of the target monitoring indicator. Specifically, embodiments of the present invention calculate the confidence level by comprehensively considering the internal confidence level of the model, the feature consistency score, and the prediction stability assessment.

[0154] The model's internal confidence level is calculated as follows: For tree models, the variance of the leaf node sample distribution is used to calculate the model's internal confidence score: 1 - (variance of predicted values ​​ / preset maximum variance threshold), normalized to the range of 0-1.

[0155] Feature consistency scores are calculated as follows: Calculate the similarity between the current feature vector and the features in the training set: 1. Find the K most similar samples in the training set (K=10); 2. Calculate the standard deviation σ of the actual values ​​of these samples; 3. Feature consistency score = 1 - min(σ / threshold, 1.0).

[0156] The predicted stability is calculated in the following way: Check the changes in the last 3 predicted values: Stability score = 1 - (standard deviation of the last 3 predicted values ​​ / preset threshold).

[0157] Finally, the total confidence score (i.e., the confidence score of the current estimate) = 0.5 × model confidence score + 0.3 × feature consistency score + 0.2 × stability score. The total confidence score is categorized as follows: High confidence: ≥0.8; Medium confidence: 0.6-0.8; Low confidence: <0.6.

[0158] Based on this, embodiments of the present invention can determine whether to trigger a pre-alarm based on the confidence level of the current estimated value. Specifically: Case 1: High confidence level (≥0.8) and the predicted value exceeds the threshold, triggering a yellow pre-alarm; Case 2: Medium confidence level (0.6-0.8) and the predicted value exceeds the threshold, triggering a blue alert; Case 3: Low confidence level (<0.6) or the predicted value is within the normal range, no alarm is triggered, and the predicted value is only recorded in the background.

[0159] Furthermore, the method also includes: The actual detection results of the target monitoring indicator are compared with the corresponding predicted values, and the short-term prediction model of the monitoring indicator is calibrated based on the comparison results.

[0160] Specifically, considering that the prediction model uses historical data during training, but each surgery and each patient is unique, the intraoperative situation may differ from the distribution of the training data, leading to a decrease in prediction performance. The online calibration mechanism enables the model to adapt to the specific context of the current surgery in real time, transforming a general model into a personalized model and significantly improving prediction accuracy for the current patient. More specifically, if the deviation between the actual detection result and the corresponding predicted value exceeds a preset threshold, calibration is performed. The specific calibration method can be online incremental learning calibration, Bayesian model update calibration, or any other feasible calibration method; this embodiment of the invention does not specifically limit this.

[0161] Step 102: Collect monitoring data based on the dynamically adjusted monitoring parameters, and perform multi-level anomaly detection based on the monitoring data.

[0162] Specifically, to further improve the accuracy of anomaly identification, this embodiment of the invention performs multi-level anomaly detection based on the monitoring data. More specifically, performing multi-level anomaly detection based on the monitoring data includes: Based on the monitoring data, single-indicator real-time threshold comparison, multi-indicator correlation analysis, time window-based trend analysis, and machine learning model-based predictive detection are performed to determine the confidence score of intraoperative abnormalities. Based on the intraoperative abnormality confidence score, it is determined whether an intraoperative abnormality exists.

[0163] Specifically, the formula for calculating the intraoperative abnormality confidence score is as follows: Score = α×S1+ β×S2+ γ×S3+ δ×S4; Wherein: S1 is the normalized score for the degree of anomaly of a single indicator, S2 is the consistency score for the correlation of multiple indicators, S3 is the score for the persistence of trends, and S4 is the confidence score for the prediction model; α, β, γ, and δ are dynamic weight coefficients that are adjusted according to the current surgical stage and risk level.

