An atrial fibrillation precise treatment auxiliary decision system based on a large language model

Abstract

The application relates to the technical field of medical auxiliary decision-making, in particular to a precise atrial fibrillation treatment auxiliary decision-making system based on a large language model, which comprises the following modules: a data acquisition module for acquiring patient medical data; an encryption module for dynamically performing AES encryption on sensitive information; a data processing and calculation analysis module for intelligently cleaning and standardizing the data based on rules and a knowledge base to generate structured analysis data; a large language model analysis engine for performing reasoning analysis on the structured data by using machine-readable atrial fibrillation guidelines and clinical path rules to generate decision support information containing recommendations, basis and warnings; and a report generation module for integrating the decision information and the data into a visual clinical decision-making auxiliary report. The application realizes the full-process automation from the secure management of multi-source heterogeneous medical data, the fusion reasoning of evidence-based knowledge to the intuitive report generation, and the precision, standardization and clinical work efficiency of the atrial fibrillation diagnosis and treatment decision-making are improved.

Description

Technical Field

[0001] This invention relates to the field of medical decision support technology, and in particular to a decision support system for the precise treatment of atrial fibrillation based on a large language model. Background Technology

[0002] In current clinical practice, atrial fibrillation (AF) treatment decisions heavily rely on physicians' integration and analysis of massive amounts of heterogeneous patient data from multiple sources, as well as the application of complex guidelines and rules. Traditional Clinical Decision Support Systems (CDSS) are mostly based on static rule engines, which struggle to effectively handle unstructured text, long-term monitoring waveforms, and other data, resulting in "data silos" and "information overload," leading to inefficient decision-making and the potential for overlooking crucial information. Furthermore, existing systems often lack robust security protection for sensitive medical information throughout the entire data preprocessing process, and their data cleaning and standardization processes are insufficiently intelligent, making it difficult to guarantee the quality of data input to the analysis engine. More importantly, the reasoning process of traditional systems is opaque, failing to provide coherent explanations that align with clinical thinking and are based on evidence, resulting in limited trust and adoption rates among physicians for their recommendations. Therefore, there is an urgent need for an auxiliary system that deeply integrates secure data governance, intelligent information extraction, interpretable reasoning based on authoritative medical knowledge, and the generation of intuitive decision reports using large language models to improve the accuracy and standardization of AF treatment. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides a decision support system for the precise treatment of atrial fibrillation based on a large language model.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention discloses a decision support system for precision treatment of atrial fibrillation based on a large language model, comprising: The data acquisition module is used to acquire patient medical data; The encryption module is used to perform AES encryption on sensitive information in the patient's medical data; The data processing and computational analysis module is used to clean and standardize the patient's medical data to generate structured analysis data. A large language model analysis engine is used to receive the structured data, perform reasoning analysis on the structured data through atrial fibrillation guidelines and clinical pathway rules, and generate decision support information. The report generation module is used to integrate the structured data with the decision support information to generate a visualized clinical decision support report.

[0005] The patient medical data includes basic patient information, clinical diagnosis and medical history data, multimodal examination and monitoring data, long-term physiological monitoring data, and treatment and follow-up data.

[0006] Furthermore, sensitive information in the patient's medical data is encrypted using AES, specifically as follows: From the collected raw patient medical data stream, fields containing personal identifiers and treatment conclusions are identified and separated in real time as plaintext to be encrypted; The plaintext to be encrypted is formatted and length-padded to form a standardized data block to be encrypted; Generate a random obfuscation factor uniquely bound to this data operation, and obfuscate the random obfuscation factor with the data block to be encrypted to obtain primary obfuscated data; Using a key derivation function based on a cryptographically secure pseudo-random number generator, combined with the system master key and the aforementioned random confusion factor, a dynamic session key and initialization vector specifically for this encryption are derived. Using the dynamic session key and initialization vector, the primary obfuscated data is encrypted, and ciphertext and integrity authentication tags are generated simultaneously. The obfuscation factor, the initialization vector, the integrity authentication tag, and the ciphertext are assembled in a preset order to construct a final secure data packet that can independently complete decryption and integrity verification.

[0007] Specifically, the plaintext to be encrypted is formatted and padded to form a standardized data block to be encrypted, as follows: Obtain and parse the database associated with the standardized medical terminology system, obtain the semantic type, length constraints and encoding rules of the sensitive data fields defined therein, and generate field formatting templates; Based on the field formatting template, each data field in the plaintext to be encrypted is converted into a standardized data element conforming to the HL7FHIR resource format and its version identifier is marked. Based on the version identifier and semantic type, a type code is assigned to each standardized data element, and all type codes are sorted according to a preset priority to generate a type code sequence; The standardized data elements are rearranged and concatenated according to their corresponding type codes in the type code sequence to generate intermediate concatenated data. Calculate the original byte length of the intermediate concatenated data, and determine the target padding length based on the requirements of the AES block cipher algorithm; Using the PKCS#7 padding scheme, padding bytes are added to the end of the intermediate concatenated data until its length reaches the target padding length, thereby generating the standardized data block to be encrypted.

[0008] Furthermore, the patient's medical data is cleaned and standardized to generate structured analysis data, specifically as follows: Receive the original patient medical data set decrypted and restored from the encryption module, and load the preset clinical data governance rule base; Based on the data integrity, logical consistency, and value range compliance rules in the clinical data governance rule base, the original data set is scanned to generate a data quality tag vector with error type labels; Based on the data quality marker vector, locate and extract all data items with logical conflicts or abnormal value ranges to form a queue to be repaired. The built-in medical terminology standardization engine is invoked, and based on the standard terminology mapping dictionary in the clinical data governance rule base, the non-standard terms in the queue to be repaired are conceptually mapped and encoded for replacement, generating a terminology unification queue. Using the medical logic verification rules defined in the clinical data governance rule base, contextual logic reasoning and contradiction resolution are performed on the data in the terminology unification queue, and logically compliant corrected data is output. The corrected data is reassembled according to a preset clinical data model and associated with its corresponding data source identifier and timestamp to generate standardized intermediate data objects. Based on the entity relationships defined in the clinical data model, multiple intermediate data objects are associated, integrated, and serialized to output the structured analysis data for use by the large language model analysis engine.

