Online diagnosis and treatment data supervision method and system based on artificial intelligence

Abstract

The invention discloses an online diagnosis and treatment data supervision method and system based on artificial intelligence, which is used for carrying out real-time intelligent supervision on free text diagnosis and treatment records generated by an online medical platform, and specifically comprises the following steps: carrying out semantic coding on text fields such as basic information, chief complaint, present medical history, past history and diagnosis of a patient; executing necessary filling and logic consistency verification based on a predefined multi-level rule base; clustering analysis is carried out on the semantic vectors by adopting a density clustering algorithm and the like, and semantic abnormity or duplicate records are identified; filtering junk texts and test data by using keyword extraction and named entity recognition technologies; and finally, outputting a structured supervision report and adding a quality label to each record. Through the method, various quality and compliance problems in the diagnosis and treatment data can be found and labeled in real time, and powerful support is provided for compliance supervision of an internet medical platform, improvement of service quality and guarantee of data credibility; meanwhile, a rule-driven and AI model-driven'double-track 'framework is adopted in the method, and an active learning mechanism is combined, so that the system can continuously perform self-learning and iterative evolution and continuously adapt to a new data mode and a new supervision requirement.

Description

Technical Field

[0001] This invention relates to the field of online medical data supervision technology, specifically to an artificial intelligence-based method and system for online medical data supervision. Background Technology

[0002] With the rise of internet healthcare and online diagnosis and treatment models, a large amount of medical data is generated and stored through online platforms. This data mainly records patients' personal information, symptoms, chief complaints, medical history, and diagnoses in free text format. During the data collection and entry process, due to the lack of face-to-face review by the system, various quality problems frequently occur, affecting service quality and posing compliance risks. Typical problems include: missing data (such as incomplete medical records due to blank diagnostic fields), duplicate data (redundancy caused by multiple entries of the same consultation content), test / fake data (non-real information such as "Test" or "test" entered for debugging), junk data (containing meaningless character combinations such as random letter strings), and logical errors (such as impossible situations like a male recording "pregnancy" related symptoms). If the above-mentioned poor-quality data is not detected and rectified in a timely manner, it may lead to statistical analysis biases, affecting medical decisions and platform compliance supervision. Therefore, ensuring the accuracy, completeness, and compliance of online diagnosis and treatment data is crucial for the safe operation and regulatory requirements of internet healthcare platforms. Summary of the Invention

[0003] The purpose of this invention is to provide an online medical data supervision method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the first aspect of this application proposes an online medical data monitoring method based on artificial intelligence, comprising the following steps:

[0005] S1, Data Access and Cleaning: Obtain medical records from the online medical platform and preprocess the medical records;

[0006] S2, Compliance rule verification: The preprocessed medical records are verified using a rule base to obtain a set of abnormal data tags;

[0007] S3, Text Semantic Encoding: The compliance records that pass the compliance rule verification are vectorized using a pre-trained language model to obtain semantic feature vectors;

[0008] S4, Anomaly Clustering Detection: Perform unsupervised density clustering analysis on semantic feature vectors to obtain a set of semantic anomaly and duplicate record labels;

[0009] S5, Spam / Test Data Filtering: Perform semantic hierarchical analysis on the text content of each preprocessed medical record to obtain a set of invalid or noisy data labels;

[0010] S6, Regulatory Report Generation and Labeling: Summarize the above-mentioned abnormal data label set, semantic abnormal and duplicate record label set, and invalid or noisy data label set to generate a structured data regulatory report.

[0011] According to one embodiment of this application, the medical record is a field, which includes: visit ID, patient ID, age, gender, chief complaint, present illness history, past medical history, diagnosis, medical orders, and timestamp.

[0012] According to one embodiment of this application, the preprocessing includes cleaning and normalizing the collected text fields.

[0013] According to one embodiment of this application, the pre-trained language model includes: a medical domain pre-trained model.

[0014] According to one embodiment of this application, the unsupervised density clustering method is DBSCAN. DBSCAN forms clusters based on the density distribution of vectors in the feature space and identifies discrete points that do not belong to any core cluster as outliers.

[0015] According to one embodiment of this application, the spam / test data filtering further includes: extracting keywords from each record using the TF-IDF algorithm, identifying strings with abnormal frequency or suspected meaninglessness; if the keywords are mostly composed of random characters, repeated symbols, or words without medical significance, then the record is determined to be spam data; at the same time, a medical domain named entity recognition model is used to extract medical entities from the text; if a record has an abnormally low number of entities and contains words that are obviously unrelated to medicine, then it is determined to be test data.

[0016] According to one embodiment of this application, the regulatory report generation and structured annotation includes: the report listing the identification information and corresponding problem type tags for each abnormal record in the form of a list or table.

