Intelligent trigger matching and closed-loop quality control method and system for brain tumor intraoperative wake-up decision

Abstract

This application discloses an intelligent triggering, matching, and closed-loop quality control system and method for intraoperative awakening decision-making in brain tumor surgery, relating to the fields of medical artificial intelligence and clinical decision support technology. The system comprises five modules: data acquisition, feature extraction, disease matching, report generation, and closed-loop feedback. It achieves full-process intelligence through two-layer rule-based screening of target patients, multi-dimensional weighted similarity matching of historical cases, generation of clinical decision reports using a large language model, and closed-loop data accumulation at multiple postoperative time points. The method operates according to the system's execution, with physician quality control links at each key node. This invention solves the problems of traditional decision-making relying on experience, data fragmentation, and lack of closed-loop feedback, improving the objectivity and accuracy of intraoperative awakening decisions, constructing a human-machine collaborative quality control system, and enabling the system's autonomous learning and evolution.

Description

Technical Field

[0001] This application relates to the fields of medical artificial intelligence and clinical decision support technology, specifically to an intelligent triggering, matching, and closed-loop quality control system and method for intraoperative awakening decision-making in brain tumor surgery. Background Technology

[0002] In functional brain tumor surgery, intraoperative awake anesthesia is a crucial means of balancing tumor resection extent with neurological function preservation, and the quality of its decision-making directly affects the patient's postoperative neurological integrity and surgical efficacy. Current clinical practice generally relies on the individual experience of physicians to select intraoperative awakening strategies, lacking an objective and standardized decision support system, resulting in systemic flaws in the decision-making process. Specifically: The lack of a quantitatively driven automated triggering mechanism in patient screening leads to delayed clinical event responses that are susceptible to subjective interference, resulting in the omission of potentially beneficial patients or the mis-inclusion of unsuitable patients, causing misallocation of medical resources and irregularities in process initiation. Multiple sources of clinical data are scattered across heterogeneous platforms such as hospital information systems, electronic medical record systems, and image archiving and communication systems. Quantitative analysis of neuroimaging requires manual intervention, and neurological functional status information is deeply buried in unstructured text medical records. Physicians struggle to efficiently integrate accurate imaging measurement data with standardized neurological function scores, making it impossible to construct a complete patient assessment profile. Historical case matching lacks multi-dimensional deep similarity calculation capabilities, preventing physicians from obtaining real-time prognostic comparison data between the awake and non-awake groups that highly match the target patients, resulting in a lack of real-world evidence to support the decision-making process. Postoperative key outcome data, such as the extent of tumor resection and neurological function recovery indicators, are stored in isolated, unstructured form, disconnected from the preoperative decision-making process, creating a decision-outcome information gap and preventing valuable clinical experience from being transformed into reusable knowledge assets.

[0003] Existing technical solutions only focus on optimizing a single link. Independent medical image analysis tools cannot be linked to medical record text information. Keyword-based medical record retrieval systems lack multimodal data fusion capabilities. Clinical databases only support simple case queries and cannot achieve intelligent linkage of the entire decision-making process. As a result, intraoperative awakening decisions are always in a dilemma of data fragmentation, lack of evidence, and failure of feedback, which seriously restricts the objectivity and accuracy of clinical decisions. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent triggering, matching, and closed-loop quality control system and method for intraoperative awakening decision-making in brain tumor surgery. It has the advantages of automatically triggering patient screening, efficiently integrating multi-source clinical data, accurately matching historical cases, generating intelligent decision reports, and realizing closed-loop feedback, thereby significantly improving the objectivity, accuracy, and efficiency of intraoperative awakening decision-making.

[0005] This application provides an intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decisions in brain tumor surgery, including: The data acquisition module, serving as the system process initiation and data hub, is deeply integrated with the hospital information system, electronic medical record system, and image archiving communication system. It is used for automatic screening of target patients and parallel acquisition of multi-source clinical data. The feature extraction module is used to standardize and extract features from multi-source clinical data, and output structured feature vectors and standardized neurological function assessment information. The disease matching module, as the core intelligent decision engine of the system, is used to build a structured historical case knowledge base, perform deep similarity matching between target patients and historical cases, and output prognostic comparison data. The report generation module is used to generate clinical decision reports based on prognostic comparison data using a large language model, and to optimize the clinical decision reports and iterate the large language model based on physician feedback. The closed-loop feedback module, as the central hub for system learning and evolution, is used to automatically collect, correlate and integrate postoperative multi-time point outcome data, and deposit the integrated data into a structured historical case knowledge base to form a closed-loop feedback between decision-making and results.

[0006] Furthermore, the data acquisition module has a built-in logically related two-layer rule filtering engine, which includes a surgical indication screening layer and a patient physiological and cooperation screening layer. The surgical indication screening layer is used to automatically retrieve the corresponding surgical request form in response to the clinical event submitted by the anesthesia assessment form. Based on the ICD-9-CM-3 surgical operation code and the ICD-10 disease diagnosis code, it selects an initial cohort of patients who meet the criteria for Class 01 craniotomy, brain and meningotomy and C71 brain malignant tumor diagnosis. The patient physiological and cooperation screening layer is used to analyze the preoperative anesthesia assessment form based on the initial patient cohort and screen out target patients who are conscious and have excellent doctor-patient cooperation assessment. The data acquisition module is also used to concurrently trigger multi-source data acquisition tasks after locking onto the target patient, and to simultaneously capture the target patient's recent surgical date head multimodal MRI image data from the image archiving communication system and the current hospitalization unstructured text medical record data from the electronic medical record system.

[0007] Furthermore, the feature extraction module includes an image intelligent quantization and feature extraction submodule and a text structured parsing and information extraction submodule; The intelligent image quantization and feature extraction submodule is used to build a fully automated standardized processing pipeline for preoperative and postoperative cranial MRI images. It sequentially performs spatial registration and standardization of multi-sequence images, image segmentation based on deep learning, automatic calculation of three-dimensional volume and spatial coordinate annotation, quantitative output and visualization rendering of key image biometric indicators, and also supports doctors to review and fine-tune through a dedicated quality control interface, outputting structured feature vectors. The text structuring parsing and information extraction submodule is used to introduce a large language model to perform deep semantic understanding and key information extraction on unstructured text medical records. Following the principle of the last record closest to the surgery on the timeline, it extracts items for neurological examinations, maps and fills the extracted information into a preset postoperative neurological assessment single structured data template to generate an assessment draft, and pushes a review reminder to the doctor before the surgery time node.

[0008] Furthermore, the disease matching module is used to build a structured historical case knowledge base that includes basic clinical information of patients, intraoperative awakening implementation tags, and key tags for various postoperative prognostic outcomes; The disease matching module is also used to perform refined matching calculations between target patients and historical cases based on a preset multi-dimensional weighted similarity algorithm. The quantitative evaluation formula for the multi-dimensional weighted similarity algorithm is as follows: Disease similarity score = Σ(feature similarity × weight); The disease matching module also has a matching output mechanism, which is used to calculate the similarity score between the target patient and the intraoperative awakening group and the non-intraoperative awakening group in the historical database, respectively, and to filter the high similarity case subset according to the preset threshold and output the corresponding structured prognostic statistical indicators.

[0009] Furthermore, the core feature terms and weight configurations of the multi-dimensional weighted similarity algorithm are as follows: The weightings for tumor location (40%), preoperative tumor volume (25%), age group (10%), gender (5%), motor function (5%), language function (5%), sensory function (5%), and higher cortical function (5%) were as follows: Motor function was assessed using the MRC muscle strength grading system, language function was assessed using the BNT naming test, sensory function was assessed using the light touch / pinprick sensation rating system, and higher cortical function was assessed using the MoCA rating system. The preset threshold is ≥0.6. The disease matching module is used to sort the subset of highly similar cases in descending order of similarity.

[0010] Furthermore, the structured prognostic statistical indicators include four categories: case size statistics, baseline characteristic distribution, surgical outcome indicators, and functional prognostic comparison. Case size statistics include the number of matched cases and their proportion in the corresponding historical groups. Baseline characteristic distribution includes tumor location, preoperative tumor volume, age group, and gender composition. Surgical outcome indicators include tumor resection rate, fiber bundle injury rate, and total volume of new edema / damage area. Functional prognostic comparison includes quantitative comparison values ​​of motor, language, sensory, and higher cortical functions before and after surgery.

[0011] Furthermore, the report generation module is used to receive and integrate the comparison data of the target patient and the high similarity intraoperative awakening group and non-intraoperative awakening group output by the disease matching module, and inject clinical analysis task instructions into the large language model through prompt word engineering; The report generation module is also used to drive the large language model to generate a comparative analysis report on intraoperative awakening strategies and neurological prognosis, which includes visual comparison tables, in-depth data interpretation and clinical inference, and targeted anesthesia clinical recommendations. The report generation module features an interactive interface with two-way selection buttons for adoption and regeneration. It also automatically records interactive data and calculates the clinical adoption accuracy of the large model output. The formula for calculating the clinical adoption accuracy is: Adoption accuracy rate = (Number of adoptions / (Number of adoptions + Number of regenerations)) × 100%.

[0012] Furthermore, the regenerated options are configured with preset feedback reason categories, including factual errors, logical flaws, key omissions, unfounded fabrications, deviation from instructions, inoperability, and others; The report generation module is used to use the recorded interaction data as a key basis for iterative fine-tuning of the prompt word strategy or large language model, driving the continuous evolution of the system's decision support capabilities.

[0013] Furthermore, the closed-loop feedback module is used to automatically start a multi-threaded data acquisition and processing process after the operation, sequentially executing the steps of demographic and key decision information collection, automatic quantification of multi-time point imaging prognostic indicators, structured acquisition of neurological function prognostic data, data association and integration and knowledge base update, and forming a self-evolving decision support closed loop. When the closed-loop feedback module performs the automatic quantification step of multi-time point radiological prognostic indicators, it calls the fully automatic standardization processing pipeline of the image intelligent quantization and feature extraction submodule.

[0014] Furthermore, the automated quantification step of multi-time point imaging prognostic indicators is used to retrieve the patient's first follow-up cranial MRI images from the image archiving communication system at preset key follow-up time points after surgery. The automated processing pipeline accurately calculates the residual tumor volume, the volume of newly developed FLAIR high signal area, and the fiber bundle damage rate after surgery. The key follow-up time points include postoperative days 1-3, 7-14, and 28-40. The structured data collection process for neurological prognosis is set up with a T+1 reminder rule. This rule sends an update reminder to the attending physician on the first day after the task is triggered at each preset postoperative assessment time point. The postoperative neurological assessment form, which has been reviewed and confirmed before the operation, is used as a benchmark to guide the physician to complete the update and confirmation of the postoperative assessment form. The system also automatically calculates the changes in various neurological function scores relative to the preoperative baseline.

[0015] Furthermore, data association and integration, and knowledge base updates include: The collected and calculated postoperative data is automatically linked and integrated with the target patient's complete preoperative feature dataset and intraoperative decision-making process record. The postoperative data includes demographic information, intraoperative decision labels, multi-time point imaging indicators, and multi-time point neurological function scores. The closed-loop feedback module is used to store the integrated complete case data with decision-outcome labels in a highly structured form into a structured historical case knowledge base, continuously enriching and updating the knowledge base of similarity matching, and improving the matching accuracy and recommendation reliability of the disease matching module.

[0016] This application also proposes an intelligent triggering, matching, and closed-loop quality control method for intraoperative awakening decisions in brain tumor surgery. The method, implemented by the aforementioned intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decisions in brain tumor surgery, includes the following steps: Step S1: In response to the clinical event of the anesthesia assessment form submission, the two-layer rule screening engine automatically screens and locks in target patients who meet the requirements of surgical indications and physiological compliance. Step S2: The intelligent image quantization and feature extraction submodule is started in parallel to process MRI images and extract quantization features; the text structure parsing and information extraction submodule is started to process text medical records and extract structured physical examination information. Step S3: Calculate the disease similarity score between the target patient and historical cases based on the multi-dimensional weighted similarity algorithm, and output the high similarity control cases in the intraoperative awakening group and the non-intraoperative awakening group, as well as their structured prognostic statistics. Step S4: Use a large language model to transform the matching data into a comparative analysis report of intraoperative awakening strategies and neurological function prognosis, which includes a visual comparison table, in-depth interpretation, and clinical suggestions. Record the doctor's adoption or regenerate feedback and calculate the clinical adoption accuracy. Step S5: At multiple preset follow-up time points after surgery, automatically collect the patient's follow-up MRI images and clinical assessment data of neurological function, call the image processing pipeline to calculate key postoperative imaging indicators, and calculate the change value of neurological function score relative to the preoperative baseline. Step S6: Link and integrate postoperative outcome data with intraoperative decision tags with preoperative feature data and intraoperative decision records, and deposit the integrated complete case data into a structured historical case knowledge base to update and enrich the knowledge base of similarity matching.