[0009] Specifically, based on the data quality marker vector, all data items with logical conflicts or abnormal value ranges are located and extracted to form a queue to be repaired. The data quality marker vector is parsed to identify sites marked as logical conflicts or value range anomalies, and an index list of anomaly sites is generated. Load a pre-defined medical data anomaly pattern knowledge base, which defines the pattern characteristics and associated data field rules for various anomalies; Based on the abnormal site index list, the corresponding original data segment is located in the original data set, and pattern matching is performed according to the medical data abnormal pattern knowledge base to identify the specific abnormal pattern to which each segment belongs, generating a set of abnormal data items with pattern labels. For each data item in the set of abnormal data items, the corresponding value range correction rule is extracted from the medical data abnormality pattern knowledge base according to its pattern label, and a strategy-bound data item is generated. Iterate through all strategy-bound data items, distinguish them from "clinical diagnostic assertions" or "objective record data" based on their clinical assertion context, and attach a category label to each data item; Based on the category labels and pattern labels, a correlation clustering analysis is performed on all strategy-bound data items to group different abnormal data items involving the same clinical event or the same physiological parameter into the same repair unit, forming a repair unit set. The repair unit set is sorted according to a preset priority rule, and the sorted repair unit sequence is output as a structured queue to be repaired.

[0010] Furthermore, the structured data is analyzed and reasoned through atrial fibrillation guidelines and clinical pathway rules to generate decision support information, specifically: The structured analysis data is parsed to identify the standardized data elements contained therein and the medical semantic relationships between them; The standardized data elements are mapped to corresponding clinical entities and numerical features, and an initial patient clinical profile is constructed based on the medical semantic relationships. Using the clinical entity as the search key, the atrial fibrillation guidelines are traversed to obtain the pathophysiological nodes, treatment intervention nodes, and potential complication nodes directly associated with them, generating a candidate medical concept set. Simultaneously, based on the numerical characteristics and the threshold conditions in the clinical pathway rules, the matching degree of the current patient relative to each standard pathway is calculated, and its main pathway anchor point and deviation status are determined. The candidate medical concept set is fused with the main path anchor point and deviation state to generate an enhanced clinical context vector; Using the decision logic tree defined in the clinical pathway rules, reasoning is performed on the enhanced clinical context vector to disambiguate and filter out the treatment options and risk assessment results that meet the conditions, forming structured decision elements; Based on the evidence-based relationship links in the atrial fibrillation guidelines, the structured decision elements are subjected to interpretive correlation and confidence fusion to generate the decision support information containing recommended solutions, justifications, and warnings.

[0011] Furthermore, the structured data is integrated with the decision support information to generate a visualized clinical decision support report, specifically: The recommended solutions, their basis, and warnings in the decision support information are extracted as core decision elements. The structured data is parsed, and the clinical entities and numerical features therein are mapped to original report fragments according to a preset clinical report template; Based on the semantic association rules between the core decision elements and the clinical report template, the core decision elements are dynamically inserted into the original report fragment to generate primary integrated content containing a decision-evidence chain; The initial integrated content is optimized for readability to generate a narrative text summary for clinicians; Extract the temporal physiological parameters from the structured data and the risk assessment results from the decision support information, perform multi-scale feature extraction and periodic pattern analysis on the temporal physiological parameters, and perform probability distribution and sensitivity analysis on the risk assessment results. Encode the analysis results into graphical syntax instructions to generate dynamic trend charts and risk matrix charts. The narrative text summary, the dynamic trend chart, and the risk matrix chart are arranged spatially and visually associated according to the layout rules of the clinical report template to generate a visual report draft. The confidence level and clinical pathway anchor points attached to the decision support information are embedded as metadata, and the visualized clinical decision support report is output.

[0012] This invention addresses the technical deficiencies in the prior art and offers the following advantages: By integrating standardized medical terminology and data models, and executing an intelligent governance process that includes abnormal pattern recognition and contextual cleansing, the completeness, consistency, and semantic standardization of input data are improved, providing a reliable data foundation for accurate analysis. Furthermore, by deeply integrating structured patient data with machine-readable authoritative atrial fibrillation guidelines and clinical pathway rules, it can generate evidence-based, personalized decision support information containing explanations and warnings, thereby improving the standardization, accuracy, and clinical acceptability of treatment recommendations. Finally, by dynamically integrating decision logic and patient data into a visualized report containing a "decision-evidence chain," key conclusions and risk trends are presented intuitively in a graphic and textual format, reducing the information interpretation and decision-making burden on clinicians and improving diagnostic efficiency and quality. The entire system achieves end-to-end intelligent support from raw data to clinical decision assistance, helping to improve the accuracy, consistency, and efficiency of atrial fibrillation diagnosis and treatment. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0014] Figure 1 This is a module framework diagram of the system; Figure 2 This is a flowchart of the overall technical process of this system; Figure 3 This is a diagram showing the learning stages of the large language model in this system. Detailed Implementation

[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0017] like Figure 1 , Figure 2 as well as Figure 3 As shown, this invention discloses a decision support system for precise treatment of atrial fibrillation based on a large language model, comprising: The data acquisition module is used to acquire patient medical data; The encryption module is used to perform AES encryption on sensitive information in the patient's medical data; The data processing and computational analysis module is used to clean and standardize the patient's medical data to generate structured analysis data. A large language model analysis engine is used to receive the structured data, perform reasoning analysis on the structured data through atrial fibrillation guidelines and clinical pathway rules, and generate decision support information. The report generation module is used to integrate the structured data with the decision support information to generate a visualized clinical decision support report.