[0017] A second aspect of this application proposes an online medical data monitoring system based on artificial intelligence, comprising:

[0018] The data access and cleaning module is used to acquire medical records from the online medical platform and preprocess the medical records.

[0019] The compliance rule verification module uses a rule base to verify the pre-processed medical records and obtain a set of abnormal data tags.

[0020] The text semantic encoding module is used to vectorize compliant medical records that have passed compliance rule verification using a pre-trained language model to obtain semantic feature vectors.

[0021] The anomaly clustering detection module is used to perform unsupervised density clustering analysis on semantic feature vectors to obtain a set of semantic anomaly and duplicate record labels.

[0022] The spam / test data filtering module is used to perform semantic hierarchical analysis on the text content of each preprocessed medical record to obtain a set of invalid or noisy data labels;

[0023] The regulatory report generation and annotation module is used to summarize the above-mentioned abnormal data label set, semantic abnormal and duplicate record label set, and invalid or noisy data label set to generate a structured data regulatory report.

[0024] According to one embodiment of this application, the regulatory report generation and structured annotation includes: the report listing the identification information and corresponding problem type tags for each abnormal record in the form of a list or table.

[0025] A third aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI-based online medical data monitoring method as described in any of the preceding claims.

[0026] The online medical data quality control method and system provided by this invention have the following outstanding advantages and technical effects:

[0027] 1. This invention is specifically designed for free text data on online medical platforms, combining deep semantic understanding with rule verification to construct an integrated intelligent monitoring framework. Unlike traditional quality control methods for hospital EMR, this method considers the characteristics of online medical services (unstructured format, susceptibility to forged / tested records, etc.), and achieves accurate identification and processing of various anomalies such as omissions, redundancy, falsification, spam, and logical contradictions by integrating multimodal analysis techniques.

[0028] 2. This method employs a dual-track architecture of rule-driven and AI model-driven approaches. Expert-defined compliance rules are used for basic screening and issue labeling, while pre-trained language models and unsupervised clustering are used to achieve semantic-level anomaly detection and pattern discovery. The two complement each other: rules provide interpretable compliance justification, while the AI ​​model is adaptive to complex semantic errors and unknown patterns; and through an active learning mechanism, it continuously absorbs expert feedback for iterative upgrades, balancing algorithm transparency and continuous evolution capabilities.

[0029] 3. This system is characterized by real-time, intelligence, and adaptability, enabling dynamic monitoring and quality supervision of online medical data. Through the generated structured regulatory reports and data annotations, it supports compliant platform operation and improves the quality of medical services, significantly enhancing data credibility and regulatory efficiency. Simultaneously, the system reserves scalable compliance regulatory interfaces, supports multi-source data fusion and platform-level audit log functions, and can meet higher-level regulatory needs and traceability requirements. Attached Figure Description

[0030] Figure 1 This is a flowchart of an artificial intelligence-based online medical data supervision method. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Figure 1 This invention provides an online medical data monitoring method based on artificial intelligence, comprising the following steps:

[0033] S1, Data Access and Cleaning: Obtain medical records from the online medical platform and preprocess the medical records;

[0034] S2, Compliance rule verification: The preprocessed medical records are verified using a rule base to obtain a set of abnormal data tags;

[0035] S3, Text Semantic Encoding: The compliance records that pass the compliance rule verification are vectorized using a pre-trained language model to obtain semantic feature vectors;

[0036] S4, Anomaly Clustering Detection: Perform unsupervised density clustering analysis on semantic feature vectors to obtain a set of semantic anomaly and duplicate record labels;

[0037] S5, Spam / Test Data Filtering: Perform semantic hierarchical analysis on the text content of each preprocessed medical record to obtain a set of invalid or noisy data labels;

[0038] S6, Regulatory Report Generation and Labeling: Summarize the above-mentioned abnormal data label set, semantic abnormal and duplicate record label set, and invalid or noisy data label set to generate a structured data regulatory report.

[0039] In the above embodiments, by deeply integrating technologies such as semantic analysis, compliance rules, and anomaly detection through data access and cleaning, compliance rule verification, text semantic encoding, anomaly clustering detection, spam / test data filtering, and regulatory report generation and annotation, intelligent and real-time monitoring of online diagnosis and treatment data is achieved, providing reliable data protection for the compliant operation of medical platforms.

[0040] According to one embodiment of this application, the medical record is a field, which includes: visit ID, patient ID, age, gender, chief complaint, present illness history, past medical history, diagnosis, medical orders, and timestamp.