[0017] Furthermore, in step S1, the dual-layer rule-based screening engine first completes the surgical indication screening based on ICD-9-CM-3 and ICD-10 codes, and then completes the patient's physiological and cooperation screening; in step S2, after preprocessing, the doctor reviews and confirms the image quantification results and physical examination information draft, forming the first quality control node in the entire process; in step S3, after the similarity score is calculated, a subset of high-similarity cases is selected based on a preset threshold of ≥0.6 and sorted in descending order of similarity; in step S4, if the doctor chooses to regenerate the report, the feedback reason is selected from the preset categories, and the system uses the feedback data for iterative fine-tuning of the large language model; in step S5, the postoperative data collection follows the T+1 reminder rule, and a postoperative quality control node is formed after the image index calculation and neurological function score statistics; in step S6, the knowledge base update is an incremental update, and the newly accumulated case data enables the system to form a self-learning decision support closed loop, and the method sets doctor review and correction links at each key node of image measurement, physical examination extraction, report review, and postoperative assessment, and records the entire process operation log to achieve traceable quality control.

[0018] This application has the following beneficial effects: Through the above implementation methods, this application effectively solves the problems of data accuracy, decision reliability, and continuous system optimization in the intraoperative awakening decision-making process for brain tumor surgery. By introducing a tiered screening order in the patient screening stage, screening efficiency and accuracy are improved. Physician review and quality control nodes are set up in the data preprocessing and postoperative data collection stages, ensuring data quality and reliability from the source and end, avoiding decision bias caused by data errors. The refined screening and sorting mechanism in the disease matching stage enables physicians to obtain high-value reference information more efficiently. The structured feedback mechanism introduced in the report generation module provides a clear direction for the continuous iterative fine-tuning of the large language model, significantly improving the clinical usability and accuracy of its generated reports. The incremental updates and self-learning capabilities of the knowledge base enable the system to continuously evolve with the accumulation of clinical practice, continuously improving the accuracy of decision support. Furthermore, setting up physician review and correction links at each key node and recording operation logs constructs a comprehensive human-machine collaborative quality control system, which not only improves the safety of decision-making but also provides a basis for clinical responsibility traceability, thereby comprehensively improving the intelligence, accuracy, and safety level of intraoperative awakening decision-making for brain tumor surgery. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of an intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery, provided in an embodiment of this application.

[0020] Figure 2 This is a flowchart illustrating an intelligent triggering, matching, and closed-loop quality control method for intraoperative awakening decision-making in brain tumor surgery, provided as an embodiment of this application. Detailed Implementation

[0021] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Traditional high-pressure bushing oil arc discharge testing devices cannot comprehensively assess the arc withstand capability of monitoring sensors when simulating different arc energies affecting the sensors. Existing devices also suffer from inconsistencies in material selection, sensor type compatibility, visualization, and high-temperature resistance, and lack unified calibration, data acquisition, and judgment standards, resulting in poor repeatability of test results and making it difficult to provide reliable sensor selection criteria.

[0024] Firstly, such as Figure 1 As shown, this application proposes an intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decisions in brain tumor surgery.

[0025] The system includes a data acquisition module. This module serves as the system's workflow initiator and data hub, deeply integrating with the hospital information system, electronic medical record system, and image archiving communication system for automatic screening of target patients and parallel acquisition of multi-source clinical data. For example, this module can be configured to periodically scan newly admitted patient records in the hospital information system, using manually set rules such as patient diagnosis codes or surgical types to initially identify patients who may require intraoperative awakening decisions. For parallel acquisition of multi-source clinical data, the module can be configured to automatically request a copy of the patient's text medical record from the electronic medical record system and request the patient's image data from the image archiving communication system after identifying a potential target patient via a preset interface or script. This data can be acquired at different times to avoid excessive system load.

[0026] The system also includes a feature extraction module. This module performs standardization processing and feature extraction on the multi-source clinical data, outputting structured feature vectors and standardized neurological function assessment information. For example, for image data obtained from an image archiving communication system, the feature extraction module can use traditional image processing algorithms, such as edge detection and threshold segmentation, to initially delineate the tumor region, and then manually measure its size and location to extract image features. For unstructured text medical record data obtained from an electronic medical record system, the module can use methods such as keyword matching or regular expressions to identify descriptions related to neurological function from the text, such as good limb movement and clear speech, and then manually convert them into preset standardized scores to form standardized neurological function assessment information. These extracted features and assessment information are then integrated into a structured feature vector.

[0027] The system further includes a disease matching module. This module, serving as the core intelligent decision engine, constructs a structured historical case knowledge base, performs deep similarity matching between the target patient and historical cases, and outputs prognostic comparison data. For example, this module can pre-build a database containing a large amount of historical brain tumor patient data. Each case in this database includes the patient's basic information, tumor characteristics, treatment plan, and postoperative outcome. When new target patient data is input, the module can use simple distance metrics such as Euclidean distance or cosine similarity to calculate the similarity between the target patient and each historical case in the knowledge base on several key features, such as age, tumor size, and tumor location. Then, based on the calculated similarity, the module selects the group of historical cases most similar to the target patient and statistically analyzes the postoperative prognoses of these cases with and without intraoperative awakening, generating prognostic comparison data.

[0028] The system also includes a report generation module. This module generates clinical decision reports using a large language model based on the prognostic comparison data, and optimizes the report and iterates the large language model based on physician feedback. For example, the module can receive prognostic comparison data from the disease matching module as input, combine it with a pre-set simple template, and generate a preliminary text report using the large language model. This report may include a brief description of the target patient's characteristics, prognostic statistics of similar historical cases, and a preliminary decision-making tendency based on these results. After reviewing the report, physicians can provide feedback verbally or in writing, such as pointing out unclear points or suggesting additional information. The system can collect this feedback and allow manual adjustments to the large language model's prompts or parameters to better meet physicians' needs in subsequent report generation.

[0029] In addition, the system includes a closed-loop feedback module. This module serves as the system's learning and evolution hub, automatically collecting and integrating postoperative multi-timepoint outcome data. The integrated data is then stored in a structured historical case knowledge base, forming a closed-loop feedback mechanism for decision-making and outcomes. For example, this module can be configured to retrieve postoperative follow-up imaging reports, neurological function assessment results, and other outcome data at several preset postoperative time points through manual input or simple database queries. This data can then be easily correlated with the patient's preoperative characteristic data and intraoperative decision records, such as by matching patient ID. The integrated and correlated complete case data, including preoperative, intraoperative, and postoperative information, can be added to the structured historical case knowledge base, continuously expanding its size and providing more references for future disease matching.

[0030] Compared to traditional decision-making methods that rely on doctors' subjective judgment for patient screening and process initiation, this system, through its data acquisition module's automatic screening mechanism, can proactively identify potential target patients and trigger subsequent processes based on preset rules. For example, in User A's case, the system can automatically respond to the anesthesia assessment form submission event, rather than waiting for the doctor to manually identify the patient. This effectively avoids situations where potentially beneficial patients are overlooked or unsuitable patients are included in the assessment, improving the objectivity and efficiency of decision-making and solving the problem of traditional methods lacking objective basis for decision triggering.

[0031] In terms of data processing, existing technologies scatter preoperative assessment data across different information systems, and quantitative analysis of neuroimaging and extraction of neurofunctional status information rely on manual intervention, resulting in low data integration efficiency. This system achieves automated and efficient data integration through a data acquisition module that performs parallel acquisition of multi-source clinical data and a feature extraction module that performs standardized processing and feature extraction on this data. For example, in User A's case, the system can acquire MRI images and text medical records in parallel, and through manually assisted image processing and keyword matching, transform the raw data into structured feature vectors and standardized neurofunctional assessment information, greatly reducing the burden on doctors to manually integrate and analyze data and solving the problem of low data integration efficiency.

[0032] To address the lack of real-world evidence in traditional decision-making, this system constructs a structured historical case knowledge base through a disease matching module and performs deep similarity matching, providing strong scientific evidence for clinical decision-making. In User A's case, the system can quickly identify a group of patients with highly similar conditions to User A from the vast historical case database through simple similarity calculations, and provide comparative prognostic data for these similar cases under different treatment options. This allows doctors to move beyond relying solely on personal experience when making decisions, providing quantitative references based on real-world data, significantly improving the scientific rigor and accuracy of decision-making, and solving the problem of insufficient real-world evidence for decision-making references.

[0033] Furthermore, traditional methods fail to directly link postoperative outcome data with preoperative decisions, hindering the systematic accumulation of valuable clinical experience. This system addresses this by automatically collecting, linking, and integrating postoperative outcome data from multiple time points through a closed-loop feedback module, and then storing it in a structured historical case knowledge base, thus forming a closed loop between decision-making and outcome. In User A's case, User A's postoperative recovery and outcome data are automatically collected and integrated back into the knowledge base, enabling the system to continuously learn and evolve. This mechanism ensures the systematic accumulation and feedback of clinical experience, continuously optimizing the system's decision support capabilities and resolving the problem of ineffective postoperative knowledge accumulation.

[0034] Finally, this system organically integrates multiple stages, including data collection, feature extraction, disease matching, report generation, and closed-loop feedback, forming a collaborative and intelligent solution across the entire process. This overcomes the limitations of existing technologies that can only solve problems in isolated stages, realizing a complete workflow from patient screening to decision generation and effect feedback, ensuring the objectivity, accuracy, and sustainable optimization of intraoperative awakening decisions for brain tumors.

[0035] In some implementations, the data acquisition module incorporates a logically linked two-layer rule-based filtering engine, comprising a surgical indication screening layer and a patient physiological and cooperation screening layer. The surgical indication screening layer, in response to clinical events submitted with an anesthesia assessment form, automatically retrieves the corresponding surgical request form and, based on ICD-9-CM-3 surgical operation codes and ICD-10 disease diagnosis codes, selects an initial patient cohort meeting the criteria for Class 01 craniotomy, cerebral, and meningotomy, and the diagnosis of C71 malignant brain tumors. The patient physiological and cooperation screening layer, based on the initial patient cohort, analyzes the preoperative anesthesia assessment form to select target patients who are conscious and have excellent doctor-patient cooperation. The data acquisition module, after identifying the target patient, concurrently triggers a multi-source data acquisition task, simultaneously capturing recent surgical date cranial multimodal MRI image data from the image archiving communication system and unstructured text medical record data from the current hospitalization from the electronic medical record system.