[0164] This invention addresses two core problems inherent in traditional intraoperative monitoring systems that rely on single-threshold alarms: a high false alarm rate due to transient interference or individual differences triggering unnecessary alarms, and a risk of missed alarms because slowly developing anomalies fail to reach a fixed threshold in time. By introducing a four-level progressive verification mechanism, the system moves from isolated single-point judgments to comprehensive pattern recognition, significantly improving the overall accuracy of anomaly identification. Specifically, these four levels cover different stages of anomaly development: trend analysis and predictive detection capture early signs of deterioration, correlation analysis confirms multi-indicator synergistic anomaly patterns, and threshold comparison provides final confirmation of clearly serious anomalies. This structure forms a full-time monitoring network. Crucially, the dynamic weighting coefficient mechanism allows the system to intelligently adapt to different surgical scenarios, such as emphasizing real-time threat response in high-risk phases and early trend warnings in stable phases. Ultimately, all evidence is integrated into a unified quantitative confidence score, providing an objective and interpretable basis for deciding whether to trigger an alarm and its level, effectively balancing the sensitivity and specificity of monitoring.

[0165] Specifically, the calculation method for the Level 4 test score is as follows: Level 1: Calculation of real-time threshold comparison score for a single indicator (S1).

[0166] Input: Real-time measured values ​​of all monitoring indicators and their corresponding dynamically adjusted alarm thresholds.

[0167] step: a. For each metric, calculate the degree to which its current value deviates from the personalized safety threshold. If the current value exceeds the upper limit, the deviation is calculated as: (current value - upper limit threshold) / (upper limit threshold - normal baseline value for this metric). If it is below the lower limit, the deviation is: (lower limit threshold - current value) / (normal baseline value - lower limit threshold). If it is within the normal range, the deviation is 0.

[0168] b. Map the calculated deviation to a range of 0 to 1 using an S-shaped function (such as the sigmoid function) to obtain an anomaly score for the indicator. This mapping results in low scores for small deviations and near-perfect scores (1 point) for severe deviations.

[0169] c. Among all monitoring indicators, select the highest calculated anomaly score as the final output result S1 for this level.

[0170] Level 2: Calculation of the multi-indicator association consistency score (S2).

[0171] Input: Real-time values ​​and trends of all monitored indicators, as well as current surgical context information.

[0172] step: a. The intraoperative abnormality recognition device has a built-in rule base for abnormality patterns defined by clinical knowledge. Each rule describes the typical change pattern of a set of indicators that should appear simultaneously for a specific abnormality (such as active hemorrhage, brainstem stimulation, or hypoventilation).

[0173] b. For each rule in the rule base, check the degree to which its preset conditions are met at present, and calculate a satisfaction score of 0 to 1 for each condition.

[0174] c. Calculate the geometric mean of the satisfaction scores for all conditions under this rule to obtain the "activation degree" of the rule.

[0175] d. Iterate through all rules and take the highest activation value as the association consistency score S2. If a rule is activated in multiple consecutive detection periods, its activation value will receive an additional bonus to reflect the persistence of the abnormal pattern.

[0176] Level 3: Calculation of trend persistence score (S3) based on time window.

[0177] Input: Historical data sequence of key monitoring indicators within a past time window (e.g., 2 minutes).

[0178] step: a. For each key indicator, use linear regression to fit its trend over the time window to obtain the trend slope (e.g., how many millimeters of mercury the blood pressure drops per minute) and its statistical significance (p-value).

[0179] b. Compare the trend slope with the preset "clinically significant change threshold" of the indicator, and calculate an initial trend risk score based on the statistical significance of the trend.

[0180] c. Assess the persistence of the trend over a longer time window. If the trend persists, make a positive revision to the initial risk score.

[0181] d. Select the highest value from the trend risk scores of all key indicators as the output S3 for this level.

[0182] Level 4: Calculation of predictive detection score (S4) based on machine learning model.

[0183] Input: Joint time series features composed of multiple key indicators in recent times (e.g., the past 30 seconds).

[0184] step: a. Analyze the current joint feature state using a lightweight temporal anomaly detection model (such as an isolated forest or an LSTM autoencoder).

[0185] b. The model outputs a raw anomaly score of 0 to 1, reflecting the degree of deviation between the current overall state of multiple indicators and the historical "normal" pattern.

[0186] c. The original anomaly score is calibrated by combining the model's own confidence level with the frequency (typicality) of the current feature in the training data to obtain the final predictive detection score S4. This score will be further improved if the model predicts that the state will exceed the safe range in the next short period (e.g., 30 seconds).