[0018] The patient medical data includes basic patient information, clinical diagnosis and medical history data, multimodal examination and monitoring data, long-term physiological monitoring data, and treatment and follow-up data.

[0019] Furthermore, sensitive information in the patient's medical data is encrypted using AES, specifically as follows: From the collected raw patient medical data stream, fields containing personal identifiers and treatment conclusions are identified and separated in real time as plaintext to be encrypted; The plaintext to be encrypted is formatted and length-padded to form a standardized data block to be encrypted; Generate a random obfuscation factor uniquely bound to this data operation, and obfuscate the random obfuscation factor with the data block to be encrypted to obtain primary obfuscated data; Using a key derivation function based on a cryptographically secure pseudo-random number generator, combined with the system master key and the aforementioned random confusion factor, a dynamic session key and initialization vector specifically for this encryption are derived. Using the dynamic session key and initialization vector, the primary obfuscated data is encrypted, and ciphertext and integrity authentication tags are generated simultaneously. The obfuscation factor, the initialization vector, the integrity authentication tag, and the ciphertext are assembled in a preset order to construct a final secure data packet that can independently complete decryption and integrity verification.

[0020] Specifically, upon receiving the original patient medical data stream, the system identifies and separates sensitive fields such as "personal identifiers" (e.g., name, ID number) and "treatment conclusions" as plaintext to be encrypted in real time. To achieve efficient and secure encryption, this plaintext is preprocessed: based on the field semantics, length, and encoding rules defined in the meta-model, the data is converted into standardized data elements and sorted. Then, standard padding schemes such as PKCS#7 are used to form data blocks of regular length to be encrypted. To resist replay attacks and ensure the uniqueness of each encryption, a random obfuscation factor (e.g., generated from timestamps, operation sequence numbers, etc.) uniquely bound to this data operation (e.g., a data upload transaction for a specific patient) is generated. This factor is then obfuscated with the data block to be encrypted (e.g., by XOR or concatenation) to obtain primary obfuscated data, effectively increasing the randomness of the ciphertext.

[0021] Next, the encryption module employs a key derivation function based on a cryptographically secure pseudo-random number generator (CSPRNG) (such as a KDF based on HMAC). Using the system's securely stored master key and the aforementioned random obfuscation factor as input, it derives a dynamic session key and initialization vector (IV) specific to this session, achieving "one-time pad" and enhancing key security. Using this dynamic session key and IV, the initial obfuscated data is encrypted using an AES encryption algorithm (such as AES-GCM mode). This mode simultaneously generates ciphertext and an authentication tag for integrity verification. Finally, all necessary components (obfuscation factor, IV, authentication tag, ciphertext) are assembled in a preset fixed order to form a self-contained final secure data packet. This data packet design allows for independent decryption and data integrity verification at the decryption end, without relying on external additional information, using only this data packet and the system master key. This ensures the confidentiality, integrity, and traceability of medical data during transmission and storage, effectively meeting the high-level security protection and privacy compliance requirements for medical and health data.

[0022] Specifically, the plaintext to be encrypted is formatted and padded to form a standardized data block to be encrypted, as follows: Obtain and parse the database associated with the standardized medical terminology system, obtain the semantic type, length constraints and encoding rules of the sensitive data fields defined therein, and generate field formatting templates; Based on the field formatting template, each data field in the plaintext to be encrypted is converted into a standardized data element conforming to the HL7FHIR resource format and its version identifier is marked. Based on the version identifier and semantic type, a type code is assigned to each standardized data element, and all type codes are sorted according to a preset priority to generate a type code sequence; The standardized data elements are rearranged and concatenated according to their corresponding type codes in the type code sequence to generate intermediate concatenated data. Calculate the original byte length of the intermediate concatenated data, and determine the target padding length based on the requirements of the AES block cipher algorithm; Using the PKCS#7 padding scheme, padding bytes are added to the end of the intermediate concatenated data until its length reaches the target padding length, thereby generating the standardized data block to be encrypted.

[0023] It should be noted that the purpose of this step is to transform raw sensitive fields from diverse sources and with varying formats into data blocks with a unified structure, regular length, and clear semantics to meet the input requirements of AES encryption. Specifically, this involves acquiring and parsing a database associated with a standardized medical terminology system. This database refers to a knowledge base or metadata storage system that integrates internationally recognized medical terminology standards such as SNOMED CT (Systematic Clinical Medical Terminology) and LOINC (Logical Naming and Coding System for Observational Indicators), and defines the mapping relationships between these standards and specific clinical data fields (such as "patient name," "diagnosis," and "drug code"). By parsing this database, the system can clearly identify the standardized semantic type (such as "Person.Name," "Condition.Code"), the maximum / minimum length constraints of its value, and the specific encoding rules to be followed (such as the date format being YYYY-MM-DD), and generate a structured field formatting template accordingly.

[0024] Based on this template, each raw data field in the plaintext to be encrypted is converted into standardized data elements conforming to the HL7 FHIR (Rapid Medical Interoperability Resource) resource format. For example, a free-text diagnostic description is mapped and encapsulated into an FHIR Observation resource fragment containing standard LOINC codes, display text, and status information, while also marking the FHIR version identifier it follows (such as R4). Then, based on the data element's version identifier and semantic type, a unique type code is assigned to each data element, and all type codes are sorted according to clinical importance or processing priority (e.g., personal identifiers take precedence over general medical data) to generate a type code sequence.