[0036] The two-layer rule-based screening engine is a phased, logically progressive automated screening mechanism designed to progressively screen patients using pre-defined rules. Its role is to improve the accuracy and efficiency of patient screening, ensuring that only patients meeting specific criteria proceed to subsequent data processing and decision-making. This engine can be implemented as a standalone software module, operating by configuring rule sets; alternatively, it can be integrated into the data acquisition module as one of its core functions, implementing rule judgment and execution through programming logic. The surgical indication screening layer is the first hurdle of the two-layer rule-based screening engine, primarily used to initially identify patients who may meet the surgical indications from a broad patient population. Its function is to conduct preliminary, large-scale screening based on medical coding standards, quickly eliminating irrelevant cases. This layer can execute structured query language (SQL) queries and match ICD codes through a database interface with the Hospital Information System (HIS) or Electronic Medical Record System (EMR); alternatively, it can utilize Natural Language Processing (NLP) technology to perform semantic analysis on the unstructured text in the surgical request form, extracting key information and comparing it with pre-defined indication rules. The patient physiological and cooperation screening layer is the second layer of the dual-rule screening engine, further refining the patient screening process based on the surgical indication screening layer. Its role is to ensure that patients not only meet the surgical indications but also possess the physiological conditions and cooperation capabilities required for intraoperative awake surgery. This layer can be used to analyze structured anesthesia assessment data from the electronic medical record system to extract key indicators such as consciousness level scores and cooperation assessment results; alternatively, it can combine machine learning models to comprehensively evaluate the patient's physiological indicators and behavioral descriptions to predict their cooperation level. Concurrent triggering of multi-source data acquisition tasks is a data acquisition strategy where, after identifying the target patient, the system can simultaneously initiate multiple data acquisition processes to obtain the required information from different data sources. Its purpose is to improve the efficiency and completeness of data acquisition, shorten data preparation time, and provide timely and comprehensive data support for subsequent feature extraction and decision-making. This can be achieved through asynchronous programming models, such as using message queues or multi-threading / multi-process technology to simultaneously send data requests to the image archiving communication system and the electronic medical record system; alternatively, it can adopt a microservice architecture, encapsulating different data acquisition tasks as independent services and using a coordinator for concurrent invocation. Parallel retrieval of recent surgery date multimodal cranial MRI images from the image archiving communication system and unstructured text medical record data from the current hospitalization from the electronic medical record system refers to the system's ability to simultaneously acquire specific types of data from different medical information systems. Its purpose is to ensure the diverse and timely nature of the acquired data, providing a foundation for a comprehensive assessment of the patient's condition.Imaging data can include MRI images of various sequences such as T1, T2, FLAIR, and DWI, which are acquired through the PACS interface; unstructured text medical record data can include progress notes, ward round records, consultation records, etc., which are acquired through the EMR interface and may require preliminary text parsing to identify relevant documents.

[0037] For example, when a doctor submits an anesthesia assessment form in the hospital information system, the system's data acquisition module immediately responds to this clinical event. The built-in two-layer rule-based filtering engine then activates. First, the surgical indication screening layer automatically queries surgical request forms associated with the anesthesia assessment form by calling the API interface of the hospital information system database. This layer executes a pre-defined SQL query, for example, searching for patient records where the surgical operation code belongs to the 01.xx series (representing craniocerebral and meningiotomy and resection) and the disease diagnosis code is C71.x (representing malignant brain tumors), thereby generating an initial patient queue. Next, the patient physiological and cooperation screening layer traverses this initial queue and, for each patient, parses the preoperative anesthesia assessment form stored in their electronic medical record system. This layer can use regular expressions or pre-trained natural language processing models to extract key fields such as state of consciousness and doctor-patient cooperation assessment from unstructured text. For example, if the state of consciousness field shows "awake" and the doctor-patient cooperation assessment is "excellent," then the patient is identified as the target patient. Once a target patient is identified, the data acquisition module simultaneously sends data requests to both the image archiving communication system and the electronic medical record system via a message queue mechanism. For example, it sends a request to the PACS to retrieve all cranial MRI sequences (such as T1WI, T2WI, FLAIR, and DWI) performed on the patient within the past week; simultaneously, it sends another request to the EMR to retrieve all unstructured text data, including progress notes, ward round records, and consultation records, from the patient's current hospitalization. These data requests are executed concurrently to ensure that all necessary information is obtained in the shortest possible time.

[0038] The above implementation method enables precise and efficient screening of patients for intraoperative awakening decisions in brain tumor surgery. The introduction of a two-layer rule-based screening engine allows the system to quickly identify a preliminary cohort of patients meeting surgical indications from massive amounts of patient data. Subsequently, it further refines the screening to identify target patients with good physiological conditions and cooperation, significantly improving the accuracy and relevance of patient screening. This phased screening mechanism effectively avoids including unsuitable patients in subsequent complex decision-making processes, thereby reducing invalid data processing and improving the overall system efficiency. Furthermore, after identifying target patients, concurrently triggered multi-source data acquisition tasks can acquire patient image and text medical record data in a timely and comprehensive manner, ensuring the integrity and timeliness of the data required for subsequent feature extraction and intelligent decision-making. This lays a solid foundation for generating high-quality clinical decision reports, thereby enhancing the scientific rigor and safety of intraoperative awakening decisions.

[0039] In some implementations, the feature extraction module includes an image intelligent quantization and feature extraction submodule and a text structured parsing and information extraction submodule. The image intelligent quantization and feature extraction submodule is used to construct a fully automated standardized processing pipeline for preoperative and postoperative cranial MRI images. It sequentially performs spatial registration and standardization of multi-sequence images, image segmentation based on deep learning, automatic calculation of three-dimensional volume and spatial coordinate annotation, quantitative output and visualization rendering of key image biometric indicators, and supports doctors to review and fine-tune through a dedicated quality control interface, outputting structured feature vectors. The text structured parsing and information extraction submodule is used to introduce a large language model to perform deep semantic understanding and key information extraction on unstructured text medical records. It extracts neurological examination items according to the principle of the last record closest to the surgery on the timeline, maps and fills the extracted information into a preset postoperative neurological assessment single structured data template to generate an assessment draft, and pushes a review reminder to the doctor before the surgery time node.

[0040] The intelligent image quantification and feature extraction submodule is a component specifically designed for processing medical image data. Its core objective is to transform raw, complex MRI images into structured, quantifiable biometric indicators. This submodule ensures standardized and efficient image data processing through automated workflows, providing accurate image features for subsequent disease matching and decision-making. Implementation can include integrating open-source medical image processing libraries (e.g., the Python-based SimpleITK library or ITK-SNAP), utilizing their algorithms for image reading, preprocessing, registration, segmentation, and quantization; or leveraging cloud computing platforms to process large-scale image data using distributed computing resources, combined with specialized medical imaging workstations for visualization and physician review.

[0041] Constructing a fully automated, standardized pipeline for preoperative and postoperative cranial MRI image processing refers to a predefined, automated image processing workflow that requires no human intervention. The aim is to ensure that each image processing session follows the same standards and procedures, thereby improving processing efficiency and consistency and reducing human error. This pipeline can be implemented by writing scripting languages ​​(e.g., Python) to chain together different image processing algorithms and tools to form an automated workflow; alternatively, it can utilize specialized medical image processing platforms or software, configuring their built-in automated processing templates or macros.

[0042] Spatial registration and normalization of multi-sequence images refers to the spatial alignment of MRI images from different sequences (e.g., T1, T2, FLAIR, etc.) and the normalization of intensity or contrast. This step aims to eliminate differences caused by different scanning parameters, equipment, or time points, ensuring the accuracy of subsequent analysis. Spatial registration can employ image registration algorithms based on mutual information and normalized cross-correlation, combined with affine transformation or non-rigid deformation models for spatial alignment; intensity normalization can be achieved through methods such as histogram matching, Z-score normalization, or white matter normalization.

[0043] Deep learning-based image segmentation utilizes deep neural network models (e.g., U-Net, V-Net) to automatically identify and separate specific regions (e.g., tumors, edema, important functional areas) in images. This technology improves segmentation accuracy and automation, reducing the workload of manual delineation. Implementation methods can include training a convolutional neural network (CNN) model using a large amount of labeled medical image data to automatically identify and segment target regions; alternatively, pre-trained deep learning models can be used, combined with transfer learning or fine-tuning techniques, to adapt to specific segmentation tasks.

[0044] Automatic 3D volume calculation and spatial coordinate annotation involves automatically calculating the 3D volume of the segmented region after image segmentation and annotating its position in standard anatomical space. This provides quantified information such as tumor size, extent of edema, and its relative position to important brain regions (e.g., motor cortex, language area). Volume calculation can be achieved by multiplying the number of pixels / voxels in the segmented region by the actual physical size of a single pixel / voxel; spatial coordinate annotation can be achieved by mapping the segmentation results to a standard brain atlas (e.g., MNI space).

[0045] The quantitative output and visualization of key imaging biometric indicators involves outputting calculated imaging features such as volume, location, and morphology in numerical form and displaying them intuitively through a graphical interface. This facilitates doctors' rapid understanding of imaging features and supports clinical decision-making. Quantitative output can use structured data formats such as CSV and JSON; visualization can be implemented in a browser using 3D reconstruction software (e.g., VTK, ParaView) or web front-end technologies (e.g., WebGL).

[0046] A dedicated quality control interface supports physician review and fine-tuning. It provides an interactive user interface that allows physicians to review, correct, and confirm the results of automated processing. This ensures the accuracy of the automated processing results, and combined with physician experience for final confirmation, improves system reliability. The interface offers functions such as image overlay, manual highlighting, and parameter adjustment; alternatively, it can be integrated into a PACS system or a standalone medical imaging workstation.

[0047] The text structuring parsing and information extraction submodule is a component specifically designed for processing unstructured medical records. Its purpose is to extract key clinical information and transform it into structured data. Its function is to convert free text recorded by doctors into standardized information that can be understood and analyzed by machines. This can be achieved using methods based on rule matching, regular expressions, and natural language processing (NLP) techniques; alternatively, pre-trained language models (e.g., BERT, GPT series) can be used for tasks such as named entity recognition (NER) and relation extraction.

[0048] Introducing large language models for deep semantic understanding and key information extraction from unstructured medical records leverages the powerful semantic understanding capabilities of large pre-trained language models (LLMs) to identify, extract, and summarize useful clinical information from complex medical texts. This overcomes the limitations of traditional NLP methods in handling the complexity, diversity, and ambiguity of medical texts, improving the accuracy and comprehensiveness of information extraction. Specific prompts can be designed to guide the large language model in identifying key entities and relationships in medical records, such as disease diagnoses, symptom descriptions, signs, and treatment plans; alternatively, the large language model can be fine-tuned using medical domain data to better adapt it to the characteristics and terminology of medical texts.

[0049] Extracting neurological examination entries based on the principle of selecting the most recent record from the timeline closest to the surgery means choosing the most relevant neurological examination record from the patient's multiple medical records. This ensures that the extracted information is up-to-date and best reflects the patient's preoperative condition. This can be achieved by analyzing the timestamps of the medical records and combining them with the surgery time for sorting and filtering; alternatively, the temporal reasoning capabilities of large language models can be utilized to identify and extract the most relevant last record.

[0050] The process of mapping extracted information to a pre-defined structured data template for postoperative neurological assessment to generate an assessment draft involves organizing and filling in key information extracted from text according to a predefined structured template. This aims to transform unstructured information into a standardized data format that is easy to process and analyze later. The structured data template can be constructed by defining a JSON Schema or XML Schema, and a program can be written to map the extracted entities to the corresponding fields in the template; alternatively, the generative capabilities of a large language model can be utilized to directly generate text or data structures conforming to the template format based on the extracted information.

[0051] Sending a review reminder to the doctor before the surgery timeframe means that the system automatically pushes the generated draft assessment to the doctor before the surgery begins, reminding them to review it. This ensures the accuracy and timeliness of the preoperative assessment, providing the doctor with a basis for decision-making. Messages can be sent via the hospital information system (HIS) notification module, email, or mobile application; alternatively, prominent to-do reminders can be set on the system interface.

[0052] For example, in the intelligent image quantization and feature extraction submodule, the SimpleITK library based on Python can be used for reading and preprocessing MRI image data. Spatial registration can be performed using ANTs (Advanced Normalization Tools) software to align T1, T2, and FLAIR sequence images to a unified space. Deep learning image segmentation can deploy a 3D U-Net model trained on a PyTorch or TensorFlow framework. This model is pre-trained on a public dataset containing a large number of brain tumor MRI images and their expert annotations (e.g., the BraTS dataset) to identify the tumor core, edema area, and enhanced tumor region. 3D volume calculation can be obtained directly by counting the number of voxels in the segmentation mask and multiplying by the voxel size. Spatial coordinate annotation can register the segmentation results to the MIST (Multimodal Image Segmentation and Tracking) brain atlas and output the position of the tumor center point in the MIST coordinate system. Visualization rendering can be implemented using the VTK.js library on the web, allowing doctors to view the 3D reconstructed brain structure and tumor region in a browser. A dedicated quality control interface can be a desktop / web application developed based on Qt or React, providing drawing tools, eraser tools, and threshold adjustment sliders for doctors to manually correct segmentation results. In the text structure parsing and information extraction submodule, a BERT-based large language model (e.g., BioBERT or ClinicalBERT) fine-tuned with medical domain data can be introduced. This model processes unstructured text medical records through prompt word engineering, such as "Extract the patient's neurological examination results from the following medical record text, including motor function, language function, sensory function, and higher cortical function, and indicate the assessment time." When extracting neurological examination entries, the system can first identify all paragraphs in the medical record containing keywords such as "neurological examination" and "physical examination," and then filter out the records closest to the surgery date based on the timestamps of these paragraphs. The extracted information, such as left limb muscle strength grade IV, right limb muscle strength grade V, normal naming test, and decreased pain sensation, will be mapped into a preset JSON format template, for example: `{"Assessment Date":"YYYY-MM-DD", "Motor Function": {"Left Side": "Grade IV", "Right Side": "Grade V"}, "Language Function": "Normal", "Sensory Function": "Decreased Pain Sensation"}`. Before the surgery time, the system can send a message to the attending physician through the hospital's internal instant messaging system or electronic medical record system's to-do list: Please review the preoperative neurological assessment draft for patient [Patient Name].