[0187] The adjustment method for the dynamic weighting coefficients (α, β, γ, δ) is as follows: The adjustment of dynamic weighting coefficients is based on two core factors: the pre-set surgical stage and the real-time calculated risk level.

[0188] 1. Preset weights based on surgical stage: This embodiment of the invention pre-sets a set of basic weights for each surgical stage, reflecting the focus of the monitoring strategy for that stage. For example: in the incision or closure stage (low risk), the weights are biased towards trend analysis (γ) and predictive detection (δ) to achieve early warning; in the lesion resection stage (high risk), the weights are significantly tilted towards real-time threshold comparison (α) and correlation analysis (β) to ensure rapid response to immediate threats; and in the hemostasis stage, extremely high attention is paid to clear abnormal indicators (high α).

[0189] 2. Dynamic adjustment based on real-time risk levels: In this embodiment of the invention, the current comprehensive risk level R (a value between 0 and 1) is obtained from an independent risk assessment module.

[0190] An adjustment vector is calculated based on the risk level: at high risk, the weight increments of the real-time threshold and correlation analysis are increased (Δα, Δβ are positive), while the weight increments of the trend and prediction are decreased (Δγ, Δδ are negative), forcing the intraoperative anomaly detection device to focus on immediate threats.

[0191] The temporary weights are obtained by adding the base weights of the current surgical stage to the risk adjustment vector.

[0192] The temporary weights are normalized to ensure that the sum of the four coefficients α, β, γ, and δ is always 1, thus obtaining the final dynamic weights for application.

[0193] The confidence score calculation and anomaly detection process is as follows: 1. Calculate the overall confidence score (Score): Use the formula Score = α×S1+ β×S2+ γ×S3+ δ×S4 to sum the four levels of scores according to dynamic weights.

[0194] 2. Dynamic Threshold for Anomaly Detection: The threshold for anomaly detection is not a fixed value. The base threshold is typically set at 0.7. This threshold is dynamically adjusted based on the current risk level: the higher the risk level, the lower the threshold will be, making anomaly detection more sensitive. The doctor's focus is also considered; if the doctor's focus is insufficient, the threshold will be appropriately increased to reduce potential interfering alarms.

[0195] 3. Final decision logic: If the score is greater than or equal to the dynamic judgment threshold, an abnormality is determined to exist. The intraoperative abnormality identification device will further analyze the sub-scores from S1 to S4 to determine the specific type of abnormality (such as bleeding, nerve damage, etc.) and its urgency.

[0196] The level of urgency is mainly determined by whether S1 and S2 are extremely high (indicating a serious immediate threat) or whether the total score far exceeds the threshold.

[0197] If the score does not reach the alarm threshold, but the S3 or S4 score is high, the intraoperative abnormality identification device will not trigger an alarm, but may prompt in the background that monitoring needs to be strengthened, reflecting early attention to potential risks.

[0198] Step 103: If an intraoperative abnormality is determined, a graded alarm signal is output based on the type of abnormality, severity level, and current surgical context.

[0199] Specifically, the step of outputting a graded alarm signal based on the abnormality type, severity level, and current surgical context includes: Based on the type and severity of the abnormality, combined with the current stage of surgery, the patient's real-time condition, the doctor's focus, and the quality of the surgical field of vision, a comprehensive urgency score is calculated. The alarm level is determined based on the comprehensive urgency score, and the alarm sensory channel and information presentation density are adaptively selected based on the alarm level, the current surgical stage, and the doctor's focus, and a corresponding graded alarm signal is generated and output.

[0200] More specifically, this invention completely transforms the traditional "one-size-fits-all" approach of alarm systems, achieving a leap from "mechanical alarms" to "intelligent notifications." By deeply integrating the attributes of the anomaly itself (type and severity) with the real-time surgical context (stage, patient, doctor, field of vision), a quantified "comprehensive urgency" is calculated. This makes the urgency assessment of alarms more accurate, greatly reducing false alarms caused by detachment from the context. More importantly, the system can classify urgency based on this urgency level and intelligently select the most appropriate alarm method (such as a combination of visual, auditory, and tactile) and information density, ensuring that critical alarms are effectively received by the doctors who need the most attention in the first instance, while avoiding unnecessary shock or information overload during delicate procedures. This ultimately forms a closed loop of "conveying the right information at the right time, in the right way," significantly improving the work efficiency and emergency response capabilities of the surgical team while ensuring patient safety.

