An ai-assisted decision platform for personalized treatment of dry eye disease

Abstract

The application relates to the technical field of ophthalmic treatment, in particular to an AI-assisted decision platform for personalized treatment of dry eye disease. The platform comprises a multi-source data integration unit, an AI diagnosis and treatment analysis unit, a personalized scheme generation unit and an efficacy feedback unit based on the physiological cycle law of dry eye disease condition monitoring and an improved statistical analysis logic. The application calculates the variance inflation factor to remove high collinearity features and selects the top 30% core features in importance through a feature screening module, adopts multi-factor regression analysis to establish a core feature quantitative correlation model, calculates the matching degree score of each subtype according to the clinically recognized dry eye disease classification standard to determine the subtype of the patient, and constructs a disease progression risk function by taking the corneal staining score and the disease duration as the independent variables through Kaplan-Meier analysis, calculates the progression probability at different time nodes and divides the risk level, and realizes effective evaluation of the precise subtype classification and disease progression risk of the dry eye disease.

Description

Technical Field

[0001] This invention relates to the field of ophthalmic treatment technology, and more specifically, to an AI-assisted decision-making platform for personalized treatment of dry eye disease. Background Technology

[0002] Dry eye disease is a common ocular surface disease in clinical practice. Its pathogenesis is related to multiple factors, including ocular surface function, overall health, environmental exposure, and lifestyle behaviors. Clinical diagnosis and treatment rely on multi-source data, including clinical history, objective ocular examination, and lifestyle behaviors. However, in the traditional model, multi-source data is scattered across different media such as electronic medical records, examination equipment, and patient self-records, lacking a standardized integration mechanism. Furthermore, manual analysis is inefficient and lacks sufficient classification accuracy. It also lacks the ability to dynamically evaluate treatment efficacy and iterate treatment plans, making it difficult to meet the needs of personalized treatment.

[0003] In the existing technology, relevant patents have explored the system construction and AI-assisted diagnosis of dry eye. For example, Chinese patent CN202011315653.0 discloses a three-terminal dry eye diagnosis system platform based on intelligent AI, including a patient terminal (containing an AI diagnosis unit, with sub-units such as consultation, dynamic vision diagnosis, and a diagnosis report unit), a doctor terminal (viewing diagnosis data and guiding management), an administrator terminal (managing data and summarizing and exporting), and a backend (containing data sending, AI analysis, and AI diagnosis intelligent evolution units, receiving requests from each terminal and generating diagnosis reports). This aims to solve the inconvenience of dry eye treatment and achieve treatment without needing to go to the hospital. Another Chinese patent CN202411766363.6 discloses a dry eye diagnosis and treatment system based on deep neural networks, including input, diagnosis, data matching, and output unit modules. The input unit inputs the patient's dry eye symptoms and OSDI scale results; the diagnosis unit obtains the correspondence between symptoms, OSDI scale, lower tear river height, and mild, moderate, and severe symptoms; the data matching module matches features to generate a graded diagnosis report; and the output unit outputs the graded results to improve the efficiency of graded diagnosis and reduce missed diagnoses and misdiagnoses.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: First, insufficient multi-source data integration and security management capabilities: While Chinese patent CN202011315653.0 constructs a three-terminal architecture encompassing the patient, doctor, and administrator ends, it fails to mention connecting the hospital's electronic medical record system and ophthalmology-specific examination equipment through a unified medical data standardization protocol, nor does it perform structured processing on the collected data, and it lacks a secure storage guarantee mechanism for medical data; Chinese patent CN202411766363.6 relies solely on patients' dry eye symptoms and OSDI scale results as data input, resulting in a single data dimension that does not cover patients' clinical history and lifestyle data, and it does not consider the security protection requirements during data storage; Second, insufficient accuracy in disease classification and disease risk assessment: Chinese patent CN202011... The backend of patent 315653.0 can only generate diagnostic reports, failing to achieve subtyping based on clinically recognized dry eye disease classification standards, and also lacking a disease progression risk assessment model. Chinese patent CN202411766363.6 focuses only on the graded diagnosis of mild, moderate, and severe dry eye, without in-depth analysis of disease subtype differences, and lacks calculations of the probability of disease progression at different time points, thus failing to provide risk warning support for clinical practice. Thirdly, personalized treatment plan generation and dynamic feedback mechanisms for efficacy are lacking: Chinese patent CN202011315653.0 does not generate targeted treatment plans based on individual patient conditions, nor does it establish a closed loop for efficacy evaluation and plan adjustment; Chinese patent CN202411766363.6 only outputs graded diagnostic results, lacks treatment plan generation functionality, and further lacks quantitative efficacy evaluation based on follow-up data and an iterative feedback mechanism for diagnostic models and treatment plans. Therefore, we propose an AI-assisted decision-making platform for personalized treatment of dry eye disease. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-assisted decision-making platform for personalized treatment of dry eye disease, in order to solve the problems mentioned in the background art, such as insufficient multi-source data integration and security management capabilities, insufficient accuracy of disease classification and disease risk assessment, and lack of personalized treatment plan generation and dynamic feedback mechanism for efficacy.