[0053] Through the above implementation methods, the feature extraction module can efficiently and accurately transform complex and heterogeneous clinical data (e.g., MRI images and unstructured text medical records) into standardized, structured feature vectors and assessment information. The image intelligent quantification and feature extraction submodule, through a fully automated standardization processing pipeline, ensures the consistency and accuracy of image data processing and, combined with a physician review and fine-tuning mechanism, effectively improves the reliability of image feature extraction. The text structure parsing and information extraction submodule, utilizing the deep semantic understanding capabilities of a large language model, overcomes the limitations of traditional methods in processing unstructured text. It can accurately extract key neurological function examination information and structure it, providing high-quality input for subsequent intelligent decision-making. This refined and automated feature extraction capability significantly improves the accuracy and reliability of the entire system in disease matching and prognostic assessment, thus providing a more solid data foundation and more clinically significant decision support for intraoperative awakening decisions in brain tumor surgery.

[0054] In some implementations, the disease matching module is used to construct a structured historical case knowledge base containing basic clinical information of patients, intraoperative awakening implementation tags, and key tags of various postoperative prognostic outcomes. The disease matching module is also used to perform refined matching calculations between target patients and historical cases based on a preset multi-dimensional weighted similarity algorithm. The quantitative evaluation formula of the multi-dimensional weighted similarity algorithm is: disease similarity score = Σ(feature similarity × weight). The disease matching module also sets a matching output mechanism to calculate the similarity scores between the target patient and the intraoperative awakening group and the non-intraoperative awakening group in the historical database, respectively, and to filter a subset of high-similarity cases based on a preset threshold and output the corresponding structured prognostic statistical indicators.

[0055] The structured historical case knowledge base refers to a systematically organized and standardized dataset of historical patient cases. This knowledge base not only stores basic clinical information such as age, gender, diagnosis, and medical history, but also includes labels indicating whether intraoperative awakening was performed, as well as key labels for various postoperative prognostic outcomes, such as postoperative neurological function scores and imaging follow-up results. This knowledge base can be built using relational databases (such as MySQL and PostgreSQL), ensuring data consistency and integrity through strict table structures and field types; alternatively, it can use NoSQL databases (such as MongoDB and Cassandra) to store semi-structured or unstructured data, accommodating more flexible data models. The construction of this knowledge base provides a comprehensive and high-quality data foundation for subsequent similarity matching.

[0056] The pre-defined multi-dimensional weighted similarity algorithm is a computational method for quantitatively assessing the similarity between a target patient and historical cases. This algorithm comprehensively considers multiple clinical feature dimensions and assigns different weights to each feature dimension to reflect its importance in decision-making. For example, weighted variants of cosine similarity, Euclidean distance, or decision tree-based similarity calculation methods can be used. In the cosine similarity implementation, each patient's clinical features can be represented as a multi-dimensional vector. Similarity is measured by calculating the cosine of the angle between two vectors, and weighting factors are introduced accordingly. The core of this algorithm lies in its quantitative evaluation formula: Disease Similarity Score = Σ(Feature Similarity × Weight), which ensures the objectivity and interpretability of the similarity assessment.

[0057] The matching output mechanism refers to a set of rules and procedures for processing and presenting the results after the disease matching module completes similarity calculations. This mechanism aims to transform the calculated similarity scores into information that directly guides clinical decision-making. Specifically, it can calculate the similarity scores between the target patient and case groups in the historical database that underwent intraoperative awakening (awake surgery group) and case groups that did not (non-awake surgery group). Subsequently, based on a preset similarity threshold (e.g., a similarity score greater than or equal to 0.6), the mechanism selects a subset of cases highly similar to the target patient from these two groups—the high-similarity case subset. Finally, the mechanism outputs structured prognostic statistics corresponding to these high-similarity case subsets, such as the average tumor resection rate and the incidence of new postoperative neurological dysfunction in each group, for physician reference.

[0058] For example, a structured historical case knowledge base can be deployed using a distributed database system (such as Apache Cassandra) to support the storage of large-scale case data and high-concurrency queries. Each case record in the knowledge base can contain a unique patient ID and multiple attribute fields associated with that ID, such as: tumor location (encoded as a number, e.g., frontal lobe 1, temporal lobe 2, etc.), preoperative tumor volume (number, in cm³), age group (number, e.g., 0-18 years old coded as 1, 19-40 years old coded as 2, etc.), gender (binary code, male 0, female 1), intraoperative awakening implementation label (boolean value, yes 1, no 0), postoperative motor function score (number, e.g., MRC muscle strength grade), postoperative language function score (number, e.g., BNT naming test score), etc. The pre-defined multi-dimensional weighted similarity algorithm can use a machine learning-based similarity model, for example, by training a support vector machine (SVM) or neural network model to learn the contribution of different feature combinations to similarity and output a similarity score between 0 and 1. During calculation, the model can assign higher weights to key features such as tumor location and preoperative tumor volume, while assigning lower weights to secondary features such as gender. The matching output mechanism can be designed as an interactive interface. After the doctor inputs the clinical data of the target patient, the system backend will calculate the similarity score between the patient and historical cases in real time. The interface will clearly display two lists: one for high-similarity intraoperative awakening cases and the other for high-similarity non-intraoperative awakening cases. Each list will show the number of selected cases, the average similarity score, and visually present the statistical comparison of the two groups of cases in key prognostic indicators such as tumor resection rate and incidence of new neurological dysfunction in a chart format. For example, if the similarity score between the target patient and the intraoperative awakening group is generally higher than that of the non-intraoperative awakening group, and the prognostic indicators of the intraoperative awakening group are better, the system will tend to recommend intraoperative awakening.

[0059] Through the above implementation methods, a structured historical case knowledge base was constructed, containing basic patient clinical information, intraoperative awakening implementation tags, and key tags for various postoperative prognostic outcomes. This provides a comprehensive and high-quality data foundation for subsequent similarity matching. By introducing a pre-defined multi-dimensional weighted similarity algorithm and clarifying the quantitative evaluation formula, the system achieves accurate and objective quantification of the similarity between the target patient and historical cases, overcoming the subjectivity and uncertainty of traditional experience-based judgment. This enables the system to identify historical cases with highly similar clinical characteristics to the target patient, thus providing a more targeted and reliable reference for intraoperative awakening decisions. By setting a matching output mechanism, similar cases in the intraoperative awakening group and the non-intraoperative awakening group can be clearly distinguished, and a subset of highly similar cases can be selected based on a pre-defined threshold, thereby outputting structured prognostic statistical indicators. This mechanism not only improves the clinical interpretability of the matching results but also directly provides potential prognostic comparison data under two decision paths (awakening or not awakening), greatly enhancing clinicians' confidence and accuracy in making intraoperative awakening decisions in complex brain tumor surgeries, effectively reducing decision-making risks, and helping to optimize patients' postoperative functional prognosis.

[0060] In some implementations, the core features and weights of the multi-dimensional weighted similarity algorithm are configured as follows: tumor location weight 40%, preoperative tumor volume weight 25%, age group weight 10%, gender weight 5%, motor function weight 5%, language function weight 5%, sensory function weight 5%, and higher cortical function weight 5%. The motor function is assessed using the MRC muscle strength grading standardization, the language function is assessed using the BNT naming test standardization, the sensory function is assessed using the light touch / pinprick sensation score standardization, and the higher cortical function is assessed using the MoCA score standardization. The preset threshold is ≥0.6, and the disease matching module is used to sort the high-similarity case subsets in descending order of similarity.

[0061] The core feature terms and weight configurations of the multi-dimensional weighted similarity algorithm are designed to ensure that similarity calculations fully reflect key clinical factors in intraoperative awakening decisions for brain tumors. The 40% weight for tumor location and the 25% weight for preoperative tumor volume are based on their high correlation with neurological prognosis and surgical difficulty; for example, tumors located in functional areas or with large volumes are crucial for the selection of intraoperative awakening strategies. The 10% weight for age group reflects the differences in physiological reserves and recovery capabilities among patients of different ages. The 5% weight for gender considers the potential impact of gender on certain neurological functions or tumor biological behavior. The 5% weights for motor function, language function, sensory function, and higher cortical function directly focus on the patient's preoperative neurological functional status, which are core targets for intraoperative awakening monitoring and protection. By presetting these feature terms and their weights, similarity assessment can be made more clinically oriented. In other implementations, these weights can be dynamically adjusted according to different tumor types, surgical purposes, or specific clinical research results; for example, the weight for language function can be appropriately increased for tumors in language functional areas; or the weight configuration can be automatically optimized based on a large amount of historical data using machine learning methods. The motor function assessment employed the MRC muscle strength grading standard to provide an objective and repeatable quantitative standard for motor function. The MRC muscle strength grading is an internationally recognized assessment method, classifying muscle strength into 0-5 levels, accurately reflecting a patient's limb motor ability. This standardized assessment ensures the comparability of motor function data among different patients, providing reliable input for similarity matching. In addition to the MRC muscle strength grading, standardized tools such as the Fugl-Meyer Assessment Scale or the motor score in the National Institutes of Health Stroke Scale (NIHSS) can also be used for assessment to adapt to different clinical scenarios or assessment focuses. The language function assessment employed the BNT (Boston Naming Test) to quantify the patient's naming ability. The BNT is a widely used neuropsychological test that assesses a patient's vocabulary retrieval and naming functions through a picture naming task, and is significant for assessing language dysfunction in patients with brain tumors. Standardized testing reduces subjectivity and improves the accuracy and consistency of assessment results. Furthermore, more comprehensive language function assessment tools such as the Western Aphasia Kit (WAB) or the Boston Diagnostic Aphasia Assessment (BDAE) can be considered to obtain richer information on language function. The sensory function was assessed using a standardized light touch / pinprick scoring system to objectively evaluate the patient's superficial sensory function. Light touch and pinprick are commonly used sensory assessment items in neurological examinations. By applying light touch and pinprick stimulation to the patient's skin, their ability to perceive different sensory stimuli is assessed and quantified. This standardized assessment helps identify sensory pathway impairments and provides crucial information for similarity matching.In other implementations, two-point discrimination, vibrational perception, and other deep sensory assessment methods can be combined to provide a more comprehensive sensory function assessment. The higher cortical functions are assessed using the MoCA (Montreal Cognitive Assessment) score, designed for rapid and comprehensive screening of the patient's overall cognitive function. The MoCA score is a simple and effective cognitive function assessment tool covering multiple cognitive domains, including attention, executive function, memory, language, visuospatial ability, abstract thinking, calculation, and orientation, and can sensitively detect mild cognitive impairment. The standardized scoring ensures the comparability and reliability of cognitive function data. In addition to the MoCA score, other standardized cognitive assessment tools such as the Mini-Mental State Examination (MMSE) or the Cambridge Cognitive Assessment (ACE-III) can be used to meet different clinical needs or assessment depths. The preset threshold of ≥0.6 is used to define the range of highly similar cases. When the disease similarity score between the target patient and a historical case reaches or exceeds 0.6, the historical case is considered highly similar to the target patient and is thus included in the high-similarity case subset. This threshold is set based on clinical experience and data analysis, aiming to balance the quantity and quality of matched cases and ensure that the selected cases have sufficient reference value. In other implementations, this threshold can be adjusted according to clinical needs. For example, to obtain more stringent matching results, the threshold can be increased to 0.7 or 0.8; or the threshold can be dynamically optimized based on historical data and clinical feedback using machine learning methods. The disease matching module is used to sort a subset of highly similar cases in descending order of similarity, so as to facilitate doctors to quickly identify and prioritize cases most similar to the target patient. By sorting in descending order, the cases with the highest similarity scores will be displayed at the top of the list, allowing doctors to first review those cases that are closest to the current patient in terms of clinical characteristics, thereby obtaining valuable clinical information and prognostic references more efficiently. In addition to sorting in descending order of similarity, it can also be sorted according to specific clinical needs, such as by postoperative complication rate, postoperative functional recovery, and other indicators, to provide references from different dimensions.