[0201] The specific calculation method for the comprehensive urgency score is as follows: 1. Determining the baseline score for abnormalities: The intraoperative abnormality identification device in this embodiment of the invention has a built-in table corresponding to abnormality types and levels. Based on the identified abnormality type (such as active bleeding, nerve injury, ventilation impairment, etc.) and its severity level, a baseline score is obtained from the table. For example, "major bleeding" may be assigned a baseline score of 0.8, and "mild oxygenation decline" may be assigned a baseline score of 0.3.

[0202] 2. Calculate the surgical stage coefficient: A risk multiplier is preset for each surgical stage. The multiplier for high-inherent-risk stages (such as lesion resection) is greater than 1 (e.g., 1.5), and the multiplier for low-risk stages (such as closure) is less than 1 (e.g., 0.7). Multiplying the multiplier of the current surgical stage by the abnormal baseline score reflects the different urgency of the same abnormality at different stages.

[0203] 3. Calculate the patient status correction factor: Assess the stability of the patient's current physiological state. For example, calculate the coefficient of variation of key vital signs (blood pressure, heart rate, blood oxygen) within the recent window. The more unstable the patient's status (higher coefficient of variation), the larger the correction factor (e.g., from 1.0 to 1.2), thereby increasing the urgency level.

[0204] 4. Calculate the physician focus compensation coefficient: Compensation is based on the physician focus score assessed in real time. When the physician's focus is low (e.g., below 0.5), it is believed that stronger stimulation is needed to attract attention, so the compensation coefficient will be greater than 1 (e.g., 1.3); when the physician is highly focused, the compensation coefficient is 1 or slightly lower than 1 to avoid excessive interference.

[0205] 5. Calculate the impact coefficient of visual field quality: If the current abnormality is strongly correlated with visual monitoring (e.g., bleeding), but the surgical visual field quality score is poor (e.g., severe occlusion), it means that the doctor has difficulty directly observing and confirming it, and a more proactive alarm is needed. In this case, the impact coefficient will be greater than 1 (e.g., 1.2). If the abnormality is not visually related, this coefficient is 1.

[0206] 6. Weighted Fusion to Obtain the Final Score: The results from the five components mentioned above (abnormal baseline score, stage coefficient, patient correction, attention compensation, and visual field impact) are multiplied sequentially to obtain an initial urgency value. The initial value is then fine-tuned by considering the trend of the abnormality (e.g., whether it is rapidly deteriorating). Finally, the calculation result is limited to the range of 0 to 1 to obtain the final comprehensive urgency score.

[0207] The specific method for generating and outputting adaptive hierarchical alarm signals is as follows: 1. Determine the alarm level: This embodiment of the invention presets five alarm levels, which correspond to the urgency rating range. For example: Level 1 (Information): Urgency < 0.3, only recorded in the background; Level 2 (Attention): 0.3 ≤ Urgency < 0.5, non-intrusive prompt; Level 3 (Warning): 0.5 ≤ Urgency < 0.7, prompt requiring attention; Level 4 (Severe): 0.7 ≤ Urgency < 0.9, prompt requiring immediate action; Level 5 (Urgent): Urgency ≥ 0.9, prompt requiring emergency intervention.

[0208] 2. Adaptive selection of alarm sensory channels: Based on the determined alarm level, combined with the current surgical stage and the doctor's real-time focus, one or more combinations of visual, auditory, and tactile channels can be selected.