[0006] To address the aforementioned technical problems, the present invention aims to provide an AI-assisted decision-making platform for personalized treatment of dry eye disease, comprising:

[0007] The multi-source data integration unit is based on a medical data standardization protocol. It collects and integrates patients' clinical medical history data, objective eye examination data, and lifestyle data by connecting to the hospital's electronic medical record system, ophthalmology-specific examination equipment, and patient-end data acquisition interface. The data is stored using medical data encryption standards to construct a structured, individualized dry eye dataset.

[0008] The AI-based diagnostic and analysis unit, based on ophthalmic clinical feature engineering and statistical learning algorithms, processes the dataset output by the multi-source data integration unit. It uses variance inflation factor to screen for core features with the highest correlation to dry eye disease and employs multivariate regression analysis to establish a quantitative correlation model between features. Subtype classification is achieved based on clinically recognized classification standards and the feature weight values ​​output by the correlation model. A disease progression risk function is constructed using Kaplan-Meier analysis, with corneal staining score and disease duration as independent variables to calculate the probability of progression at different time points, thus forming a risk assessment result.

[0009] The personalized treatment plan generation unit is based on a rule base constructed from the clinical diagnosis and treatment guidelines for dry eye disease. It combines the subtyping results and risk assessments output by the AI ​​diagnosis and treatment analysis unit, screens patients' allergy history and contraindications to exclude unsuitable treatment methods, optimizes the parameters of applicable treatment methods, and generates personalized plans that include specific treatment measures, implementation guidelines, and expected effect assessments.

[0010] The efficacy feedback unit, based on the physiological cycle pattern of dry eye disease monitoring and improved statistical analysis logic, achieves quantitative evaluation and dynamic adjustment of dry eye disease treatment efficacy through multi-node objective physiological indicator collection, adaptive data distribution baseline difference algorithm, indicator-efficacy correlation driven dynamic weight algorithm, and cross-unit parameter iterative feedback. It is used to correct parameters for the correlation model and risk function of the AI ​​diagnosis and treatment analysis unit, and to provide objective data support and algorithm output basis for the optimization of treatment plan in the personalized plan generation unit.

[0011] The doctor-patient interaction unit provides an interactive interface that conforms to medical information display standards. It provides medical staff with access points for data query, treatment plan editing, and efficacy analysis, and supports manual review and adjustment of AI analysis results. It also provides patients with functions such as visual display of treatment plans, medication reminders, follow-up appointments, and symptom recording, enabling two-way transmission of medical information.

[0012] As a further improvement to this technical solution, the multi-source data integration unit includes a data acquisition module, a data integration module, and an encrypted storage module connected in sequence, wherein:

[0013] The data acquisition module is based on a medical data standardization protocol. It connects to the hospital's electronic medical record system via an HL7 FHIR interface to collect clinical medical history data (including history of systemic diseases, history of ocular surface surgery, and long-term medication records). It connects to ophthalmic examination equipment (tear secretion meter, meibomian gland imaging meter, corneal topography meter) via a DICOM standard interface to collect objective ocular examination data. It also collects lifestyle data (including average daily screen time, environmental humidity exposure records, and contact lens wearing cycle) through patient-side H5 page forms.

[0014] The data integration module is used to perform structured processing on the multi-source data output by the data acquisition module. Specifically, it includes: extracting key information from unstructured electronic medical record text using natural language processing technology; converting DICOM image metadata output by the examination device into numerical indicators; and completing the field association and format unification of multi-source data through field definition specifications based on dry eye clinical diagnosis and treatment standards (covering four core fields: patient basic information, clinical characteristics, examination indicators, and behavioral characteristics) to form a structured dataset containing a unique patient identifier.

[0015] The encrypted storage module uses the AES-256 encryption algorithm to encrypt the structured dataset at the field level. The encrypted data is stored in an encrypted database with access log auditing function according to a three-level index rule of "patient ID-data type-collection time". It also configures role-based access control permissions, which comply with medical data security standards.