[0062] Through the above implementation methods, the core feature terms and their weight configurations of the multi-dimensional weighted similarity algorithm were clarified, and an internationally recognized standardized neurofunctional assessment method was introduced. Clear similarity screening thresholds and ranking rules were also established. This addresses the problem of insufficient clinical relevance of disease matching results when specific quantitative standards are lacking. Specifically, by assigning reasonable weights to different clinical features, the clinical guidance of similarity calculation is ensured; standardized assessment guarantees the objectivity and comparability of neurofunctional data; preset thresholds effectively screen out truly valuable high-similarity cases; and descending order ranking enables physicians to efficiently obtain the most suitable clinical cases for their target patients. Therefore, this scheme significantly improves the accuracy, reliability, and clinical guidance significance of the prognostic comparison data output by the disease matching module, providing physicians with more refined and personalized support for intraoperative awakening decisions in brain tumor surgery, thereby optimizing the clinical decision-making process.

[0063] In some implementations, structured prognostic statistics include four categories: case size statistics, baseline characteristic distribution, surgical outcome indicators, and functional prognostic comparisons. Case size statistics refer to the number of matched cases and their proportion in the corresponding historical groups. Baseline characteristic distributions include tumor location, preoperative tumor volume, age group, and gender composition. Surgical outcome indicators include tumor resection rate, fiber bundle injury rate, and total volume of newly developed edema / damage area. Functional prognostic comparisons refer to the quantitative comparison values ​​of motor, language, sensory, and higher cortical functions before and after surgery.

[0064] Structured prognostic statistics refer to a set of organized, categorized, and quantified data used to assess postoperative recovery and treatment outcomes. The concept lies in transforming complex clinical prognostic information into a standardized form that is easy to understand and analyze. This can be achieved by data population using pre-defined data models and fields, or by mapping raw data to structured templates using data cleaning and transformation tools. Case size statistics refer to the number of cases highly similar to the target patient selected from a historical case knowledge base during similarity matching, and the proportion of these cases in the corresponding group (e.g., intraoperative awake group or non-intraoperative awake group) within the entire historical database. This helps physicians understand the representativeness and statistical significance of the matching results. This can be achieved by counting the number of matched cases and calculating their percentage in the total, or by visualizing the distribution of cases in different groups. Baseline feature distribution refers to the distribution of the matched subset of historical cases across key clinical features, including tumor location, preoperative tumor volume, age group, and gender composition. These features are important factors influencing prognosis, and understanding their distribution helps assess the reliability and applicability of the matching results. These can be achieved by calculating the mean, median, standard deviation, or frequency distribution of these characteristics using statistical methods, or by visualizing them using histograms, pie charts, etc. Surgical outcome indicators refer to key parameters used to quantitatively assess the direct effects of surgery, including tumor resection rate, fiber tract injury rate, and total volume of newly developed edema / damage areas. These indicators directly reflect the success of the surgery and its impact on brain tissue. They can be achieved by automatically calculating these indicators using imaging analysis software or by manually measuring them, or by using pre-set scoring criteria for graded assessment. Functional prognosis comparison refers to the quantitative assessment of various neurological functions (such as motor, language, sensory, and higher cortical functions) before and after surgery, and comparing their changes. This helps physicians comprehensively understand the trend and extent of postoperative functional recovery. This can be achieved by scoring using standardized scales (such as MRC muscle strength grading, BNT naming test, MoCA score, etc.) and calculating the differences before and after, or by presenting quantitative changes through percentage changes, absolute value changes, etc.

[0065] Through the above implementation methods, the prognostic comparison data output by the disease matching module is structured, classified, and quantified, enabling doctors to obtain and understand key prognostic information more clearly and comprehensively. This multi-dimensional and standardized presentation significantly reduces the time cost and cognitive burden on doctors in extracting effective information from complex data, avoiding decision-making biases caused by disorganized data. Doctors can intuitively compare the potential prognoses of patients under different treatment strategies, thereby enabling them to more accurately weigh surgical risks and functional protection during intraoperative awakening decision-making, and formulate treatment plans that are more suitable for individual patient conditions, improving the scientific nature and personalization of decision-making.

[0066] In some implementations, the report generation module receives and integrates comparative data of the target patient and highly similar intraoperative awakening (IWA) and non-IWA groups output by the disease matching module, and injects clinical analysis task instructions into the large language model through cue word engineering. Receiving and integrating comparative data means that the report generation module obtains structured data from the disease matching module after similarity matching and prognostic statistical processing. This data includes detailed characteristics of the target patient and various prognostic statistical indicators (such as tumor resection rate and neurological function recovery) of highly similar historical cases (divided into IWA and non-IWA groups). The purpose of integration is to provide the large language model with a unified and comprehensive input dataset. Cue word engineering refers to the careful design and optimization of text instructions input to the large language model to guide it in generating reports according to the expected format, content, and analytical depth. This may include defining the report structure, specifying the key indicators to be analyzed, setting the role of the large language model (e.g., as a clinical analysis expert), and providing contextual information. Injecting clinical analysis task instructions involves sending these cue word-engineered instructions along with the integrated comparative data to the large language model, clearly informing it of the analytical tasks to be performed and the report generation requirements.

[0067] The report generation module also drives the large language model to generate a comparative analysis report on intraoperative awakening strategies and neurological function prognosis, including visual comparison tables, in-depth data interpretation and clinical inference, and targeted anesthesia clinical recommendations. Driving the large language model to generate the report means that the report generation module, acting as a coordinator, invokes the core capabilities of the large language model, enabling it to process information, perform logical reasoning, and generate text based on received data and instructions. The generated report is rich in content. The visual comparison tables intuitively show the differences in key prognostic indicators between the target patient and two groups of historical cases, facilitating rapid data comprehension by physicians. The in-depth data interpretation and clinical inference involve the large language model professionally analyzing these comparative data to reveal potential clinical significance, risk factors, and prognostic trends, providing evidence-based clinical inferences. The targeted anesthesia clinical recommendations, combined with the patient's specific condition and prognostic analysis, provide guiding anesthesia management strategies, such as the timing of intraoperative awakening, the selection of anesthetic drugs, and the focus of neurological function monitoring.

[0068] The report generation module features an interactive interface with two-way selection buttons for "Accept" and "Regenerate." It also automatically records interaction data and calculates the clinical acceptance accuracy of the large language model output. The formula for calculating the clinical acceptance accuracy is: Acceptance Accuracy = Number of Acceptances / (Number of Acceptances + Number of Regenerations) × 100%. The interactive interface provides doctors with a direct feedback channel to the system. After reviewing the report generated by the large language model, doctors can choose to accept it to indicate their approval of the report content, or choose to regenerate it to indicate that the report needs improvement. Automatic recording of interaction data means that the system captures every doctor's action (acceptance or regeneration) in real time and records relevant information such as time, user, and report version. Calculating the clinical acceptance accuracy uses a preset formula to quantitatively evaluate the quality of the report generated by the large language model and the doctor's acceptance, providing an objective basis for subsequent system optimization.

[0069] Through the above implementation methods, the report generation module not only efficiently generates intraoperative awakening strategy neurological function prognosis comparative analysis reports that include visualized comparison tables, in-depth data interpretation and clinical inference, and targeted anesthesia clinical recommendations using a large language model, but more importantly, it effectively solves the quality assessment and continuous optimization problems of reports generated by the large language model by setting up an interactive interface with two-way selection buttons for adoption and regeneration, and automatically recording interaction data and calculating clinical adoption accuracy. This allows doctors to directly participate in the report generation and improvement process, ensuring the clinical usability and accuracy of the reports, significantly improving doctors' trust in the system output and adoption rate, and providing a valuable data foundation for subsequent iterative optimization of the large language model, thereby continuously improving the intelligence level and clinical value of the entire decision support system.

[0070] In some implementations, the regenerated options are configured with preset feedback reason categories, including factual errors, logical flaws, key omissions, unfounded fabrications, deviations from instructions, inoperability, and others; the report generation module is used to use the recorded interaction data as a key basis for iterative fine-tuning of the prompt word strategy or large language model, driving the continuous evolution of the system's decision support capabilities.

[0071] The regenerated report options are configured with preset feedback reason categories. When a doctor chooses to regenerate the report, the system provides a series of predefined options for the doctor to choose from to explain their reasons for regeneration. These categories can be fixed lists, presented to the user via dropdown menus or checkboxes, or dynamically generated or recommended by the system based on context. These categories aim to transform the doctor's subjective feedback into quantifiable and analyzable structured data for subsequent systematic improvements. The report generation module uses the recorded interaction data as a key basis for iterative fine-tuning of the prompting strategy or the large language model. This means the system not only records the doctor's choice to regenerate, but more importantly, it records the feedback reason category chosen by the doctor. This interaction data with specific reasons can be used by the system to analyze the accuracy, logic, and completeness of the large language model's report generation in different contexts. For example, the system can statistically analyze the frequency of various feedback reasons to identify weaknesses in the large language model or prompting strategy, thereby guiding the optimization of prompts (i.e., instructions issued to the large language model) to make them clearer and more specific, or it can be used for supervised fine-tuning or reinforcement learning of the large language model itself to improve its performance on specific tasks.

[0072] Through the above implementation methods, doctors can provide specific and structured feedback when choosing to regenerate reports, such as pointing out factual errors or logical flaws in the report. This refined feedback mechanism enables the report generation module to accurately identify the specific types and causes of problems in the large language model during report generation, thus providing clear and targeted basis for optimizing the prompt word strategy or iteratively fine-tuning the large language model. This significantly improves the efficiency of system learning and evolution, ensures continuous improvement in decision support capabilities, and ultimately makes the generated clinical decision reports more accurate, reliable, and better suited to actual clinical needs.

[0073] In some implementations, the closed-loop feedback module automatically initiates a multi-threaded data acquisition and processing flow after surgery. This flow sequentially performs demographic and key decision information collection, automatic quantification of multi-timepoint imaging prognostic indicators, structured acquisition of neurological functional prognostic data, data association and integration, and knowledge base updates, ultimately forming a self-evolving decision support closed loop. When performing the automatic quantification step of multi-timepoint imaging prognostic indicators, the closed-loop feedback module invokes the fully automated standardized processing pipeline of the aforementioned intelligent image quantification and feature extraction submodules.

[0074] The closed-loop feedback module automatically initiates a multi-threaded data acquisition and processing flow after surgery. This means that upon receiving a signal or event indicating the completion of surgery, the module automatically triggers a series of data acquisition and processing tasks without manual intervention. The multi-threaded data acquisition and processing flow can be understood as the system being able to execute multiple data-related operations simultaneously or concurrently to improve data processing efficiency and response speed. This can be achieved by monitoring the surgery completion status update in the Hospital Information System (HIS) or receiving surgery completion notifications through the Operating Room Management System (ORMS) interface, followed by the system's internal scheduler or workflow engine initiating a pre-defined data processing task queue. Another implementation method is that the system can be configured with a scheduled task to automatically scan relevant data sources and initiate the acquisition process some time after surgery.

[0075] The process sequentially collects demographic and critical decision information, aiming to systematically gather and organize basic patient information as well as critical information generated during intraoperative awakening decisions. Demographic information is typically obtained from the electronic medical record (EMR) system, while critical decision information may originate from decision reports, anesthesia records, or surgical records output by the report generation module. The collection process can employ data extraction, transformation, and loading (ETL) tools to consolidate data scattered across different systems into a central database. Alternatively, the system can provide a structured form interface to guide healthcare professionals in supplementing critical decision information postoperatively and automatically linking it to the patient's demographic information.