[0025] Based on the sequence, all standardized data elements are rearranged and concatenated to generate intermediate concatenated data. The original byte length of this intermediate concatenated data is calculated, and the target padding length is determined according to the block size requirements of the selected AES algorithm (such as CBC mode, typically a multiple of 16 bytes). The system uses the industry-standard PKCS#7 padding scheme, adding the appropriate number of padding bytes to the end of the intermediate concatenated data to generate a standardized data block to be encrypted with a length fully conforming to the encryption algorithm's requirements.

[0026] Furthermore, the patient's medical data is cleaned and standardized to generate structured analysis data, specifically as follows: Receive the original patient medical data set decrypted and restored from the encryption module, and load the preset clinical data governance rule base; Based on the data integrity, logical consistency, and value range compliance rules in the clinical data governance rule base, the original data set is scanned to generate a data quality tag vector with error type labels; Based on the data quality marker vector, locate and extract all data items with logical conflicts or abnormal value ranges to form a queue to be repaired. The built-in medical terminology standardization engine is invoked, and based on the standard terminology mapping dictionary in the clinical data governance rule base, the non-standard terms in the queue to be repaired are conceptually mapped and encoded for replacement, generating a terminology unification queue. Using the medical logic verification rules defined in the clinical data governance rule base, contextual logic reasoning and contradiction resolution are performed on the data in the terminology unification queue, and logically compliant corrected data is output. The corrected data is reassembled according to a preset clinical data model and associated with its corresponding data source identifier and timestamp to generate standardized intermediate data objects. Based on the entity relationships defined in the clinical data model, multiple intermediate data objects are associated, integrated, and serialized to output the structured analysis data for use by the large language model analysis engine.

[0027] It should be noted that a pre-defined clinical data governance rule base is loaded. This rule base is a set of rules built upon medical knowledge systems, industry standards (such as the structured requirements of HL7 CDA / FHIR), and clinical practice consensus. It clearly defines multi-dimensional verification rules for data integrity (such as key fields not being empty), logical consistency (such as "discharge date" not being earlier than "admission date"), and value range compliance (such as "systolic blood pressure" values ​​being within a reasonable physiological range). Based on these rules, the original dataset is automatically scanned to identify missing, contradictory, or abnormal data points, and a data quality tag vector with an error type label is generated for each data item, thereby quantifying and locating data quality issues.

[0028] Based on this vector, all abnormal data items are located and extracted to form a queue to be repaired. To address the common problem of inconsistent terminology in the queue (e.g., the same diagnosis is recorded as "atrial fibrillation," "heart fibrillation," or "AF" in different systems), the system calls the built-in medical terminology standardization engine. This engine integrates software components that map relationships of international standard medical terminology systems (such as SNOMED CT for clinical concepts, LOINC for laboratory observation indicators, and ICD for disease classification). Using a standard terminology mapping dictionary in the rule base, it performs concept recognition and code replacement on free text or hospital codes, generating a unified terminology queue to ensure that all medical concepts are represented by standardized codes.

[0029] By leveraging deeper medical logic validation rules from the rule base (e.g., "Patients receiving warfarin treatment should regularly record their INR values"), contextual logic reasoning is performed on the terminology-standardized data to identify and resolve potential clinical inconsistencies (such as conflicting medication records), outputting logically compliant corrected data. Subsequently, following a pre-defined clinical data model that characterizes clinical entities and their relationships (e.g., based on the OMOP general data model), the corrected data is reassembled into standardized intermediate data objects carrying data sources and timestamps. Based on the entity relationships defined in this model (e.g., the "patient-visit-diagnosis-medication" chain), multiple intermediate data objects are associated, integrated, and serialized, outputting a highly structured, semantically clear structured analytical data set that directly serves the deep processing and reasoning of the large language model analysis engine.

[0030] Specifically, based on the data quality marker vector, all data items with logical conflicts or abnormal value ranges are located and extracted to form a queue to be repaired. The data quality marker vector is parsed to identify sites marked as logical conflicts or value range anomalies, and an index list of anomaly sites is generated. Load a pre-defined medical data anomaly pattern knowledge base, which defines the pattern characteristics and associated data field rules for various anomalies; Based on the abnormal site index list, the corresponding original data segment is located in the original data set, and pattern matching is performed according to the medical data abnormal pattern knowledge base to identify the specific abnormal pattern to which each segment belongs, generating a set of abnormal data items with pattern labels. For each data item in the set of abnormal data items, the corresponding value range correction rule is extracted from the medical data abnormality pattern knowledge base according to its pattern label, and a strategy-bound data item is generated. Iterate through all strategy-bound data items, distinguish them from "clinical diagnostic assertions" or "objective record data" based on their clinical assertion context, and attach a category label to each data item; Based on the category labels and pattern labels, a correlation clustering analysis is performed on all strategy-bound data items to group different abnormal data items involving the same clinical event or the same physiological parameter into the same repair unit, forming a repair unit set. The repair unit set is sorted according to a preset priority rule, and the sorted repair unit sequence is output as a structured queue to be repaired.

[0031] Specifically, the process begins by parsing the data quality marker vector to generate an index list of anomalous sites that identify their locations. Then, a pre-defined medical data anomaly pattern knowledge base is loaded. This knowledge base is a rule system built upon clinical guidelines, medical logic, and historical data quality analysis to systematically define the pattern characteristics of various common anomalies and their associated field rules. For example, for the "contradictory vital signs" pattern, its characteristics might be associated with fields such as "heart rate," "blood pressure," and "recording time," and "resting heart rate greater than 200 beats / min" is defined as a specific pattern instance.