[0016] As a further improvement to this technical solution, the AI ​​diagnosis and analysis unit includes a feature selection module, which is used to select core features from the structured dataset output by the multi-source data integration unit; the core feature selection by the feature selection module includes the following steps:

[0017] S210.1 Calculate the variance inflation factor (VIF) of each feature in the dataset and set a VIF threshold. ;

[0018] S210.2, Compare the VIF of each feature with Excluding VIF High collinearity;

[0019] S210.3. Based on the importance weighting table of clinical characteristics of dry eye, select those with VIF not exceeding [a certain value]. Furthermore, the top 30% of the most important features form the core feature set. .

[0020] As a further improvement to this technical solution, the AI ​​diagnosis and analysis unit also includes an association model construction module, which is used to construct a core feature set based on the feature selection module output by the feature selection module. A quantitative correlation model is established using multifactor regression analysis; the correlation model construction module establishes the quantitative correlation model through the following steps:

[0021] S220.1, From the core feature set Continuous features were extracted and used as independent variables. ;

[0022] S220.2, Quantification of dry eye symptoms severity Using the least squares method to solve for the parameters of the regression model, we obtain the constant term. and the weight values ​​of each feature ;

[0023] S220.3 Output the weight values ​​of each feature. , The larger the absolute value, the greater the influence of the corresponding feature on the dry eye condition.

[0024] As a further improvement to this technical solution, the AI ​​diagnosis and analysis unit also includes a risk analysis module, which is used to classify dry eye disease subtypes and assess the risk of disease progression. The risk analysis module performs subtype classification and risk assessment by including the following steps:

[0025] S230.1. Based on clinically recognized classification criteria for dry eye disease, the classification criteria for dry eye disease include aqueous hypoplasia. Excessive evaporation type Hybrid Combined with the feature weight values ​​output by the association model construction module Calculate the matching score for each subtype. ,in Feature weight values absolute value and characteristics and fractals Related indicator variables The sum of the products of (1 when associated, 0 otherwise);

[0026] S230.2 Compare the matching scores of each subtype. The highest-scoring subtype is determined to be the patient's dry eye disease subtype;

[0027] S230.3 Constructing a disease progression risk function using Kaplan-Meier analysis. Corneal staining score Duration of illness As the independent variable, according to a preset time interval calculate Probability of corneal epithelial damage progression at any given time ;

[0028] S230.4, according to The values ​​are used to classify risk levels, forming a system that includes risk level and corresponding time points. The risk assessment results.

[0029] As a further improvement to this technical solution, the personalized solution generation unit includes a treatment guideline structuring module. This module is used to convert clinical treatment guidelines for dry eye disease into computable structured data, specifically including:

[0030] Natural language processing technology was used to segment the guideline text, identify entities, and extract relationships to extract the correspondence between treatment methods and classifications, as well as the correlation threshold between treatment intensity and risk level.

[0031] The extracted relationships are converted into a triplet data structure and stored in the medical knowledge graph. The subject is the subtype or risk level, the object is the treatment method or parameter, and the relationship is a logical association of "adaptation" or "correspondence".

[0032] By comparing the structured data with more than 5 years of clinical diagnosis and treatment data, the matching degree between the structured data and actual diagnosis and treatment behavior is verified. When the matching degree reaches 90% or above, the calibration is completed, and a guideline knowledge base that can be directly called is formed.

[0033] As a further improvement to this technical solution, the personalized solution generation unit further includes a dynamic solution generation module. This dynamic solution generation module is used to generate personalized solutions by combining multi-source data, specifically including:

[0034] Contraindication screening: By comparing allergy history records in the multi-source data integration unit with treatment components or operational characteristics in the guideline knowledge base through a field matching algorithm, contraindication options with a matching degree of 80% or higher are automatically removed;

[0035] Parameter optimization: Based on the classification results output by the AI ​​diagnosis and analysis unit, the baseline parameters of the corresponding classification in the guide knowledge base are called, and a parameter adjustment model is constructed by combining the patient's average daily eye use time and environmental humidity data.

[0036] Solution output: Integrate and optimize treatment measures, implementation guidelines, and expected effect evaluation indicators to generate a structured solution document with a timestamp, and link it to the guideline knowledge base reference nodes and parameter adjustment basis.