[0076] Automated quantification of multi-timepoint imaging prognostic indicators refers to the system's ability to automatically acquire and quantify patients' imaging data at different time points after surgery to assess postoperative recovery and prognostic indicators. Automated quantification means that manual measurement is unnecessary; instead, image processing and feature extraction are performed using pre-defined algorithms and models. This may include calling professional medical imaging processing software interfaces or utilizing built-in image analysis algorithms to perform operations such as image data segmentation, registration, and volume calculation.

[0077] Structured acquisition of neurological prognostic data aims to systematically collect and record postoperative neurological function assessment data from patients and transform it into a standardized, computer-processable structured format. Neurological function assessments are typically performed by clinicians and involve examinations of the patient's motor, language, sensory, and cognitive functions. Structured acquisition can be achieved by providing pre-set electronic assessment forms containing standardized scoring items and options, which physicians simply select or input values ​​for. Alternatively, the system can utilize Natural Language Processing (NLP) technology to extract key neurological function descriptions from unstructured text recorded by physicians and map them to predefined structured fields.

[0078] Data association and integration, along with knowledge base updates, involves the unified association and integration of the aforementioned collected demographic information, key decision-making information, multi-timepoint imaging prognostic indicators, and neurological functional prognostic data. The purpose of this association and integration is to aggregate these scattered data fragments into a complete case record with a time sequence and logical relationships. The integrated data is then used to update the structured historical case knowledge base. This can be achieved through relational queries and data insertion operations in a database management system (DBMS), ensuring the integrity of new data with existing cases. Knowledge base updates can employ an incremental update strategy, adding only newly generated case data or updating the latest status of existing cases, thereby continuously enriching the content of the knowledge base.

[0079] The steps to form a self-evolving decision support closed loop describe how, through the aforementioned data collection, processing, integration, and knowledge base updates, the system can establish a mechanism for continuous learning and improvement. When new postoperative outcome data is integrated into the knowledge base, this data serves as new training samples to optimize the similarity algorithm of the disease matching module, the large language model of the report generation module, and so on. This continuous feedback and learning process enables the system to continuously adjust and improve its decision support capabilities based on the latest clinical practice and patient outcome data, thereby achieving self-evolution.

[0080] When performing the automatic quantification of multi-timepoint imaging prognostic indicators, the closed-loop feedback module invokes the fully automated standardized processing pipeline of the aforementioned intelligent image quantification and feature extraction submodules. This clarifies that when processing postoperative imaging data, the closed-loop feedback module does not develop a new processing flow independently, but rather reuses or invokes the fully automated standardized processing pipeline built by the existing intelligent image quantification and feature extraction submodules within the feature extraction module. This reuse mechanism ensures consistency, standardization, and efficiency in preoperative and postoperative imaging data processing. The invocation can be achieved through application programming interfaces (APIs) for inter-module communication, or through shared function libraries or services.

[0081] For example, the closed-loop feedback module can be configured to listen for updates to the surgical status field in the Hospital Information System (HIS). For instance, when the surgical status changes from in progress to completed, a message queue service (such as Kafka or RabbitMQ) is triggered to send a surgical completion event to the closed-loop feedback module. Upon receiving this event, the closed-loop feedback module will start a data processing coordinator based on a microservice architecture. The coordinator concurrently launches multiple independent microservices: a demographic information collection service is responsible for obtaining patient information such as ID, age, gender, and diagnosis from the HIS and Electronic Medical Record (EMR) systems via a RESTful API interface, and obtaining anesthesia plans and intraoperative events for intraoperative awakening decisions from the anesthesia recording system; an image prognosis quantification service automatically retrieves the patient's cranial MRI images from the Picture Archiving and Communication System (PACS) via the DICOM protocol according to preset follow-up time points, and calls the deep learning model pipeline based on the TensorFlow or PyTorch framework provided by the aforementioned image intelligent quantification and feature extraction submodule to automatically calculate indicators such as residual tumor volume and volume of newly developed FLAIR high signal areas; a neurological function acquisition service pushes an electronic form link to the attending physician. This form is designed based on the HL7 standard and includes standardized assessment items for motor function (such as MRC muscle strength grading), language function (such as BNT naming test), sensory function (such as light touch / pinprick sensation score), and higher cortical function (such as MoCA score). After the physician completes the form, the data is automatically submitted. All collected data is then processed uniformly by a data integration service. This service uses the patient's unique identifier (such as hospital number) as the primary key to associate data from different sources and convert it into structured data packets in JSON format. Finally, these structured data packets are written to a structured historical medical record knowledge base built on PostgreSQL or MongoDB via database connectors (such as JDBC or ODBC), enabling incremental updates to the knowledge base.

[0082] Through the above implementation methods, a comprehensive, automated, and continuously optimized postoperative feedback mechanism is established. After surgery, the system can automatically and efficiently collect and process postoperative outcome data from multiple sources and time points, including demographic information, key decision-making information, imaging prognostic indicators, and neurological function assessment data. This automated data flow and structured processing significantly reduces manual intervention and improves the timeliness and accuracy of data collection. More importantly, by associating and integrating these real postoperative outcome data with preoperative characteristic data and intraoperative decision-making processes, and continuously updating the structured historical case knowledge base, the system can continuously learn and adapt to new clinical situations. This allows the disease matching module to refer to richer and more realistic case data when performing similarity matching, thereby significantly improving the accuracy of matching and the reliability of prognostic prediction. Ultimately, this self-evolving decision support closed loop ensures that the system can continuously optimize its decision-making model based on the latest clinical feedback and patient outcomes, providing physicians with more valuable and clinically guiding intraoperative awakening strategy suggestions, thereby improving the safety and effectiveness of brain tumor surgery.

[0083] In some implementations, the automatic quantification step of multi-timepoint imaging prognostic indicators is used to retrieve the patient's first follow-up cranial MRI images from the image archiving communication system at preset key follow-up time points after surgery. The automated processing pipeline accurately calculates the residual tumor volume, the volume of newly developed FLAIR high signal areas, and the fiber bundle damage rate after surgery. The key follow-up time points include postoperative days 1-3, 7-14, and 28-40. The structured acquisition step of neurological function prognostic data sets a T+1 reminder rule to send an update reminder to the attending physician on the first day after the task is triggered at each preset postoperative assessment time point. The reminder guides the physician to complete the update and confirmation of the postoperative neurological assessment form based on the preoperatively reviewed and confirmed postoperative assessment form, and automatically calculates the changes of various neurological function scores relative to the preoperative baseline.

[0084] The automated quantification step of multi-timepoint imaging prognostic indicators aims to automatically analyze postoperative imaging data at multiple preset time points to obtain key prognostic indicators. Its role is to provide objective and quantitative information on postoperative imaging changes, which is crucial for assessing surgical outcomes, monitoring disease progression, and updating the historical case knowledge base. This step can be implemented by integrating professional medical imaging processing software or developing customized image analysis algorithms. For example, it can utilize traditional image processing techniques or machine learning methods to identify and quantify tumor residues, edema areas, etc. Alternatively, it can employ deep learning-based image segmentation models to perform pixel-level analysis of MRI images, thereby accurately calculating various volumetric indicators and damage rates. The preset key follow-up time points are critical time points that the system automatically triggers for image data retrieval and processing. These time points are set based on the patterns of postoperative recovery and complication observation in clinical practice. These time points can be pre-coded in the system configuration, for example, managed through a timetable or calendar module; or they can be implemented through a configurable rule engine, allowing clinicians or system administrators to dynamically adjust these follow-up time points according to different disease types, surgical methods, or individual patient conditions. Retrieving the patient's initial follow-up cranial MRI images from the image archiving and communication system is the source and method of data acquisition, ensuring authoritative and complete original image data. This can be achieved through interface integration with the PACS system via the standard DICOM protocol, enabling automatic querying and transmission of image data; alternatively, it can be achieved indirectly through API interfaces provided by the hospital information system or electronic medical record system, obtaining the storage path or identifier of the image data, and then downloading the corresponding image files from the PACS storage server via file transfer protocols. The automated processing pipeline refers to a series of predefined, automatically executed image processing steps designed to standardize and quantify the raw MRI images. This pipeline can consist of multiple independent software modules; for example, one module might handle image preprocessing, another image segmentation, and yet another volume calculation and feature extraction. Alternatively, it can be built using a unified software platform or framework, leveraging its rich image processing algorithm library and visualization tools to integrate various processing functions into a unified interface. Accurate calculation of postoperative residual tumor volume, newly developed FLAIR high-signal area volume, and fiber bundle damage rate are specific imaging prognostic indicators, directly reflecting the immediate surgical outcome and potential complications. The residual tumor volume and the volume of the FLAIR high-signal area can be identified by image segmentation algorithms to pinpoint the lesion region, and then calculated cumulatively based on pixel or voxel size. The calculation of fiber tract injury rate typically requires combining diffusion tensor imaging (DTI) data. By tracing fiber tracts in DTI data, key nerve fiber tracts are reconstructed, and then the integrity or degree of damage of these fiber tracts in the surgical area is analyzed.

[0085] The structured acquisition step for neurological prognostic data aims to systematically and structurally collect postoperative neurological function assessment data from patients. Its purpose is to provide objective evidence of the patient's functional recovery, forming a comprehensive prognostic assessment system together with imaging indicators. This step can be implemented through an electronic assessment form system with preset neurological function assessment items; doctors simply need to select or enter the corresponding scores on the interface. Alternatively, it can be integrated with the hospital's electronic medical record system to automatically extract structured neurological function assessment records from medical records, or use natural language processing technology to extract information from unstructured text and transform it into structured data. The T+1 reminder rule is a time management and task reminder mechanism designed to ensure that doctors complete neurological function assessments promptly after the preset assessment time. This can be implemented through a task scheduling system that automatically generates a reminder task after each preset assessment time and sends it to the attending physician via SMS, email, or system in-app message on the first day after the task is triggered. Alternatively, it can be integrated into the hospital's clinical workflow management system, where the system automatically adds the corresponding assessment task to the attending physician's to-do list when a patient's postoperative assessment time is approaching or approaching. The system guides physicians through updating and confirming postoperative neurological assessments based on preoperatively approved forms, emphasizing standardization and continuity of assessment. The system can automatically load pre-filled templates of confirmed preoperative neurological assessments, allowing physicians to modify or supplement postoperative results. Alternatively, an interactive interface clearly displays preoperative assessment results and provides input boxes or selection lists for physicians to fill in postoperative assessment data. The system automatically calculates the changes in various neurological function scores relative to the preoperative baseline, a crucial step in quantitatively analyzing neurological function assessment data. After physician confirmation of the postoperative assessment form, the system automatically retrieves the corresponding preoperative assessment data from the database and performs simple subtraction or percentage change calculations on the scores. More complex statistical methods can also be employed; for example, for ordinal scores, it can calculate grade changes; for continuous scores, it can calculate mean differences or standardized effect sizes.

[0086] The above implementation methods ensure timely, accurate, and standardized acquisition of postoperative multi-timepoint imaging and neurological function data. The automatic quantification of multi-timepoint imaging prognostic indicators, combined with the setting of key follow-up time points, enables the system to objectively and continuously monitor postoperative residual tumor, edema, and fibrous bundle damage, providing reliable quantitative evidence for assessing surgical outcomes and disease progression. Simultaneously, the T+1 reminder rule and preoperative baseline-based assessment mechanism introduced in the structured acquisition step of neurological function prognostic data effectively improve the timeliness and accuracy of neurological function assessment and achieve quantitative comparison of functional recovery. These mechanisms collectively address potential issues of lag, inconsistency, or inaccuracy in postoperative data acquisition, thereby providing high-quality postoperative outcome data for the closed-loop feedback module. This significantly enriches and updates the structured historical case knowledge base, thereby significantly improving the matching accuracy and recommendation reliability of the disease matching module. Ultimately, this allows the entire system to continuously optimize its intelligent decision-making capabilities based on more comprehensive and accurate feedback information.