[0032] After locating the original data fragment based on the index list, pattern matching is performed using the aforementioned knowledge base to identify "value range anomalies" (e.g., blood glucose value 999) or "logical conflicts" (e.g., "diagnosis: diabetes" but "glycated hemoglobin: 4.0%)), and to determine whether they belong to specific anomaly patterns such as "extreme value anomalies," "contradictory assertions," or "temporal anomalies." A pattern label is then attached to each anomaly data item, forming a set of anomaly data items with semantic understanding. Based on the pattern labels, pre-configured value range correction rules are extracted from the knowledge base. These value range correction rules are automated or semi-automated processing strategies preset for specific anomaly patterns. For example, for "extreme value anomalies," the rule is "replace with a recent reasonable value from the same patient" or "mark as requiring manual review"; for "unit inconsistency" contradictions, the rule is "convert imperial units to international standard units based on the field identifier."

[0033] To further optimize the repair logic, all data items with bound repair strategies are traversed and categorized into "clinical diagnostic assertions" or "objective recorded data" based on their clinical context (e.g., whether from a doctor's diagnostic record or device monitoring output), with corresponding category labels. This distinction is crucial because the correction of subjective diagnostic assertions requires greater caution and often necessitates corroboration with more evidence. Subsequently, correlation clustering analysis is performed based on category and pattern labels, grouping multiple anomalies involving the same clinical event (e.g., a specific visit) or the same physiological parameter (e.g., continuous blood pressure monitoring) into the same repair unit. For example, contradictory "diagnosis" and "medication" anomalies from the same visit are clustered together for easier overall assessment and correction. Finally, the repair unit set is sorted according to predefined priority rules (e.g., anomalies affecting treatment safety take priority, and objective data anomalies take priority over subjective assertion anomalies), outputting a clearly structured and prioritized queue of items to be repaired.

[0034] Furthermore, the structured data is analyzed and reasoned through atrial fibrillation guidelines and clinical pathway rules to generate decision support information, specifically: The structured analysis data is parsed to identify the standardized data elements contained therein and the medical semantic relationships between them; The standardized data elements are mapped to corresponding clinical entities and numerical features, and an initial patient clinical profile is constructed based on the medical semantic relationships. Using the clinical entity as the search key, the atrial fibrillation guidelines are traversed to obtain the pathophysiological nodes, treatment intervention nodes, and potential complication nodes directly associated with them, generating a candidate medical concept set. Simultaneously, based on the numerical characteristics and the threshold conditions in the clinical pathway rules, the matching degree of the current patient relative to each standard pathway is calculated, and its main pathway anchor point and deviation status are determined. The candidate medical concept set is fused with the main path anchor point and deviation state to generate an enhanced clinical context vector; Using the decision logic tree defined in the clinical pathway rules, reasoning is performed on the enhanced clinical context vector to disambiguate and filter out the treatment options and risk assessment results that meet the conditions, forming structured decision elements; Based on the evidence-based relationship links in the atrial fibrillation guidelines, the structured decision elements are subjected to interpretive correlation and confidence fusion to generate the decision support information containing recommended solutions, justifications, and warnings.

[0035] It should be noted that, based on the structured analysis data obtained from the previous steps, the medical concepts represented by standardized codes (such as SNOMED CT, LOINC) and their relationships (such as "having" and "causing") are identified, and an initial clinical profile of the patient is constructed, which includes the patient's demographic characteristics, symptoms, signs, examination results, comorbidities and previous treatments.

[0036] The atrial fibrillation guidelines refer to a knowledge base of clinical practice guidelines published by authoritative medical institutions (such as the American Heart Association (AHA) and the European Society of Cardiology (ESC)) and processed in a structured, machine-readable manner. It systematically encodes evidence-based recommendations regarding atrial fibrillation classification, stroke risk assessment (such as CHA2DS2-VASc), rhythm and heart rate control strategies, indications for anticoagulation therapy, and ablation surgery recommendations. The clinical pathway rules refer to standardized treatment process definitions based on the aforementioned guidelines and combined with the specific medical institution's treatment procedures. These definitions include clear admission criteria, key decision points, temporal task chains, and outcome evaluation thresholds, and are formally expressed as decision logic trees, state machines, or rule sets.

[0037] During inference, the engine performs dual alignment: On one hand, using entities in the patient's clinical profile (e.g., "heart failure") as keys, it traverses the guideline knowledge base to extract all relevant pathological mechanisms, recommended drugs (e.g., "ARNI"), and risk warnings (e.g., "bleeding risk"), generating a candidate medical concept set. On the other hand, it matches the patient's numerical characteristics (e.g., "LVEF=35%)" with threshold conditions in the clinical pathway rules to determine the patient's most suitable standard pathway anchor point (e.g., "heart failure with atrial fibrillation management pathway") and specific deviation status (e.g., "meets ablation assessment criteria"). The candidate concept set is then fused with the pathway anchor point and deviation status to form an enhanced clinical context vector representing the current patient's specific situation. Using the decision logic tree explicitly defined in the clinical pathway rules, this vector is used for automated inference, filtering out treatment options and risk assessment results that meet guideline recommendations at specific pathway nodes, forming structured decision elements.

[0038] Finally, by tracing back the evidence-based relationships in the atrial fibrillation guidelines (such as citing relevant clinical trial results), each decision element is accompanied by the source of evidence, recommendation level (and potential conflict warnings). Through confidence fusion, the final decision support information is generated, providing clear medical evidence and safety warnings. This achieves transparent and credible auxiliary decision-making, effectively supporting clinicians in developing individualized treatment plans.

[0039] Furthermore, the structured data is integrated with the decision support information to generate a visualized clinical decision support report, specifically: The recommended solutions, their basis, and warnings in the decision support information are extracted as core decision elements. The structured data is parsed, and the clinical entities and numerical features therein are mapped to original report fragments according to a preset clinical report template; Based on the semantic association rules between the core decision elements and the clinical report template, the core decision elements are dynamically inserted into the original report fragment to generate primary integrated content containing a decision-evidence chain; The initial integrated content is optimized for readability to generate a narrative text summary for clinicians; Extract the temporal physiological parameters from the structured data and the risk assessment results from the decision support information, perform multi-scale feature extraction and periodic pattern analysis on the temporal physiological parameters, and perform probability distribution and sensitivity analysis on the risk assessment results. Encode the analysis results into graphical syntax instructions to generate dynamic trend charts and risk matrix charts. The narrative text summary, the dynamic trend chart, and the risk matrix chart are arranged spatially and visually associated according to the layout rules of the clinical report template to generate a visual report draft. The confidence level and clinical pathway anchor points attached to the decision support information are embedded as metadata, and the visualized clinical decision support report is output.