[0041] Specifically, the system retrieves medical records in XML format from the online medical platform via an HTTP interface. Fields include: visit ID, patient ID, age, sex, chief complaint, present illness, past history, diagnosis, doctor's advice, and timestamp.

[0042] According to one embodiment of this application, the preprocessing includes cleaning and normalizing the collected text fields.

[0043] Specifically, the collected text fields undergo cleaning and standardization processes, including: removing HTML tags, control characters, and emoticons; unifying full-width / half-width characters; and converting between simplified and traditional Chinese characters. A professional medical sentence segmenter is used for sentence segmentation and punctuation correction, and a medical dictionary and Hidden Markov Model are applied for word segmentation and part-of-speech tagging. These processes eliminate input noise and extract semantic units, providing clean and structurally consistent input data for subsequent analysis modules.

[0044] Specifically, the input consists of raw medical records collected from internet healthcare platforms, denoted as a set:

[0045] R = {r1, r2, ..., r} n}

[0046] Where, r i This represents the i-th original medical record, with the following structure:

[0047] r i =(visit_id,chief_complaint,diagnosis,…)

[0048] Among them, visit_id is the unique identifier of the medical record; chief_complaint is the chief complaint information; diagnosis is the diagnosis content; and there are other fields such as medical history, doctor's orders, patient information, etc.

[0049] Processing procedure:

[0050] Standardized medical records after cleaning

[0051] Where, r i Original medical records; This is a structured extraction function that parses XML input and extracts fields. This is a cleaning and standardization function, which includes tag removal, simplified / traditional character conversion, character cleaning, synonym normalization, and terminology standardization.

[0052] Output:

[0053] After structured extraction and cleaning, the standardized medical record set is represented as follows:

[0054]

[0055] in, A standardized collection of medical records; For the i-th medical record after structured and cleaned processing, the corresponding original record is r. i ;

[0056] According to one embodiment of this application, the rule base includes two main categories: integrity rules and logical consistency rules.

[0057] The rule base is a predefined rule base by domain experts.

[0058] According to one embodiment of this application, each standardized medical record is sequentially validated using rule functions from the rule base, resulting in a standardized set of medical records. Based on the verification results, it is divided into: a set of compliance records. Non-compliance record collection For the set of non-compliant records, an abnormal data label set is further generated.

[0059] Specifically, a predefined rule base defined by domain experts is used to perform hard validation on medical records. The rule base includes two main categories: integrity rules and logical consistency rules. For example, integrity rules require key fields such as chief complaint and diagnosis to be non-empty, age to be within the range [0, 120], gender to be within the set {M, F}, and the diagnosis code to be retrieved and approved in the ICD-10 standard code table. Logical consistency rules are used to identify inconsistencies within the records, such as age-disease validation (e.g., age < 12 and abnormal markers for a diagnosis of prostate disease), gender-symptom validation (abnormal markers for males such as "pregnancy" or "cervix"), and time sequence validation (treatment end time must not be earlier than start time). These rules are executed through the Drools engine, generating preliminary problem labels (e.g., "missing fields," "logical conflicts") for each record.

[0060] By rapidly screening standardized medical records through a pre-defined rules engine, the system can efficiently identify and filter out obviously non-compliant or formatted medical data, thereby significantly reducing the burden on subsequent semantic analysis and anomaly detection. Simultaneously, the rules base supports a dynamic update mechanism, continuously iterating and optimizing based on the latest regulatory policy changes and feedback from clinical experts to adapt to evolving compliance requirements and medical practice standards, ensuring the system possesses excellent adaptability and scalability.

[0061] Input: A collection of standardized medical records after cleaning. Rule base F = {f1, f2, ..., f k}, where each rule function f i A judgment logic corresponding to a violation or anomaly type is used to perform compliance screening on input data.

[0062] Processing procedure:

[0063] The system processes each standardized medical record The rule functions from the rule base are applied sequentially for validation. For any rule function f j ∈F, its effect is defined as:

[0064]

[0065] In this system, 1 indicates compliance with the rules, and 0 indicates violation or abnormality.

[0066] Each record is validated sequentially:

[0067] like Non-compliant

[0068] Output:

[0069] The rule validation module will validate the standardized medical record set. Based on the verification results, it is divided into:

[0070] Compliance Record Collection

[0071] Non-compliance record collection

[0072] For the set of non-compliant records, the system further generates a set of abnormal samples and their corresponding tags:

[0073] Non-compliance records and tag collection in, This indicates anomaly samples identified through rule-based validation; i These are exception labels, such as "Missing field: diagnosis" or "Missing field: chief complaint".

[0074] According to one embodiment of this application, the pre-trained language model includes: a medical domain pre-trained model.