[0087] In some implementations, data association and knowledge base updates include automatically associating and integrating the collected and calculated postoperative data with the target patient's complete preoperative feature dataset and intraoperative decision-making process records. The postoperative data includes demographic information, intraoperative decision labels, multi-timepoint imaging indicators, and multi-timepoint neurological function scores. The closed-loop feedback module is used to store the integrated complete case data with decision-outcome labels in a highly structured form into the structured historical case knowledge base, continuously enriching and updating the knowledge base of similarity matching, and improving the matching accuracy and recommendation reliability of the disease matching module.

[0088] The data association and integration, along with the knowledge base update process, aims to logically connect and unify patient data scattered across different time points and modules, and then use this data to update the system's core knowledge base. The concept is to establish inherent connections between data, ensuring the integrity and consistency of information, thereby providing a more comprehensive basis for subsequent intelligent decision-making. This can be achieved through cross-system data matching and merging using unique patient identifiers; or through predefined data models and mapping rules, mapping data fields from different sources to a unified structured format. Postoperative data acquisition and calculation refers to patient-related information acquired automatically or semi-automatically and preliminarily processed (e.g., quantified) after surgery. Its role is to provide objective evidence of the patient's actual treatment effect and recovery status, and is crucial for evaluating the effectiveness of intraoperative decisions. Implementation methods may include automatically retrieving postoperative follow-up records from the hospital information system, obtaining follow-up imaging reports from the image archiving communication system, and using postoperative assessment forms entered by specialists. The complete preoperative feature dataset of the target patient refers to the collection of all relevant clinical characteristics possessed by the patient before surgery. Its role is to provide initial conditions and background information for decision-making, and is the basis for the system's similarity matching and decision analysis. The implementation methods can include patient demographic information, medical history, preoperative imaging results, preoperative neurological function assessment results, and laboratory test data. This data is typically standardized and structured in the data acquisition and feature extraction modules. Intraoperative decision-making process records are detailed accounts of key treatment strategies and interventions taken during the surgery. Their purpose is to clarify whether intraoperative awakening was implemented and the specific operational details, providing a basis for assessing the causal relationship between decisions and outcomes. Implementation methods can include anesthesia records, surgical records, intraoperative neurophysiological monitoring reports, etc., which may contain explicit decision labels. Demographic information comprises the patient's basic identity characteristics and background data, such as age, sex, and ethnicity. Its purpose is to serve as part of the patient's individual characteristics for disease matching and prognostic analysis, helping the system identify characteristic differences among different populations. Implementation methods can include automatic extraction from hospital information systems or electronic medical record systems and standardization processing. Intraoperative decision labels are classification markers used to identify whether a patient received a specific treatment strategy (e.g., intraoperative awakening) during surgery. Its function is to serve as a key categorical variable, distinguishing patient groups from different treatment pathways for subsequent prognostic comparative analysis and efficacy evaluation. It can be implemented as a pre-set field in the surgical or anesthesia record, which the physician selects or fills in during or after the operation. Multi-timepoint imaging indicators refer to quantitative data obtained through imaging examinations (such as MRI) at different pre-set postoperative time points (e.g., days 1-3, 7-14, and 28-40).Its function is to dynamically monitor tumor residue, edema, and fibrous bundle damage, assess surgical outcomes and complications, and provide objective imaging evidence. This can be achieved through automated processing and quantification of MRI images at different time points using an intelligent image quantification and feature extraction submodule. Multi-timepoint neurological function scoring refers to standardized assessments of the patient's neurological function at different preset postoperative time points. Its function is to objectively reflect the recovery or impairment of the patient's postoperative neurological function, serving as an important basis for assessing the patient's quality of life and functional prognosis. This can be achieved by the attending physician performing a neurological examination at preset postoperative assessment time points, scoring the patient using a standardized scale, and then structurally inputting these scores into the system. The structured historical case knowledge base is a meticulously organized and standardized database storing a large amount of complete clinical data and treatment outcomes from historical patients. Its function is to serve as the knowledge base for the system's similarity matching and decision support, providing rich reference cases for decision-making for new patients. This can be implemented using a relational database or a NoSQL database, with strictly defined data schemas and fields to ensure data queryability and analyzability. The disease matching module is the core intelligent decision-making engine of the system. It is responsible for finding highly similar cases in a structured historical case knowledge base based on the characteristics of the target patient. Its role is to provide clinicians with prognostic references and decision-making basis based on real-world data, assisting them in developing personalized treatment plans. This can be achieved by using a pre-set multi-dimensional weighted similarity algorithm to calculate the similarity score between the target patient and historical cases, and then filtering out highly similar cases.

[0089] Through the above implementation methods, scattered postoperative outcome data are deeply correlated and integrated with patients' preoperative characteristics and intraoperative decision-making processes, thereby constructing a structured historical case knowledge base containing a complete decision-making-outcome chain. This continuous knowledge base update mechanism enables the disease matching module to perform similarity matching based on richer and more real-world feedback data, significantly improving its matching accuracy and recommendation reliability. The system no longer makes decisions solely based on static historical data, but can learn and evolve from the actual results of each clinical practice, thus providing clinicians with more forward-looking and personalized intelligent decision support, effectively solving the problem of the difficulty in continuously optimizing and improving the system's decision-making capabilities.

[0090] Secondly, such as Figure 2As shown, this application proposes an intelligent triggering, matching, and closed-loop quality control method for intraoperative awakening decisions in brain tumor surgery. This method is based on the aforementioned intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decisions in brain tumor surgery. Through the coordinated operation of six key steps, this method constructs a complete decision support closed loop, effectively solving the core problems in traditional decision-making processes, such as the lack of objective basis for decision triggering, low data integration efficiency, lack of real-world evidence for decision reference, and the inability to effectively accumulate postoperative knowledge.

[0091] In practice, in step S1, in response to the clinical event of anesthesia assessment form submission, the method automatically filters and identifies target patients through a two-layer rule-based screening engine. This engine objectively identifies patient groups that meet the surgical indications and physiological compliance requirements based on preset medical coding rules and patient physiological state indicators. This avoids the screening bias caused by the reliance on subjective judgment by doctors in traditional methods, ensuring that the decision-making process has quantifiable objective evidence. For example, the system can automatically link anesthesia assessment forms and surgical request forms, completing preliminary patient cohort screening based on standardized coding rules, significantly improving the standardization and efficiency of patient screening.

[0092] In step S2, the intelligent image quantization and feature extraction submodule is launched in parallel to process MRI images and extract quantification features. Simultaneously, the text structured parsing and information extraction submodule is launched to process medical records and extract structured physical examination information. The intelligent image quantization and feature extraction submodule performs spatial registration, image segmentation, and volume calculation on multimodal MRI images through an automated processing pipeline, outputting structured feature vectors. The text structured parsing and information extraction submodule utilizes semantic understanding technology to extract standardized neurological function assessment information from unstructured medical records. This parallel processing mechanism achieves efficient integration of multi-source data from hospital information systems, electronic medical record systems, and image archiving communication systems, transforming previously scattered image and text data into unified structured features, significantly shortening data preparation time, and solving the problem of low data integration efficiency caused by manual processing in traditional methods.

[0093] In step S3, a multi-dimensional weighted similarity algorithm is used to calculate the disease similarity score between the target patient and historical cases, outputting highly similar control cases (intraoperative awakening group and non-intraoperative awakening group) and their structured prognostic statistics. This algorithm quantitatively assesses the similarity of key clinical features such as tumor location and preoperative tumor volume, selecting a subset of historical cases with a similarity ≥ 0.6, and outputting structured prognostic statistics in descending order of similarity. These statistics include case size statistics, baseline feature distribution, surgical outcome indicators, and functional prognostic comparisons. This process provides a quantitative reference system based on real-world evidence for clinical decision-making, enabling physicians to intuitively compare the prognostic differences of different treatment options in similar patient groups, effectively compensating for the lack of scientific data support in traditional decision-making.

[0094] In step S4, the matching data is transformed into a comparative analysis report of neurological function prognosis for intraoperative awakening strategies using a large language model. This report includes a visual comparison table, in-depth interpretation, and clinical recommendations. The system records the physician's adoption or regeneration feedback to calculate the clinical adoption accuracy. The system injects clinical analysis instructions into the large language model through prompt word engineering. The generated report not only presents a comparison of prognostic data between the intraoperative awakening group and the non-intraoperative awakening group but also provides professional interpretation and targeted anesthesia recommendations. The interactive interface features a two-way selection button for adoption and regeneration. When the physician chooses to regenerate, they must select a feedback reason from a preset category. The system automatically calculates the clinical adoption accuracy based on this and uses it for model iteration and optimization. This mechanism ensures the clinical usability and continuous improvement capability of the decision report, significantly improving the accuracy of decision support.

[0095] In step S5, the system automatically collects follow-up MRI images and neurological function clinical assessment data at multiple preset follow-up time points after surgery. It then uses the image processing pipeline to calculate key postoperative imaging indicators and the change in neurological function scores relative to the preoperative baseline. The image processing pipeline precisely quantifies indicators such as postoperative residual tumor volume and the volume of newly developed FLAIR high-signal areas. Simultaneously, the system pushes assessment update reminders to the attending physician according to the T+1 reminder rule, guiding the completion of the structured postoperative neurological assessment form. This step achieves automated collection and standardized processing of postoperative outcome data, ensuring the accurate acquisition of key prognostic indicators and providing a reliable data foundation for closed-loop feedback.

[0096] In step S6, postoperative outcome data with intraoperative decision tags is linked and integrated with preoperative characteristic data and intraoperative decision records. The integrated complete case data is then stored in a structured historical case knowledge base, updating and enriching the knowledge base for similarity matching. The system automatically links demographic information, intraoperative decision tags, multi-timepoint imaging indicators, and neurological function scores to form a complete case file with decision-outcome tags, and incrementally updates the knowledge base in a highly structured format. This closed-loop feedback mechanism enables the system to continuously accumulate real-world experience, constantly optimize the accuracy of the disease matching module, and achieve autonomous evolution of decision support capabilities.

[0097] In step S1, the dual-layer rule-based screening engine first completes the screening of surgical indications based on ICD-9-CM-3 and ICD-10 codes, and then completes the screening of patient physiology and cooperation. In step S2, after the preprocessing is completed, the doctor reviews and confirms the image quantification results and physical examination information drafts, forming the first quality control node in the entire process. In step S3, after the similarity score is calculated, a subset of high-similarity cases is selected based on a preset threshold of ≥0.6 and sorted in descending order of similarity. In step S4, if the doctor chooses to regenerate the report, the feedback reason is selected from the preset categories, and the system uses the feedback data for iterative fine-tuning of the large language model. In step S5, the postoperative data collection follows the T+1 reminder rule, and a postoperative quality control node is formed after the image index calculation and neurological function score statistics. In step S6, the knowledge base is updated incrementally, and the newly accumulated case data enables the system to form a self-learning decision support closed loop. Moreover, the method sets doctor review and correction links at each key node of image measurement, physical examination extraction, report review, and postoperative assessment, and records the entire process operation log to achieve traceable quality control.

[0098] The execution order of the two-layer rule-based screening engine is explicitly defined as follows: first, surgical indication screening, followed by patient physiological and compliance screening. Surgical indication screening is based on the International Classification of Diseases (ICD-9-CM-3 surgical procedure codes and ICD-10 disease diagnosis codes), aiming to quickly identify patients who meet the surgical criteria at a macro level. Patient physiological and compliance screening, building upon this, conducts a more detailed assessment of the patient's level of consciousness and compliance to ensure safe and effective intraoperative awakening. This tiered screening sequence effectively improves screening efficiency and accuracy, avoiding subsequent detailed physiological assessments for patients who do not meet basic surgical criteria, thereby optimizing resource allocation. For example, automated coding screening can be performed first, followed by physiological and compliance assessments by medical staff; alternatively, the system can perform both layers of screening simultaneously in the background, but logically prioritizes indication determination.