[0040] It should be noted that decision elements are extracted from the decision support information, and the structured patient data is parsed. Based on a pre-defined report template conforming to clinical documentation standards (such as SOAP format: subjective, objective, assessment, and planning), the patient's clinical entities and numerical characteristics are automatically filled in to generate original report fragments (such as the vital signs list in the "objective" section). Based on predefined semantic association rules (e.g., associating "anticoagulation recommendation" with the "planning" section of the report and related "laboratory test results"), core decision elements are dynamically inserted into the corresponding original report fragments to form primary integrated content. This process is not a simple splicing, but rather constructs a clear "decision-evidence chain," ensuring that each recommendation is supported by adjacent data. Next, the primary content is optimized for readability, for example, by translating coded medical concepts (such as "Drug:Apixaban") into clinical narrative language to generate a coherent narrative text summary.

[0041] To further enhance the report's insight, the system extracts temporal physiological parameters (such as continuous heart rate monitoring data) from structured data and risk assessment results from decision-making information. Through multi-scale feature extraction (such as calculating daily average heart rate and nighttime heart rate variability) and periodic pattern analysis (identifying the diurnal rhythm of atrial fibrillation load), and presenting the risk assessment results probabilistically, the system utilizes graphical syntax commands (such as command sets based on D3.js or similar libraries) to generate visualization components such as dynamic trend charts and risk matrix diagrams.

[0042] Finally, following the layout and visual design rules of the clinical report template, the narrative text summary and various visualization charts are spatially arranged, and their inherent connections are strengthened through visual association annotations (such as highlighting conclusions in the text and corresponding data points in the charts with the same color), generating a draft visualization report. The final report will embed the confidence level and clinical pathway anchor points attached to the decision support information as searchable metadata, outputting a visualized clinical decision support report that combines clinical readability, decision traceability, and in-depth insights to serve the clinical diagnosis and treatment process.

[0043] This implementation also includes: Extract the standardized set of data elements and the logical reasoning path summary contained in the decision support information to generate a unique identifier for the decision record containing an intervention option summary vector and a decision context fingerprint. Based on the intervention option summary vector, the preset physiological indicator change trajectory and theoretical clinical event occurrence probability corresponding to each intervention option are mapped from the preset medical knowledge base, and encoded as an expected data pattern template. Based on the unique identifier of the decision record, the follow-up protocol is actively triggered, and the raw follow-up data stream containing time-series physiological signals and structured questionnaire responses is asynchronously collected from the connected portable monitoring device and the patient terminal through the interface. The data processing and computational analysis module is invoked to clean and extract features from the original follow-up data stream, generating an observation data pattern vector aligned with the dimension of the expected data pattern template. The observed data pattern vector is matched point-by-point with the expected data pattern template and the difference is calculated to generate a pattern deviation sequence. The pattern deviation sequence is then fused with the decision context fingerprint to form a traceable association evaluation vector. Using the unique identifier of the decision record as the primary key, a bidirectional index is established in the database to link the associated evaluation vector with the corresponding original decision support information, and the information is stored in association.

[0044] It should be noted that after the system generates decision support information, it extracts its core standardized data element set and a summary of the key logical reasoning path from which the conclusion is derived. Based on this information, a unique identifier for the decision record is generated using techniques such as hash algorithms, containing an intervention option summary vector and a decision context fingerprint. This identifier is unique and cannot be tampered with. According to the intervention option summary vector, the system maps the pre-set physiological indicator change trajectory (such as the target range curve of the INR value after anticoagulant medication) and the theoretical clinical event probability (such as the expected annual stroke risk reduction value under treatment) corresponding to each intervention option from a pre-set medical knowledge base containing knowledge of pharmacokinetics, disease natural history, etc. Based on this unique identifier for the decision record, a follow-up protocol is actively triggered. This protocol asynchronously collects raw follow-up data streams containing time-series physiological signals (such as daily heart rate and heart rhythm segments) and structured questionnaire responses (such as bleeding symptom self-assessment and medication adherence) from connected portable ECG monitoring devices, wearable devices, and patient terminal applications through a standardized interface.

[0045] Then, the existing data processing and analysis modules in the system are invoked to clean, denoise, and extract features from the raw follow-up data stream, generating an observation data pattern vector that is fully aligned with the expected data pattern template in terms of time dimension and indicator type. The observation data pattern vector is then matched point-by-point with the expected data pattern template, and the difference is calculated to generate a pattern deviation sequence reflecting the deviation between the actual effect and the expectation. For example, the conformity between the actual INR value sequence and the target range curve is calculated. This sequence is fused with the decision context fingerprint that identifies the initial decision intention to form a traceable correlation evaluation vector, which not only includes "how effective it was" but also "which decision it was effective against".

[0046] Ultimately, the system uses the unique identifier of the decision record as the primary key of the database, and performs bidirectional indexing and associated storage of the aforementioned associated evaluation vectors and the original decision support information generated initially. This allows clinicians or the system itself to review any historical decision at any time and view its subsequent real-world effect data, providing a data foundation for timely adjustments to treatment plans and iterative optimization of decision models. This achieves a complete closed loop from intelligent decision-making to effect verification, enhancing the dynamic adaptability and clinical credibility of the decision support system.