[0209] Example of decision-making rules: Level 2 / 3 alarms: The visual channel is usually selected by default (e.g., a flashing icon in a specific area of ​​the screen). If the doctor's concentration is low, a short, gentle audible alert can be added. Level 4 alarms: Must include an auditory channel (continuous specific tone) and an enhanced visual channel (bright flashing in the center of the screen). If the doctor is operating through a microscope, the tactile channel integrated on the instrument handle (gentle vibration) can be additionally activated to avoid sound startle. Level 5 alarms: Activate all channels. Includes a unique mandatory auditory alarm, a visual warning box covering the main field of vision, and may include a voice announcement of the abnormality type via the operating room broadcast. Special situations: During extremely delicate procedures such as "suturing blood vessels," even for Level 4 alarms, the audible alert may be temporarily replaced with stronger tactile feedback to prevent hand tremors.

[0210] 3. Adaptive Adjustment of Information Presentation Density: High-Density Mode: When the doctor's focus is high and distraction is permissible during the surgical phase (such as tissue suturing), alarm information can display more detailed content, such as abnormal values, trend graphs, and a list of possible causes. Low-Density Mode: When the doctor's focus is low or during high-risk procedures (such as tumor removal), alarm information will be greatly simplified, presenting only the most critical elements, such as the abnormality name, the most critical value, and the primary response action (such as "bleeding - quick check"), to minimize cognitive load and facilitate rapid decision-making.

[0211] 4. Signal Generation and Output: The intraoperative anomaly detection device translates the aforementioned decisions (level, channel, information content) into specific control commands. These commands are sent in real time to the corresponding output devices: a graphical interface driver (controlling visual elements in specific areas of the display screen), an audio controller (controlling the sound from speakers or headphones), and a haptic feedback driver (controlling the vibration motors of instrument handles or seats). Ultimately, the surgeon and surgical team will perceive alarm signals with appropriate levels, channels, and information density through these devices, and thus take corresponding actions.

[0212] The solution provided by this invention acquires multi-dimensional intraoperative status information in real time and dynamically adjusts the monitoring parameters for the current surgical stage based on this information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each indicator, and the actual abnormal threshold for each indicator. The multi-dimensional status information includes current surgical stage identification information, real-time patient physiological parameters, physician focus assessment information, and surgical field quality assessment information. Monitoring data is collected based on the dynamically adjusted monitoring parameters, and multi-level abnormality detection is performed based on this data. When an intraoperative abnormality is detected, a graded alarm signal is output based on the abnormality type, severity level, and current surgical context. This enables intelligent allocation of monitoring resources, significantly reducing system load and manual intervention requirements while ensuring safety, and simultaneously improving the accuracy and timeliness of abnormality identification, reducing the risk of missed and false alarms.

[0213] Figure 4 This invention provides a structural block diagram of an intraoperative anomaly identification device based on multimodal data, as shown in the embodiment of the invention. Figure 4 As shown, the system includes: The monitoring parameter dynamic adjustment module 301 is used to acquire intraoperative multidimensional status information in real time and dynamically adjust the monitoring parameters of the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators of the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes the current surgical stage identification information, the patient's real-time physiological parameters, the doctor's focus assessment information, and the surgical field quality assessment information. Anomaly detection module 302 is used to collect monitoring data based on dynamically adjusted monitoring parameters, and to perform multi-level anomaly detection based on the monitoring data; The alarm module 303 is used to output a graded alarm signal based on the type of abnormality, severity level and current surgical context when an intraoperative abnormality is determined to exist.

[0214] The solution provided by this invention acquires multi-dimensional intraoperative status information in real time through a monitoring parameter dynamic adjustment module 301, and dynamically adjusts the monitoring parameters for the current surgical stage based on the monitoring parameter dynamic adjustment module 301. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multi-dimensional status information includes the current surgical stage identification information, the patient's real-time physiological parameters, the doctor's focus assessment information, and the surgical field quality assessment information. An abnormality detection module 302 collects monitoring data based on the dynamically adjusted monitoring parameters and performs multi-level abnormality detection based on the monitoring data. When an intraoperative abnormality is determined, an alarm module 303 outputs a graded alarm signal based on the abnormality type, severity level, and current surgical context. This enables intelligent allocation of monitoring resources, greatly reduces system load and manual intervention requirements while ensuring safety, and significantly improves the accuracy and timeliness of abnormality identification, reducing the risk of missed and false alarms.