[0037] As a further improvement to this technical solution, the efficacy feedback unit includes a follow-up data acquisition module and an adaptive baseline difference analysis module, wherein:

[0038] The follow-up data acquisition module is used to collect tear secretion data by connecting to specialized ophthalmic examination equipment at preset time points after treatment initiation (which correspond to the physiological cycle of dry eye disease changes). Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four categories of objective physiological indicators; each collected objective physiological indicator is associated with a unique patient ID and a collection timestamp. Data is transmitted to an encrypted database via the standard interface of the multi-source data integration unit. Storage, compatible with device interfaces and data storage specifications of multi-source data integration units;

[0039] The adaptive baseline difference analysis module is used to execute an adaptive data distribution baseline difference algorithm, specifically including: baseline data of objective physiological indicators collected at the beginning of treatment. For reference, the indicator data at each follow-up point of the same patient. Conduct normality assessment and within-group difference analysis; use test methods to assess the normality of the data and output the assessment results. According to the judgment result Select statistical methods If the data conforms to a normal distribution, then For paired t-tests, otherwise Wilcoxon signed-rank test; based on statistical methods Calculate follow-up data Compared with baseline data The within-group differences are calculated, and the statistical results of the differences are output. It generates a trend curve with time nodes on the horizontal axis and index values ​​on the vertical axis. .

[0040] As a further improvement to this technical solution, the efficacy feedback unit also includes a dynamic weight evaluation module and a cross-unit parameter feedback module, wherein:

[0041] The dynamic weight evaluation module is used to execute a dynamic weight algorithm driven by the correlation between indicators and treatment efficacy, specifically including: historical clinical case data stored based on the multi-source data integration unit. (Sample size ≥ 1000 cases) Tear secretion volume collected by the follow-up data acquisition module Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four types of objective physiological indicators were used to calculate the correlation coefficient between each objective physiological indicator and the degree of relief of dry eye disease using Pearson correlation analysis. ;by As a dynamic weight for each objective physiological indicator, combined with the improvement rate of each indicator. (including tear secretion) Tear film breakup time As a positive indicator, its For follow-up data Compared with baseline data Relative rate of change; corneal fluorescein staining score Concentration of inflammatory factors in tears As a negative indicator, its Baseline data With follow-up data (relative rate of change) to calculate the comprehensive efficacy index ;

[0042] The cross-unit parameter feedback module is used to realize cross-unit parameter iterative feedback, specifically including: the comprehensive efficacy index and the statistical results of the differences output by the adaptive baseline difference analysis module The data is transmitted to the AI ​​diagnostic analysis unit, where the feature weights and regression coefficients of the risk function of the associated model are corrected using gradient descent; when two consecutive follow-up nodes... Below the dynamic threshold of the corresponding node ( For follow-up nodes corresponding to historical valid cases in the multi-source data integration unit When the 75th percentile (obtained based on historical data statistics) is reached, a plan adjustment trigger signal is sent to the personalized plan generation unit. The trigger signal includes the improvement rate of each indicator. and correlation coefficient This provides data to support the optimization of treatment parameters for personalized treatment plans.

[0043] As a further improvement to this technical solution, the doctor-patient interaction unit includes a medical staff interaction module and a patient interaction module, wherein:

[0044] The medical staff interaction module is based on an encrypted database of the patient's unique identifier ID and a multi-source data integration unit. It uses a structured interface and operation log retention technology to enable multi-source data association query, manual review of AI analysis results, and operation traceability;

[0045] The patient-side interaction module is based on the structured treatment plan generated by the personalized plan generation unit and the appointment interface of the hospital's diagnosis and treatment system. It uses timeline visualization, multi-terminal push and standardized data collection technologies to realize the display of treatment information, medication follow-up reminders and patient symptom data collection.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention integrates multi-source data through a multi-source data integration unit based on the HL7FHIR interface of the medical data standardization protocol, the DICOM standard interface, and patient-side H5 page forms to connect with hospital electronic medical record systems, ophthalmic examination equipment, and patient-side data. The data integration module performs structured processing on the multi-source data, extracts key information from unstructured electronic medical record text through natural language processing, converts DICOM image metadata into numerical indicators, and completes field association and format unification according to the field definition specifications of dry eye clinical diagnosis and treatment standards. Then, the encrypted storage module uses the AES-256 encryption algorithm for field-level encryption, stores the data in an encrypted database according to a three-level index of patient ID-data type-collection time, and configures role-based access control permissions, thereby realizing the standardized integration of multi-source data for dry eye diagnosis and treatment and storage management that complies with medical data security specifications.

[0048] 2. This invention uses the feature screening module of the AI ​​diagnosis and analysis unit to calculate the variance inflation factor to eliminate highly collinear features and screen the top 30% of core features based on importance. The association model construction module uses multivariate regression analysis to establish a quantitative association model of core features. The risk analysis module calculates the matching score of each subtype based on the clinically recognized dry eye disease classification standards of aqueous hypoxia, evaporative hyperxia, and mixed subtype to determine the subtype to which the patient belongs. Furthermore, it constructs a disease progression risk function with corneal staining score and disease duration as independent variables through Kaplan-Meier analysis, calculates the probability of progression at different time points and classifies the risk level, thus achieving accurate subtype classification of dry eye disease and effective assessment of disease progression risk.