[0047] This implementation also includes: While the report generation module integrates and generates clinical decision support reports, it extracts the sequence of clinical path rule nodes traversed and the index of atrial fibrillation guideline provisions referenced in the reasoning analysis process output by the large language model analysis engine to generate a reasoning path fingerprint. Extract the key data source identifiers and version hashes from the structured analysis data to generate a set of data traceability identifiers; The inference path fingerprint and the data source identification set are concatenated and fused with timestamps to form an initial audit vector; and the initial audit vector is then subjected to feature compression and serialization to generate an audit summary. The audit summary is XORed with the final decision option summary contained in the decision support information and hashed to generate an immutable integrity verification anchor point. The audit summary, the integrity verification anchor point, and the session metadata generated by the current report are used together as log entries, encrypted by the encryption module, and then stored in association with the unique identifier of the clinical decision support report as an index.

[0048] It should be noted that, in order to address the common problems of opaque decision-making processes and unauditable conclusions in existing intelligent decision support systems, which lead to insufficient clinical trust and difficulties in medical quality review, this embodiment simultaneously initiates an audit information capture process while the report generation module integrates and generates a visual report. Specifically, from the internal reasoning process of the large language model analysis engine, the sequence of clinical path rule nodes traversed by its reasoning chain and the index of atrial fibrillation guideline provisions referenced are extracted. This information is encoded to generate a unique reasoning path fingerprint to depict the logical roadmap of this reasoning.

[0049] Simultaneously, from the structured analysis data upon which this conclusion was based, the corresponding key data source identifiers (such as the original ECG record ID and laboratory report number) and the calculated data version hash value were extracted to form a data traceability identifier set, ensuring that the conclusion can be traced back to the original data. The inference path fingerprint and the data traceability identifier set were concatenated and integrated with a precise timestamp to form an initial audit vector. This vector was then subjected to feature compression and serialization to generate a concise audit summary, which serves as a "summary" of the entire decision-making process.

[0050] To ensure that the audit record cannot be tampered with afterward, the audit summary is XORed with the final decision option summary in the decision support information, and a hash chain operation is performed on the result to generate a strongly associated, tamper-proof integrity verification anchor. Any tampering with the original decision or audit information will cause this anchor verification to fail.

[0051] Finally, the audit summary, integrity verification anchor, and session metadata generated by this report (such as operator ID and system version number) are encapsulated into a single log entry. This entry is encrypted using an encryption module and then stored in a secure log database, indexed by the unique identifier corresponding to the clinical decision support report. Thus, every output report possesses an independently verifiable, cryptographically guaranteed "birth certificate," achieving full transparency and auditability of the decision-making process and tamper-proof conclusions, enhancing the system's credibility and compliance in rigorous medical environments.

[0052] In summary, this invention effectively addresses the problems of low data quality, insufficient knowledge utilization, and opaque decision-making processes in existing clinical decision support systems by constructing a fully automated system that integrates secure access to multi-source heterogeneous medical data, intelligent governance, evidence-based reasoning, and visualization report generation. By integrating standardized medical terminology and data models, and executing an intelligent governance process that includes abnormal pattern recognition and contextual cleansing, the completeness, consistency, and semantic standardization of input data are improved, providing a reliable data foundation for accurate analysis. Furthermore, by deeply integrating structured patient data with machine-readable authoritative atrial fibrillation guidelines and clinical pathway rules, it can generate evidence-based, personalized decision support information that includes explanations and warnings, thereby improving the standardization, accuracy, and clinical acceptability of the decision-making system. Finally, by dynamically integrating decision logic and patient data into a visualized report containing a "decision-evidence chain," key conclusions and risk trends are presented intuitively in a graphic and textual format, reducing the information interpretation and decision-making burden on clinicians and improving diagnostic and treatment efficiency and quality. The entire system achieves end-to-end intelligent support from raw data to clinical decision assistance, helping to improve the accuracy, consistency, and efficiency of atrial fibrillation diagnosis and treatment.

[0053] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0054] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0055] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0056] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0057] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0058] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A decision support system for precision treatment of atrial fibrillation based on a large language model, characterized in that, include: The data acquisition module is used to acquire patient medical data; The encryption module is used to perform AES encryption on sensitive information in the patient's medical data; The data processing and computational analysis module is used to clean and standardize the patient's medical data to generate structured analysis data. A large language model analysis engine is used to receive the structured data, perform reasoning analysis on the structured data through atrial fibrillation guidelines and clinical pathway rules, and generate decision support information. The report generation module is used to integrate the structured data with the decision support information to generate a visualized clinical decision support report.

2. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 1, characterized in that: The patient's medical data includes basic patient information, clinical diagnosis and medical history data, multimodal examination and monitoring data, long-term physiological monitoring data, and treatment and follow-up data.

3. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 1, characterized in that, Sensitive information in the patient's medical data is encrypted using AES, specifically as follows: From the collected raw patient medical data stream, fields containing personal identifiers and treatment conclusions are identified and separated in real time as plaintext to be encrypted; The plaintext to be encrypted is formatted and length-padded to form a standardized data block to be encrypted; A random obfuscation factor uniquely bound to this data operation is generated, and the random obfuscation factor is obfuscated with the data block to be encrypted to obtain primary obfuscated data; Using a key derivation function based on a cryptographically secure pseudo-random number generator, combined with the system master key and the aforementioned random confusion factor, a dynamic session key and initialization vector specifically for this encryption are derived. Using the dynamic session key and initialization vector, the primary obfuscated data is encrypted, and ciphertext and integrity authentication tags are generated simultaneously. The obfuscation factor, the initialization vector, the integrity authentication tag, and the ciphertext are assembled in a preset order to construct a final secure data packet that can independently complete decryption and integrity verification.

4. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 3, characterized in that, The plaintext to be encrypted is formatted and padded to form a standardized data block to be encrypted, specifically: Obtain and parse the database associated with the standardized medical terminology system, obtain the semantic type, length constraints and encoding rules of the sensitive data fields defined therein, and generate field formatting templates; Based on the field formatting template, each data field in the plaintext to be encrypted is converted into a standardized data element conforming to the HL7 FHIR resource format and its version identifier is marked. Based on the version identifier and semantic type, a type code is assigned to each standardized data element, and all type codes are sorted according to a preset priority to generate a type code sequence; The standardized data elements are rearranged and concatenated according to their corresponding type codes in the type code sequence to generate intermediate concatenated data. Calculate the original byte length of the intermediate concatenated data, and determine the target padding length based on the requirements of the AES block cipher algorithm; Using the PKCS#7 padding scheme, padding bytes are added to the end of the intermediate concatenated data until its length reaches the target padding length, thereby generating the standardized data block to be encrypted.

5. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 1, characterized in that, The patient's medical data is cleaned and standardized to generate structured analysis data, specifically as follows: Receive the original patient medical data set decrypted and restored from the encryption module, and load the preset clinical data governance rule base; Based on the data integrity, logical consistency, and value range compliance rules in the clinical data governance rule base, the original data set is scanned to generate a data quality tag vector with error type labels; Based on the data quality marker vector, locate and extract all data items with logical conflicts or abnormal value ranges to form a queue to be repaired. The built-in medical terminology standardization engine is invoked, and based on the standard terminology mapping dictionary in the clinical data governance rule base, the non-standard terms in the queue to be repaired are conceptually mapped and encoded for replacement, generating a terminology unification queue. Using the medical logic verification rules defined in the clinical data governance rule base, contextual logic reasoning and contradiction resolution are performed on the data in the terminology unification queue, and logically compliant corrected data is output. The corrected data is reassembled according to a preset clinical data model and associated with its corresponding data source identifier and timestamp to generate standardized intermediate data objects. Based on the entity relationships defined in the clinical data model, multiple intermediate data objects are associated, integrated, and serialized to output the structured analysis data for use by the large language model analysis engine.

6. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 5, characterized in that, Based on the data quality marker vector, all data items with logical conflicts or abnormal value ranges are located and extracted, forming a queue to be repaired, specifically as follows: The data quality marker vector is parsed to identify sites marked as logical conflicts or value range anomalies, and an index list of anomaly sites is generated. Load a pre-defined medical data anomaly pattern knowledge base, which defines the pattern characteristics and associated data field rules for various anomalies; Based on the abnormal site index list, the corresponding original data segment is located in the original data set, and pattern matching is performed according to the medical data abnormal pattern knowledge base to identify the specific abnormal pattern to which each segment belongs, generating a set of abnormal data items with pattern labels. For each data item in the set of abnormal data items, the corresponding value range correction rule is extracted from the medical data abnormality pattern knowledge base according to its pattern label, and a strategy-bound data item is generated. Iterate through all strategy-bound data items, distinguish them as belonging to the "clinical diagnostic assertion" or "objective record data" category based on their clinical assertion context, and attach a category label to each data item; Based on the category labels and pattern labels, a correlation clustering analysis is performed on all strategy-bound data items to group different abnormal data items involving the same clinical event or the same physiological parameter into the same repair unit, forming a repair unit set. The repair unit set is sorted according to a preset priority rule, and the sorted repair unit sequence is output as a structured queue to be repaired.

7. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 1, characterized in that, The structured data is analyzed using atrial fibrillation guidelines and clinical pathway rules to generate decision support information, specifically: The structured analysis data is parsed to identify the standardized data elements contained therein and the medical semantic relationships between them; The standardized data elements are mapped to corresponding clinical entities and numerical features, and an initial patient clinical profile is constructed based on the medical semantic relationships. Using the clinical entity as the search key, the atrial fibrillation guidelines are traversed to obtain the pathophysiological nodes, treatment intervention nodes, and potential complication nodes directly associated with them, generating a candidate medical concept set. Simultaneously, based on the numerical characteristics and the threshold conditions in the clinical pathway rules, the matching degree of the current patient relative to each standard pathway is calculated, and its main pathway anchor point and deviation status are determined. The candidate medical concept set is fused with the main path anchor point and deviation state to generate an enhanced clinical context vector; Using the decision logic tree defined in the clinical pathway rules, reasoning is performed on the enhanced clinical context vector to disambiguate and filter out the treatment options and risk assessment results that meet the conditions, forming structured decision elements; Based on the evidence-based relationship links in the atrial fibrillation guidelines, the structured decision elements are subjected to interpretive correlation and confidence fusion to generate the decision support information containing recommended solutions, justifications, and warnings.

8. The atrial fibrillation precision treatment auxiliary decision-making system based on a large language model according to claim 1, characterized in that, The structured data is integrated with the decision support information to generate a visualized clinical decision support report, specifically: The recommended solutions, their basis, and warnings in the decision support information are extracted as core decision elements. The structured data is parsed, and the clinical entities and numerical features therein are mapped to original report fragments according to a preset clinical report template; Based on the semantic association rules between the core decision elements and the clinical report template, the core decision elements are dynamically inserted into the original report fragment to generate primary integrated content containing a decision-evidence chain; The initial integrated content is optimized for readability to generate a narrative text summary for clinicians; Extract the temporal physiological parameters from the structured data and the risk assessment results from the decision support information, perform multi-scale feature extraction and periodic pattern analysis on the temporal physiological parameters, and perform probability distribution and sensitivity analysis on the risk assessment results. Encode the analysis results into graphical syntax instructions to generate dynamic trend charts and risk matrix charts. The narrative text summary, the dynamic trend chart, and the risk matrix chart are arranged spatially and visually associated according to the layout rules of the clinical report template to generate a visual report draft. The confidence level and clinical pathway anchor points attached to the decision support information are embedded as metadata, and the visualized clinical decision support report is output.

Images