[0075] According to one embodiment of this application, the step of vectorizing the compliant medical records that have passed the compliance rule verification using a pre-trained language model to obtain semantic feature vectors includes:

[0076] Unstructured text fields are encoded using a medical pre-trained language model, and their global semantic hierarchy is extracted. i =MedBERT [CLS] (text i ); the semantic vector e of the diagnosis and treatment text i Concatenating the structured field with its one-hot encoded data yields a composite feature vector x used for rule-based decision-making and model training. i : Output a set of semantic feature vectors.

[0077] Specifically, the pre-processed diagnostic and treatment text fields are vectorized using a pre-trained language model. A medical-domain pre-trained model (such as Chinese-MedBERT) can be used to learn representations of patient complaints, present medical history, past medical history, and diagnoses. The fields are concatenated and a special delimiter is added before being input into the model. A fixed-dimensional high-dimensional sentence vector representation is generated by pooling the output vector (e.g., extracting the [CLS] vector). This representation integrates clinical terminology and contextual information, enabling the measurement of semantic similarity between different diagnostic and treatment records. To accommodate structured information, fields such as age and gender are one-hot encoded and concatenated with the text vector to form a unified feature vector. After processing by the text semantic encoding module, each record receives a semantic feature vector, which serves as the input basis for subsequent clustering and rule matching.

[0078] Specifically,

[0079] enter:

[0080] Standardized text The unstructured text fields such as chief complaint, present illness, past medical history, and diagnosis are represented as text. i .

[0081] Processing procedure:

[0082] Use a medical pre-trained language model (such as Chinese-MedBERT) to process the i-th diagnosis text. i Encode and extract its global semantic vector representation e i :

[0083] e i =MedBERT [CLS] (text i )

[0084] Among them, text i The text represents the i-th medical record, including chief complaint, medical history, diagnosis, etc.; [CLS] is a special label for the BERT model, and its corresponding output vector is widely used in sentence-level tasks (such as classification and similarity calculation). In this method, it is used to represent the overall semantic information of the medical record.

[0085] To achieve the fusion of structured and unstructured semantic information, this method integrates the semantic vector e of the diagnostic text. i The feature vector x is concatenated with the one-hot encoded structured fields (such as age and gender) to obtain the final composite feature vector x used for rule judgment and model training. i :

[0086]

[0087] Among them, OneHot(age i ) is the One-hot encoding of the age field, OneHot(sex) i ) represents the One-hot encoding of the gender field. This refers to vector concatenation.

[0088] Output:

[0089] The set of semantic feature vectors of medical records: Where x i Let be the semantic feature vector representation of the i-th medical record.

[0090] According to one embodiment of this application, the unsupervised density clustering method is DBSCAN. DBSCAN forms clusters based on the density distribution of semantic feature vectors in the feature space and identifies discrete points that do not belong to any core cluster as outliers.

[0091] According to one embodiment of this application, performing unsupervised density clustering analysis on the semantic feature vectors includes: processing the set X of structured-unstructured fused semantic feature vectors. valid Perform density clustering to generate a set of category labels: Perform a repeatability check on the clustering results. If a certain cluster C j The number of samples included exceeds a preset threshold θ, i.e., |C j If |>θ, then the cluster is marked as a duplicate cluster; thus, the abnormal data set is obtained. This results in a set of semantic anomalies and duplicate record labels.

[0092] The system employs the DBSCAN clustering algorithm, performing cluster analysis based on the density distribution of semantic feature vectors in the feature space of medical records, thereby achieving unsupervised anomaly detection and duplicate data identification. Semantically similar normal records are clustered together, while semantically isolated samples are identified as "semantic anomalies." If the number of highly similar records in a small cluster exceeds a threshold, it is determined to be "duplicate data." The anomaly clustering detection module does not rely on manual annotation and can effectively discover unknown anomaly types. To ensure adaptability, core clustering parameters (such as neighborhood radius ε and minimum sample size minPts) can be dynamically adjusted in conjunction with real-time statistics to enhance robustness to the diversity and anomaly distribution of medical data.

[0093] For example, input: compliance record R valid semantic feature vector set

[0094] Processing procedure:

[0095] This method employs the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm to process the semantic vector set X after the fusion of structured and unstructured data. valid Perform density clustering to generate a set of category labels:

[0096]

[0097] Among them, X valid ={x1,x2,…,x n} represents the set of effective diagnostic semantic feature vectors after filtering by pre-defined rules; L cls (X valid ) represents the set of category labels after clustering the set of valid semantic feature vectors; c i For sample x i Clustering labels; The integer value represents the cluster number, and -1 indicates that the clustering algorithm failed to cluster sample x. i If a point is categorized into any density cluster, it is generally considered a noise point or a potentially abnormal medical record.