[0215] In one optional embodiment of this application, the apparatus further includes a monitoring blueprint generation module, which is used to perform the following steps: Based on patient individual characteristic data, tumor characteristic data, and physician skill data, a personalized monitoring blueprint is generated for different surgical stages. The monitoring blueprint includes the initial set of monitoring indicators for each surgical stage, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold for each monitoring indicator.

[0216] In one optional embodiment of this application, generating personalized monitoring blueprints for different surgical stages based on patient individual characteristic data, tumor characteristic data, and physician skill data specifically includes: Based on a pre-constructed multi-dimensional risk assessment model, the risk assessment result of the current surgery is determined. The input parameters of the multi-dimensional risk assessment model include patient baseline parameters, tumor parameters, and surgeon parameters. The patient baseline parameters include age, severity score of underlying disease, and coagulation function indicators. The tumor parameters include location risk classification, size, blood supply richness, and distance from important functional areas. The surgeon parameters include surgical proficiency index and recent complication rate. Based on the risk assessment results of the current surgery, allocate differentiated monitoring resource budgets for different stages of the current surgery; Based on the monitoring resource budget and inherent risks of different surgical stages, the initial set of monitoring indicators, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold of each monitoring indicator are determined for each surgical stage. The initial set of monitoring indicators for each surgical stage includes a subset of core monitoring indicators and a subset of auxiliary monitoring indicators, and the initial sampling frequency of each monitoring indicator is positively correlated with the inherent risk of the corresponding surgical stage.

[0217] In one optional embodiment of this application, the step of dynamically adjusting the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information specifically includes: The current risk level and the doctor's current level of focus are determined based on the intraoperative multidimensional status information. Based on the current risk level and the doctor's current focus at the current surgical stage, at least one of the following should be adjusted: the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator.

[0218] In an optional embodiment of this application, the apparatus further includes a detection result delay processing module, which is used to perform the following steps: Determine the target monitoring indicator whose test results are delayed beyond a preset tolerance time; The historical trend data of the target monitoring indicator, the associated monitoring data of the target monitoring indicator, the current surgical stage information, and the current operation type information are input into a pre-established short-term prediction model of the monitoring indicator to obtain the current estimated value of the target monitoring indicator. Based on the confidence level of the current estimated value of the target monitoring indicator, determine whether to trigger a pre-alarm.

[0219] In an optional embodiment of this application, the detection result delay processing module is further configured to perform the following steps: The actual detection results of the target monitoring indicator are compared with the corresponding predicted values, and the short-term prediction model of the monitoring indicator is calibrated based on the comparison results.

[0220] In one optional embodiment of this application, the step of performing multi-level anomaly detection based on the monitoring data specifically includes: Based on the monitoring data, single-indicator real-time threshold comparison, multi-indicator correlation analysis, time window-based trend analysis, and machine learning model-based predictive detection are performed to determine the confidence score of intraoperative abnormalities. Based on the intraoperative abnormality confidence score, it is determined whether an intraoperative abnormality exists.

[0221] In one optional embodiment of this application, the formula for calculating the intraoperative abnormality confidence score is as follows: Score = α×S1+ β×S2+ γ×S3+ δ×S4; Wherein: S1 is the normalized score for the degree of anomaly of a single indicator, S2 is the consistency score for the correlation of multiple indicators, S3 is the score for the persistence of trends, and S4 is the confidence score for the prediction model; α, β, γ, and δ are dynamic weight coefficients that are adjusted according to the current surgical stage and risk level.

[0222] In one optional embodiment of this application, the step of outputting a graded alarm signal based on the abnormality type, severity level, and current surgical context specifically includes: Based on the type and severity of the abnormality, combined with the current stage of surgery, the patient's real-time condition, the doctor's focus, and the quality of the surgical field of vision, a comprehensive urgency score is calculated. The alarm level is determined based on the comprehensive urgency score, and the alarm sensory channel and information presentation density are adaptively selected based on the alarm level, the current surgical stage, and the doctor's focus, and a corresponding graded alarm signal is generated and output.