[0099] After preprocessing, doctors review and confirm the image quantification results and draft physical examination information. This aims to introduce a manual review mechanism to ensure the accuracy of the automated preprocessing results. The image quantification results and draft physical examination information are the foundational data for subsequent decision-making, and their accuracy is crucial. Doctor review and confirmation can correct any deviations or errors that may exist in the automated processing, such as minor errors in image segmentation or semantic ambiguities in text extraction. This step, as the first quality control node in the entire process, ensures data quality from the source, providing reliable input for subsequent disease matching and report generation. Review and confirmation can be performed through an interactive interface, allowing doctors to visually inspect and manually adjust the system-generated quantification results, or edit and supplement the draft physical examination information.

[0100] After the similarity score is calculated, a subset of highly similar cases is selected based on a preset threshold of ≥0.6 and sorted in descending order of similarity. This technical feature specifies the particular selection and sorting strategy of the disease matching module when outputting highly similar cases. The similarity score is a quantitative indicator measuring the degree of similarity between the target patient and historical cases. Setting a preset threshold (e.g., ≥0.6) ensures that the selected case subset has sufficient reference value, avoiding the introduction of cases with excessively low similarity that could interfere with decision-making. Sorting by similarity in descending order allows physicians to prioritize cases most similar to the target patient, thereby more efficiently obtaining valuable clinical information. This selection and sorting mechanism helps improve the accuracy and efficiency of decision-making. In addition to a threshold of ≥0.6, the threshold can also be set to 0.7 or 0.55, etc., based on clinical experience or model performance test results, to adapt to different clinical needs.

[0101] If a doctor chooses to regenerate the report, they select a feedback reason from a preset category. The system uses this feedback data for iterative fine-tuning of the Large Language Model (LLM). This technical feature describes a continuous optimization mechanism for the Large Language Model (LLM) in the report generation module. When a doctor is dissatisfied with the clinical decision report generated by the LLM and chooses to regenerate it, the system guides the doctor to select a specific reason from the preset feedback reason categories, such as factual errors, logical flaws, or key omissions. This structured feedback data is a valuable learning resource that can be used for targeted iterative fine-tuning of the LLM, thereby continuously improving the accuracy, logic, and usability of its generated reports. This closed-loop feedback mechanism is key to achieving intelligent evolution of the system. Feedback reason categories can include, but are not limited to: incomplete information, unclear expression, and non-practical suggestions.

[0102] Postoperative data collection follows a T+1 reminder rule, establishing postoperative quality control nodes after the calculation of imaging indicators and the statistical analysis of neurological function scores. This technical feature aims to ensure the timeliness and accuracy of postoperative data collection and introduces a postoperative quality control process. The T+1 reminder rule ensures that at each preset postoperative assessment time point, the system can send an update reminder to the attending physician on the first day after the task is triggered, prompting the physician to complete the postoperative assessment in a timely manner. The establishment of postoperative quality control nodes after the calculation of imaging indicators and the statistical analysis of neurological function scores means that these key postoperative data will undergo further verification or confirmation to ensure their accuracy, providing a reliable basis for subsequent knowledge base updates and system learning. For example, postoperative quality control nodes may include cross-checking by another physician or automatic data consistency checks by the system.

[0103] The knowledge base is updated incrementally. Newly accumulated case data enables the system to form a self-learning decision support closed loop. This technical feature clarifies the updating method of the structured historical case knowledge base and the resulting improvement in system capabilities. Incremental updates mean that the knowledge base continuously receives and integrates new case data, rather than periodically replacing it completely. This approach ensures that the knowledge base remains up-to-date and is continuously enriched with the accumulation of new cases. Newly accumulated case data, especially complete case data with decision-outcome labels, provides the system with more learning samples, enabling the disease matching module to continuously optimize its matching accuracy and recommendation reliability through self-learning, thus forming a self-evolving decision support closed loop. Besides incremental updates, a combination of periodic batch updates and real-time incremental updates can also be used.

[0104] The method incorporates physician review and correction steps at each key stage, including image measurement, physical examination extraction, report review, and postoperative assessment, and records a complete workflow log for traceable quality control. This technical feature emphasizes a consistent quality control and traceability mechanism throughout the entire process. The inclusion of physician review and correction steps at multiple key stages ensures opportunities for human intervention and correction at each stage of data processing, decision generation, and outcome evaluation, significantly improving system reliability and security. Simultaneously, recording a complete workflow log, including the specific content of each review and correction and the operator's information, makes the entire decision-making process traceable, which is crucial for clinical responsibility determination and system performance auditing. The operation log may include timestamps, operation type, operator ID, and data comparisons before and after modification.

[0105] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery, characterized in that, include: The data acquisition module is used for automatic screening of target patients and parallel acquisition of multi-source clinical data; The feature extraction module is used to standardize and extract features from the multi-source clinical data, and output structured feature vectors and standardized neurological function assessment information. The disease matching module is used to build a structured historical case knowledge base, perform deep similarity matching between target patients and historical cases, and output prognostic comparison data. The report generation module is used to generate a clinical decision report based on the prognostic comparison data using a large language model, and to optimize the clinical decision report and iterate the large language model based on doctor feedback; The closed-loop feedback module is used to automatically collect, associate and integrate postoperative multi-time point outcome data, and deposit the integrated data into the structured historical case knowledge base to form a closed-loop feedback between decision-making and results.

2. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 1, characterized in that, The data acquisition module has a built-in logically related two-layer rule filtering engine, which includes a surgical indication screening layer and a patient physiological and cooperation screening layer. The surgical indication screening layer is used to automatically retrieve the corresponding surgical request form in response to the clinical event of the anesthesia assessment form submission. Based on the ICD-9-CM-3 surgical operation code and the ICD-10 disease diagnosis code, it selects an initial patient cohort that meets the criteria for Class 01 craniotomy, brain and meningotomy and diagnosis of C71 brain malignant tumor. The patient physiological and cooperation screening layer is used to analyze the preoperative anesthesia assessment form based on the initial patient cohort and screen out target patients who are conscious and have excellent doctor-patient cooperation assessment. The data acquisition module is also used to concurrently trigger multi-source data acquisition tasks after locking onto the target patient, and to simultaneously capture the target patient's recent surgical date head multimodal MRI image data from the image archiving communication system and the current hospitalization unstructured text medical record data from the electronic medical record system.

3. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 1, characterized in that, The feature extraction module includes an image intelligent quantization and feature extraction submodule and a text structured parsing and information extraction submodule; The image intelligent quantization and feature extraction submodule is used to construct a fully automated standardized processing pipeline for preoperative and postoperative cranial MRI images. It sequentially performs spatial registration and standardization of multi-sequence images, image segmentation based on deep learning, automatic calculation of three-dimensional volume and spatial coordinate annotation, quantitative output and visualization rendering of key image biometric indicators, and also supports doctors to review and fine-tune through a dedicated quality control interface, outputting structured feature vectors. The text structuring parsing and information extraction submodule is used to introduce a large language model to perform deep semantic understanding and key information extraction on unstructured text medical records. Following the principle of the last record closest to the surgery on the timeline, it extracts neurological examination items, maps and fills the extracted information into a preset postoperative neurological assessment single structured data template to generate an assessment draft, and pushes a review reminder to the doctor before the surgery time node.

4. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 1, characterized in that, The disease matching module is used to construct a structured historical case knowledge base that includes basic clinical information of patients, intraoperative awakening implementation tags, and key tags for various postoperative prognostic outcomes; The disease matching module is further used to perform refined matching calculations between target patients and historical cases based on a preset multi-dimensional weighted similarity algorithm. The quantitative evaluation formula for the multi-dimensional weighted similarity algorithm is as follows: Disease similarity score = Σ(feature similarity × weight); The disease matching module is also equipped with a matching output mechanism, which is used to calculate the similarity score between the target patient and the intraoperative awakening group and the non-intraoperative awakening group in the historical database, respectively, and to filter the high similarity case subset according to the preset threshold and output the corresponding structured prognostic statistical indicators.

5. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 4, characterized in that, The core feature terms and weight configurations of the multi-dimensional weighted similarity algorithm are as follows: The weightings for tumor location (40%), preoperative tumor volume (25%), age group (10%), gender (5%), motor function (5%), language function (5%), sensory function (5%), and higher cortical function (5%) were as follows: The motor function was assessed using the MRC muscle strength classification, the language function was assessed using the BNT naming test, the sensory function was assessed using the light touch / pinprick sensation score, and the higher cortical function was assessed using the MoCA score. The preset threshold is ≥0.6, and the disease matching module is used to sort the high similarity case subset in descending order of similarity.

6. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 4, characterized in that, The structured prognostic statistical indicators include four categories: case size statistics, baseline characteristic distribution, surgical outcome indicators, and functional prognostic comparison. The case size statistics are the number of matched cases and their proportion in the corresponding historical group. The baseline characteristic distribution includes tumor location, preoperative tumor volume, age group, and gender composition. The surgical outcome indicators are tumor resection rate, fiber bundle damage rate, and total volume of new edema / damage area. The functional prognostic comparison is a quantitative comparison of motor, language, sensory, and higher cortical functions before and after surgery.

7. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 1, characterized in that, The report generation module is used to receive and integrate the comparison data of the target patient and the high similarity intraoperative awakening group and non-intraoperative awakening group output by the disease matching module, and inject clinical analysis task instructions into the large language model through prompt word engineering; The report generation module is also used to drive the large language model to generate a comparative analysis report on intraoperative awakening strategies and neurological prognosis, which includes visual comparison tables, in-depth data interpretation and clinical inference, and targeted anesthesia clinical recommendations. The report generation module features an interactive interface with two-way selection buttons for adoption and regeneration. It also automatically records interactive data and calculates the clinical adoption accuracy of the large model output. The formula for calculating the clinical adoption accuracy is as follows: Adoption accuracy rate = (Number of adoptions / (Number of adoptions + Number of regenerations)) × 100%.

8. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 7, characterized in that, The regenerated options are configured with preset feedback reason categories, which include factual errors, logical flaws, key omissions, unfounded fabrications, deviation from instructions, inoperability, and others. The report generation module is used to use the recorded interaction data as a key basis for iterative fine-tuning of the prompt word strategy or large language model, driving the continuous evolution of the system's decision support capabilities.

9. The intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery according to claim 3, characterized in that, The closed-loop feedback module is used to automatically start a multi-threaded data acquisition and processing process after the operation is completed, and sequentially execute the steps of demographic and key decision information collection, automatic quantification of multi-time point imaging prognostic indicators, structured acquisition of neurological function prognostic data, data association and integration and knowledge base update, and the formation of a self-evolving decision support closed loop. When performing the automatic quantification step of multi-time point imaging prognostic indicators, the closed-loop feedback module calls the fully automatic standardization processing pipeline of the intelligent image quantization and feature extraction submodule.

10. A method for intelligent triggering, matching, and closed-loop quality control of intraoperative awakening decisions in brain tumor surgery, characterized in that, The method is executed based on the intelligent triggering, matching, and closed-loop quality control system for intraoperative awakening decision-making in brain tumor surgery as described in any one of claims 1 to 9, and includes the following steps: Step S1: In response to the clinical event of the anesthesia assessment form submission, the two-layer rule screening engine automatically screens and locks in target patients who meet the requirements of surgical indications and physiological compliance. Step S2: The intelligent image quantization and feature extraction submodule is started in parallel to process MRI images and extract quantization features; the text structure parsing and information extraction submodule is started to process text medical records and extract structured physical examination information. Step S3: Calculate the disease similarity score between the target patient and historical cases based on the multi-dimensional weighted similarity algorithm, and output the high similarity control cases in the intraoperative awakening group and the non-intraoperative awakening group, as well as their structured prognostic statistics. Step S4: Use a large language model to transform the matching data into a comparative analysis report of intraoperative awakening strategies and neurological function prognosis, which includes a visual comparison table, in-depth interpretation, and clinical suggestions. Record the doctor's adoption or regenerate feedback and calculate the clinical adoption accuracy. Step S5: At multiple preset follow-up time points after surgery, automatically collect the patient's follow-up MRI images and clinical assessment data of neurological function, call the image processing pipeline to calculate key postoperative imaging indicators, and calculate the change value of neurological function score relative to the preoperative baseline. Step S6: Link and integrate postoperative outcome data with intraoperative decision tags with preoperative feature data and intraoperative decision records, and deposit the integrated complete case data into a structured historical case knowledge base to update and enrich the knowledge base of similarity matching.

Images