[0098] To identify potential duplicate medical records, the system performs a duplicate detection on the clustering results. If a certain cluster C... j The number of samples included exceeds a preset threshold θ, i.e., |C j If |>θ, then the cluster is marked as a duplicate cluster, and the records in it may have problems such as batch generation, template filling, or duplicate submission.

[0099] If |C j If |>θ, then the cluster is marked as a repeating cluster.

[0100] in, This represents a cluster obtained after DBSCAN clustering; |C j | indicates the number of samples contained in the cluster; θ is the threshold parameter for duplicate judgment (manually set, representing the upper limit of the "acceptable frequency" of a certain type of semantically similar records).

[0101] Output:

[0102] Abnormal data collection

[0103] in, This indicates outlier samples identified after cluster analysis; i For exception labels, such as "semantic exception" or "duplicate data".

[0104] According to one embodiment of this application, the method for filtering junk data and test data includes: extracting keywords, recognizing named entities, performing semantic pattern analysis and calculating statistical indicators for each medical record, and jointly identifying low-quality or non-authentic records from multiple dimensions.

[0105] According to one embodiment of this application, the filtering of junk data and test data further includes: extracting keywords from each record using the TF-IDF algorithm and detecting whether there are strings with abnormal frequency or suspected meaninglessness; if the extracted keywords are mainly composed of random characters, repeated symbols, or words without medical meaning, the record is initially judged to be junk data; applying a named entity recognition model in the medical field to extract professional medical entities from the record; if the number of entities extracted from a record is abnormally small and there are obvious non-medical terms, it is determined to be test data.

[0106] Specifically, keyword extraction and entity recognition technologies are applied to perform semantic hierarchical analysis on the text content of each record in order to identify spam and test data.

[0107] Specifically, the TF-IDF algorithm is first used to extract keywords from each record, and then the records are checked for abnormal frequency or potentially meaningless strings. If the extracted keywords are mainly composed of random characters, repeated symbols (such as "aaa" or "111"), or words without medical significance, the record is preliminarily identified as spam.

[0108] Secondly, a Named Entity Recognition (NER) model from the medical field is applied to extract professional medical entities (such as symptoms, diseases, and drug names) from the records. Normal online consultation records typically contain multiple clear and reasonable medical entities. If a record contains an unusually small number of extracted entities and contains clearly non-medical terms such as "test," "test," or "123," it is further determined to be test data.

[0109] To enhance robustness, this method also incorporates the following auxiliary indicators:

[0110] Repeated character ratio: Used to identify structured garbage input, such as continuous repetition of the same character or word;

[0111] Spam dictionary matching: A pre-built medical-irrelevant vocabulary (such as "test data", "sample text", etc.) can quickly determine typical test fill content.

[0112] Ultimately, all records identified as spam or test data are marked as anomalous samples by the system and assigned corresponding labels (such as "spam data: duplicate characters" or "test data: no entity") to support subsequent report generation and manual review.

[0113] For example, input: a standardized set of records Medical Entity Dictionary med Garbage dictionary D noise

[0114] Processing procedure:

[0115] To further support manual verification and rule updates for abnormal records, the system employs the TF-IDF method to extract keywords from standardized medical records. Specifically, for each record... Calculate the TF-IDF value of its terms and select the top k high-weight terms as its keyword set:

[0116]

[0117] Wherein, TFIDF stands for Term Frequency-Inverse Document Frequency algorithm; For the i-th medical record; keywords i={w1,w2,…,w k} indicates from The set of k extracted keywords is sorted by weight.

[0118] Extracting entity information from medical records:

[0119]

[0120] Among them, E i It is the set of entities in the i-th record; NER is the named entity recognition function, used to identify medical entities such as diseases, symptoms, drugs, etc.

[0121] like If no medical-related entities are identified in the record, it indicates that the record may have no actual medical meaning and is therefore judged as spam.

[0122] Spam detection function:

[0123] Among them, a high proportion of repeated characters is a method for detecting structured spam data, used to determine whether the input is random input, data manipulation, or template data, such as spam data like "ah ah ah ah ah", "aaaaa", "11111", etc.; if This indicates that the record contains spam, such as "test" or "asdf".

[0124] Output:

[0125] in, This indicates abnormal samples identified by the garbage data detection module; i These are abnormal tags, such as "garbage data" or "test data".

[0126] According to one embodiment of this application, the regulatory report generation and structured annotation includes: the report listing the identification information and corresponding problem type tags for each abnormal record in the form of a list or table.