[0223] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. The processor 401 can call logical instructions in the memory 403 to execute an intraoperative anomaly identification method based on multimodal data. This method includes: acquiring intraoperative multidimensional state information in real time, and dynamically adjusting the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual anomaly threshold of each monitoring indicator. The multidimensional state information includes identification information of the current surgical stage, real-time physiological parameters of the patient, physician focus assessment information, and surgical field quality assessment information. Monitoring data is collected based on the dynamically adjusted monitoring parameters, and multi-level anomaly detection is performed based on the monitoring data. If an intraoperative anomaly is determined to exist, a graded alarm signal is output based on the anomaly type, severity level, and current surgical context.

[0224] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0225] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the intraoperative anomaly identification method based on multimodal data provided by the above methods. The method includes: acquiring intraoperative multidimensional state information in real time, and dynamically adjusting the monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual anomaly threshold of each monitoring indicator. The multidimensional state information includes current surgical stage identification information, real-time physiological parameters of the patient, physician focus assessment information, and surgical field quality assessment information. The method collects monitoring data based on the dynamically adjusted monitoring parameters and performs multi-level anomaly detection based on the monitoring data. When an intraoperative anomaly is determined to exist, the method outputs a graded alarm signal based on the anomaly type, severity level, and current surgical context.

[0226] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the intraoperative anomaly identification method based on multimodal data provided by the above methods. The method includes: acquiring intraoperative multidimensional state information in real time, and dynamically adjusting monitoring parameters of the current surgical stage based on the intraoperative multidimensional state information. The monitoring parameters include the actual set of monitoring indicators of the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual anomaly threshold of each monitoring indicator. The multidimensional state information includes current surgical stage identification information, real-time physiological parameters of the patient, physician focus assessment information, and surgical field quality assessment information. The method also includes collecting monitoring data based on the dynamically adjusted monitoring parameters and performing multi-level anomaly detection based on the monitoring data. If an intraoperative anomaly is determined to exist, the method outputs a graded alarm signal based on the anomaly type, severity level, and current surgical context.

[0227] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0228] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the intraoperative anomaly identification method based on multimodal data described in various embodiments or some parts of embodiments.

[0229] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intraoperative anomaly identification based on multimodal data, used for intraoperative anomaly identification in neurosurgical tumor resection, characterized in that, include: Real-time acquisition of intraoperative multidimensional status information, and dynamic adjustment of monitoring parameters for the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes current surgical stage identification information, real-time patient physiological parameters, physician focus assessment information, and surgical field quality assessment information. Monitoring data is collected based on dynamically adjusted monitoring parameters, and multi-level anomaly detection is performed based on the monitoring data; If an intraoperative abnormality is confirmed, a graded alarm signal is output based on the type of abnormality, severity level, and current surgical context.

2. The intraoperative anomaly identification method based on multimodal data according to claim 1, characterized in that, Before the surgery begins, the method also includes: Based on patient individual characteristic data, tumor characteristic data, and physician skill data, a personalized monitoring blueprint is generated for different surgical stages. The monitoring blueprint includes the initial set of monitoring indicators for each surgical stage, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold for each monitoring indicator.

3. The intraoperative anomaly identification method based on multimodal data according to claim 2, characterized in that, The process involves generating personalized monitoring blueprints for different surgical stages based on patient individual characteristic data, tumor characteristic data, and physician skill data, specifically including: Based on a pre-constructed multi-dimensional risk assessment model, the risk assessment result of the current surgery is determined. The input parameters of the multi-dimensional risk assessment model include patient baseline parameters, tumor parameters, and surgeon parameters. The patient baseline parameters include age, severity score of underlying disease, and coagulation function indicators. The tumor parameters include location risk classification, size, blood supply richness, and distance from important functional areas. The surgeon parameters include surgical proficiency index and recent complication rate. Based on the risk assessment results of the current surgery, allocate differentiated monitoring resource budgets for different stages of the current surgery; Based on the monitoring resource budget and inherent risks of different surgical stages, the initial set of monitoring indicators, the initial sampling frequency of each monitoring indicator, and the initial abnormal threshold of each monitoring indicator are determined for each surgical stage. The initial set of monitoring indicators for each surgical stage includes a subset of core monitoring indicators and a subset of auxiliary monitoring indicators, and the initial sampling frequency of each monitoring indicator is positively correlated with the inherent risk of the corresponding surgical stage.

4. The intraoperative anomaly identification method based on multimodal data according to claim 1, characterized in that, The dynamic adjustment of monitoring parameters for the current surgical stage based on the intraoperative multidimensional status information specifically includes: The current risk level and the doctor's current level of focus are determined based on the intraoperative multidimensional status information. Based on the current risk level and the doctor's current focus at the current surgical stage, at least one of the following should be adjusted: the actual set of monitoring indicators for the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator.

5. The intraoperative anomaly identification method based on multimodal data according to claim 1, characterized in that, In the process of acquiring intraoperative multidimensional status information in real time, the method further includes: Determine the target monitoring indicator whose test results are delayed beyond a preset tolerance time; The historical trend data of the target monitoring indicator, the associated monitoring data of the target monitoring indicator, the current surgical stage information, and the current operation type information are input into a pre-established short-term prediction model of the monitoring indicator to obtain the current estimated value of the target monitoring indicator. Based on the confidence level of the current estimated value of the target monitoring indicator, determine whether to trigger a pre-alarm.

6. The intraoperative anomaly identification method based on multimodal data according to claim 5, characterized in that, The method further includes: The actual detection results of the target monitoring indicator are compared with the corresponding predicted values, and the short-term prediction model of the monitoring indicator is calibrated based on the comparison results.

7. The intraoperative anomaly identification method based on multimodal data according to claim 1, characterized in that, The multi-level anomaly detection based on the monitoring data specifically includes: Based on the monitoring data, single-indicator real-time threshold comparison, multi-indicator correlation analysis, time window-based trend analysis, and machine learning model-based predictive detection are performed to determine the confidence score of intraoperative abnormalities. Based on the intraoperative abnormality confidence score, it is determined whether an intraoperative abnormality exists.

8. The intraoperative anomaly identification method based on multimodal data according to claim 7, characterized in that, The formula for calculating the confidence score of intraoperative abnormalities is as follows: Score = α×S1+ β×S2+ γ×S3+ δ×S4; Wherein: S1 is the normalized score for the degree of abnormality of a single indicator, S2 is the consistency score for the association of multiple indicators, S3 is the score for the persistence of trends, and S4 is the confidence score for the prediction model; α, β, γ, and δ are dynamic weight coefficients that are adjusted according to the current surgical stage and risk level.

9. The intraoperative anomaly identification method based on multimodal data according to claim 1, characterized in that, The system outputs graded alarm signals based on the abnormality type, severity level, and current surgical context, specifically including: Based on the type and severity of the abnormality, combined with the current stage of surgery, the patient's real-time condition, the doctor's focus, and the quality of the surgical field of vision, a comprehensive urgency score is calculated. The alarm level is determined based on the comprehensive urgency score, and the alarm sensory channel and information presentation density are adaptively selected based on the alarm level, the current surgical stage, and the doctor's focus, and a corresponding graded alarm signal is generated and output.

10. A device for intraoperative anomaly identification based on multimodal data, used for intraoperative anomaly identification in neurosurgical tumor resection, characterized in that, include: The monitoring parameter dynamic adjustment module is used to acquire intraoperative multidimensional status information in real time and dynamically adjust the monitoring parameters of the current surgical stage based on the intraoperative multidimensional status information. The monitoring parameters include the actual set of monitoring indicators of the current surgical stage, the actual sampling frequency of each monitoring indicator, and the actual abnormal threshold of each monitoring indicator. The multidimensional status information includes current surgical stage identification information, real-time patient physiological parameters, physician focus assessment information, and surgical field quality assessment information. An anomaly detection module is used to collect monitoring data based on dynamically adjusted monitoring parameters and to perform multi-level anomaly detection based on the monitoring data. The alarm module is used to output a graded alarm signal based on the type of abnormality, severity level, and current surgical context when an intraoperative abnormality is determined to exist.

Images