[0127] According to one embodiment of this application, the regulatory report generation and structured annotation further includes: unifying the aforementioned non-compliant medical records, abnormal medical records, spam information, and test data into a final anomaly annotation set L = L. rules ∪L cluster ∪L garbage The abnormal medical record set L obtained in the previous steps is combined with multi-source labels, further mapped to a structured report format, and the final report is generated: T = {(visit_id)} i Problem type i ,describe i )}.

[0128] Specifically, after completing the above analysis, the system summarizes all detected problem information and generates a structured data monitoring report. The report lists or tables the identification information and corresponding problem type tags for each abnormal record. For example, for each problematic record, the report indicates the record's ID and a problem description (e.g., "Missing field: Diagnosis", "Duplicate data: Duplicate with ID=V123", "Junk data: Contains invalid characters", "Logical anomaly: Gender contradicts medical record", etc.). The report can also statistically analyze the frequency and proportion of various problems, helping operators assess the overall data quality. In addition to reports, the system adds a quality tag field to each medical record in the data storage, recording its pass / fail status and specific problems detected. Upper-level systems (such as electronic medical record systems and decision support systems) can automatically filter or prompt problematic data based on these tags when accessing data, achieving dynamic control of data quality. The above reports and annotations provide the foundation for the traceable management of online medical data. When the rule base or model is updated, this process can be run again to update the annotations and reports, enabling cyclical optimization of data quality control.

[0129] For example, input: The system integrates the label information from the three anomaly detection outputs into a final anomaly label set L =

[0130] L rules ∪L cluster ∪L garbage .

[0131] Among them, L rules This is a set of abnormal data tags obtained based on the compliance rule verification module; L cluster L is a set of semantic anomalies and duplicate record labels detected based on clustering algorithms. garbage This is the set of invalid or noisy data labels identified by the garbage data detection module.

[0132] Processing procedure:

[0133] The abnormal medical record set L obtained in the previous steps is combined with multi-source labels, further mapped to a structured report format, and the final report is generated: T = {(visit_id)} i Problem type i ,describe i )}.

[0134] Among them, visit_id i It is the unique identifier for this medical record; Problem type i The description is for terms such as "semantic anomaly", "duplicate data", and "junk data"; the description i is a text description generated based on keyword extraction and diagnosis content, which is convenient for manual verification.

[0135] After completing the quality control of medical records, an anomaly label is assigned to each abnormal record. i Write to its database field quality tag This is used for subsequent queries, quality statistics, and model feedback training.

[0136] Database update operations:

[0137] in, Medical records marked as abnormal; i For the corresponding exception labels, such as "semantic anomaly", "duplicate data", "junk data", etc.; quality tag This is a field in the database used to record the quality control results of this medical record.

[0138] According to one embodiment of this application, it also includes, S7, continuous model iteration and active learning.

[0139] Furthermore, after performing rule validation, cluster detection, and spam filtering, some records could not be clearly categorized. These medical records, whose clustering label was "uncertain" and which were not judged as spam data, were collected as an unknown label sample set. The records are then submitted to human experts for review. After review, the experts assign a clear anomaly type label to each record, forming a manually labeled set. For new sample sets generated by expert annotation By incrementally updating the model parameters θ by minimizing the supervised learning loss function, new rules f can be summarized from the expert annotation set when some outlier samples exhibit clear and inductively characteristic patterns. new Update the rule base.

[0140] Specifically, for novel anomalies appearing in cluster analysis or keyword recognition, the system initiates an active learning loop: such samples are pushed to domain experts for review, and after obtaining accurate annotations, they are incorporated into the training set for model retraining. Simultaneously, experts can summarize and add new rules based on newly discovered issues. For example, corresponding validation rules can be added to the rule base for new terminology or formatting errors. Through expert rule maintenance and iterative development with the AI ​​model, this invention achieves continuous self-learning and evolution of the system. Combining manual review with machine learning maintains the interpretability of the rules and continuously improves detection capabilities with the accumulation of new data, better meeting the dual requirements of regulatory authorities for algorithm transparency and continuous optimization.

[0141] enter:

[0142] After the system performs rule validation, cluster detection, and spam filtering, some records cannot be clearly categorized. These medical records, whose cluster label is "uncertain" and which were not identified as spam, are collected as an unlabeled sample set. The records are then submitted to human experts for review. After review, the experts assign a clear anomaly type label to each record, forming a manually labeled set.

[0143] Where L cls (x i ) = -1 indicates that the clustering result is "uncertain". The result indicates that the garbage is detected as "no".

[0144] in l i The label is in the abnormal record set L.

[0145] Processing procedure:

[0146] By leveraging expert annotation feedback on unlabeled samples, two updates are performed: first, the model parameters are incrementally trained to improve the ability to identify complex or boundary samples; second, new rules are extracted from the labeled data to expand the existing rule base. This "model-rule" dual-track feedback mechanism enhances the system's continuous adaptability and interpretability.

[0147] For new sample sets generated by expert annotation Incremental updates of model parameters θ are performed by minimizing the supervised learning loss function:

[0148]

[0149] Where θ represents the current model parameters; η represents the learning rate; For annotation set Constructed cross-entropy loss function; This is the gradient with respect to the parameters.

[0150] When some outlier samples exhibit clear and inductively applicable patterns, new rules f can be summarized from expert annotation sets. new Update the rule base:

[0151] F←F∪{f new}

[0152] Where F is the existing set of rule functions; f new A rule function that can be expressed programmatically, abstracted by experts from labeled samples. After the update, f new It can directly participate in subsequent rule-based verification.

[0153] Output: The updated model and rules, used for the next round of detection.

[0154] According to one embodiment of this application, it also includes functional extensions.

[0155] Specifically, in addition to the main processes mentioned above, this system can be expanded with the following functions as needed:

[0156] Compliance and Regulatory Interface: Provides an external API to output data quality analysis reports and anomaly alerts in real time to health regulatory agencies or the platform's internal compliance module, enabling connection and information sharing with external regulatory systems.

[0157] Multi-dimensional data access: In addition to medical records, it can access various data sources such as structured medical records, medical images, and wearable device monitoring data, enabling unified supervision and analysis of information across data sources and enhancing the comprehensiveness of data monitoring.

[0158] Platform-level traceability logs: Audit logs that record all data processing flows and results at the system level, including rule verification logs, model inference records, and manual feedback records, to ensure the auditability and traceability of the entire regulatory process.

[0159] Through the above implementation methods, the present invention forms a data supervision system for online diagnosis and treatment scenarios, which deeply integrates technologies such as semantic analysis, compliance rules and anomaly detection, realizes intelligent and real-time monitoring of online diagnosis and treatment data, and provides reliable data protection for the compliant operation of medical platforms.

[0160] The results of specific implementation examples are as follows:

[0161] Table 1. Statistics and Examples of Quality Control Problem Types

[0162]

[0163] illustrate:

[0164] • Number of detected records: The number of records marked as a specified problem by the algorithm and rule base.

[0165] • Percentage (%): Number of corresponding questions ÷ Total data volume × 100%.

[0166] • Example record ID: Randomly selected from the hit set for manual review and model calibration.

[0167] Table 2. Trends in online medical service data quality issues over the past four months (unit: %)

[0168]

[0169] *Total percentage of quality issues = (Number of all marked records in the current month ÷ Total number of records in the current month) × 100%.

[0170] This table can be directly incorporated into regulatory quarterly and annual reports and compared with platform product / operation improvement measures to assess the effectiveness of quality control strategies and the extent of data quality improvement.

[0171] According to a second aspect of the present invention, an artificial intelligence-based online medical data monitoring system is provided, which corresponds to an artificial intelligence-based online medical data monitoring method in the above embodiments.

[0172] According to one embodiment of this application, an online medical data monitoring system based on artificial intelligence includes:

[0173] The data access and cleaning module is used to acquire medical records from the online medical platform and preprocess the medical records.

[0174] The compliance rule verification module is used to perform hard verification on the pre-processed medical records using the rule base and filter out non-compliant medical records.

[0175] The text semantic encoding module is used to vectorize compliant medical records that have passed compliance rule verification using a pre-trained language model to obtain semantic feature vectors.

[0176] The anomaly clustering detection module is used to perform unsupervised density clustering analysis on semantic feature vectors to obtain anomaly points, thereby obtaining abnormal medical records;

[0177] The spam / test data filtering module is used to apply keyword extraction and entity recognition technologies to perform semantic hierarchical analysis on the text content of each preprocessed medical record in order to identify spam information and test data.

[0178] The regulatory report generation and annotation module is used to summarize the aforementioned non-compliant medical records, abnormal medical records, spam information, and test data to generate structured data regulatory reports.

[0179] According to one embodiment of this application, the medical record is a field, which includes: visit ID, patient ID, age, gender, chief complaint, present illness history, past medical history, diagnosis, medical orders, and timestamp.

[0180] According to one embodiment of this application, the preprocessing includes cleaning and normalizing the collected text fields.

[0181] According to one embodiment of this application, the rule base includes two main categories: integrity rules and logical consistency rules.

[0182] According to one embodiment of this application, the pre-trained language model includes: a medical domain pre-trained model.

[0183] According to one embodiment of this application, the unsupervised density clustering method is DBSCAN. DBSCAN forms clusters based on the density distribution of vectors in the feature space and identifies discrete points that do not belong to any core cluster as outliers.

[0184] According to one embodiment of this application, the spam / test data filtering further includes: extracting keywords from each record using the TF-IDF algorithm, identifying strings with abnormal frequency or suspected meaninglessness; if the keywords are mostly composed of random characters, repeated symbols, or words without medical significance, then the record is determined to be spam data; at the same time, a medical domain named entity recognition model is used to extract medical entities from the text; if a record has an abnormally low number of entities and contains words that are obviously unrelated to medicine, then it is determined to be test data.

[0185] According to one embodiment of this application, the regulatory report generation and structured annotation includes: the report listing the identification information and corresponding problem type tags for each abnormal record in the form of a list or table.

[0186] According to one embodiment of this application, it also includes continuous model iteration and active learning.

[0187] According to a third aspect of the present invention, an electronic device is provided, which may include: a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other through the communication bus. The processor may invoke logical instructions in the memory to execute an online medical data monitoring method and system based on artificial intelligence according to any of the above embodiments.

[0188] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0189] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute an online diagnosis and treatment data monitoring method and system based on artificial intelligence provided by the above methods.

[0190] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform an artificial intelligence-based online diagnosis and treatment data monitoring method and system provided by the above methods.

[0191] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0192] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for supervising online medical data based on artificial intelligence, comprising the following steps: S1, Data Access and Cleaning: Obtain medical records from the online medical platform and preprocess the medical records; S2, Compliance rule verification: The preprocessed medical records are verified using a rule base to obtain a set of abnormal data tags; S3, Text Semantic Encoding: The compliance records that pass the compliance rule verification are vectorized using a pre-trained language model to obtain semantic feature vectors; S4, Anomaly Clustering Detection: Perform unsupervised density clustering analysis on semantic feature vectors to obtain a set of semantic anomaly and duplicate record labels; S5, Spam / Test Data Filtering: Perform semantic hierarchical analysis on the text content of each preprocessed medical record to obtain a set of invalid or noisy data labels; S6, Regulatory Report Generation and Labeling: Summarize the above-mentioned abnormal data label set, semantic abnormal and duplicate record label set, and invalid or noisy data label set to generate a structured data regulatory report.

2. The method for monitoring online medical data based on artificial intelligence as described in claim 1, characterized in that: The medical record consists of fields, including: visit ID, patient ID, age, gender, chief complaint, present illness, past medical history, diagnosis, medical orders, and timestamp.

3. The method for monitoring online medical data based on artificial intelligence as described in claim 2, characterized in that: The preprocessing includes cleaning and normalizing the collected text fields.

4. The method for monitoring online medical data based on artificial intelligence as described in claim 1, characterized in that: The pre-trained language models include: medical domain pre-trained models.

5. The method for monitoring online medical data based on artificial intelligence as described in claim 1, characterized in that: The unsupervised density clustering method is DBSCAN. DBSCAN forms clusters based on the density distribution of vectors in the feature space and identifies discrete points that do not belong to any core cluster as outliers.

6. The method for monitoring online medical data based on artificial intelligence as described in claim 1, characterized in that: The filtering of spam / test data includes: extracting keywords, identifying named entities, performing semantic pattern analysis and calculating statistical indicators for each medical record, and jointly identifying low-quality or non-authentic records from multiple dimensions.

7. The method for monitoring online medical data based on artificial intelligence as described in claim 1, characterized in that: The regulatory report generation and structured annotation include: the report lists or tables the identification information and corresponding problem type tags for each abnormal record.

8. An online medical data monitoring system based on artificial intelligence, comprising: The data access and cleaning module is used to acquire medical records from the online medical platform and preprocess the medical records. The compliance rule verification module uses a rule base to verify the pre-processed medical records and obtain a set of abnormal data tags. The text semantic encoding module is used to vectorize compliant medical records that have passed compliance rule verification using a pre-trained language model to obtain semantic feature vectors. The anomaly clustering detection module is used to perform unsupervised density clustering analysis on semantic feature vectors to obtain a set of semantic anomaly and duplicate record labels. The spam / test data filtering module is used to perform semantic hierarchical analysis on the text content of each preprocessed medical record to obtain a set of invalid or noisy data labels; The regulatory report generation and annotation module is used to summarize the above-mentioned abnormal data label set, semantic abnormal and duplicate record label set, and invalid or noisy data label set to generate a structured data regulatory report.

9. The online medical data monitoring system based on artificial intelligence as described in claim 8, characterized in that: The regulatory report generation and structured annotation include: the report lists or tables the identification information and corresponding problem type tags for each abnormal record.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the artificial intelligence-based online medical data monitoring method as described in any one of claims 1 to 7.

Images