[0049] 3. This invention transforms clinical treatment guidelines for dry eye disease into a ternary structure stored in a medical knowledge graph through a structured module of the personalized treatment plan generation unit. This structure is then compared and calibrated with over 5 years of clinical treatment data to form a guideline knowledge base. The dynamic treatment plan generation module screens patients for allergies and contraindications, eliminating unsuitable treatments, and optimizes treatment parameters based on subtyping results and patient lifestyle data to generate structured personalized treatment plans. Simultaneously, the efficacy feedback unit collects objective physiological indicators according to the physiological cycle. After adaptive baseline difference analysis to determine normality, paired t-tests or Wilcoxon signed-rank tests are used to calculate intragroup differences and dynamic weight assessments. Pearson correlation analysis is used to calculate the correlation coefficient between indicators and disease remission and to calculate the comprehensive efficacy index. Finally, cross-unit parameter feedback corrects the model parameters of the AI ​​treatment analysis unit and triggers personalized plan adjustments, achieving precise generation of personalized treatment plans for dry eye disease, dynamic evaluation of efficacy, and iterative optimization of the plans.

[0050] 4. This invention enables multi-source data association queries, manual review of AI analysis results, and operation traceability through the medical staff interaction module of the doctor-patient interaction unit based on the patient's unique identifier. The patient interaction module enables the visualization of treatment information timeline, medication reminders, follow-up appointments, and collection of patient symptom data based on the structured treatment plan. This provides medical staff and patients with an interactive interface that conforms to the medical information display standards, ensuring the two-way transmission of diagnosis and treatment information and the effectiveness of manual intervention of AI analysis results. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0052] The meanings of the labels in the diagram are as follows:

[0053] 100. Multi-source data integration unit; 110. Data acquisition module; 120. Data integration module; 130. Encrypted storage module;

[0054] 200. AI Diagnosis and Analysis Unit; 210. Feature Screening Module; 220. Association Model Construction Module; 230. Risk Analysis Module;

[0055] 300. Personalized treatment plan generation unit; 310. Structured treatment guidelines module; 320. Dynamic treatment plan generation module;

[0056] 400. Treatment efficacy feedback unit; 410. Follow-up data collection module; 420. Adaptive baseline difference analysis module; 430. Dynamic weight assessment module; 441. Cross-unit parameter feedback module;

[0057] 500. Doctor-patient interaction unit; 510. Medical staff interaction module; 520. Patient interaction module. Detailed Implementation

[0058] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0059] like Figure 1 As shown, this embodiment provides an AI-assisted decision-making platform for personalized treatment of dry eye disease, including:

[0060] As a further explanation of this embodiment, after the patient completes all dry eye-related examinations (including tear secretion test, corneal fluorescein staining, meibomian gland imaging, tear film breakup time detection, etc.) and all examination data passes system verification (e.g., data format is compliant, key indicators are not missing), the multi-source data integration unit 100 automatically triggers the transmission of integrated patient data to the AI ​​diagnosis and treatment analysis unit 200, with a data transmission delay of no more than 5 minutes; after the AI ​​diagnosis and treatment analysis unit 200 completes the patient's condition classification, risk assessment, and other analyses and generates a preliminary personalized treatment plan, it immediately pushes a plan generation instruction to the personalized plan generation unit 300, with an instruction response time of no more than 10 seconds. Within seconds, when a patient completes a phase of treatment according to the personalized plan and undergoes follow-up, the efficacy feedback unit 400 will provide feedback on efficacy data (such as symptom improvement rate, changes in tear secretion, etc.) and model parameter correction requirements to the AI ​​diagnosis and analysis unit 200 within one hour after each follow-up data collection is completed (i.e., after the various efficacy indicator test data on the follow-up day are entered into the system). When a doctor manually adjusts the generated personalized diagnosis and treatment plan through the system interface and clicks the "Plan Confirm Adjustment" button, the personalized plan generation unit 300 will synchronize the adjustment information to the multi-source data integration unit 100 in real time to update the patient's diagnosis and treatment records and ensure that subsequent data integration and analysis are based on the latest plan information.

[0061] Multi-source data integration unit 100 is based on medical data standardization protocol. By connecting to the hospital's electronic medical record system, ophthalmology-specific examination equipment and patient-end data collection interface, it collects and integrates patients' clinical medical history data, objective eye examination data and lifestyle data, and stores them using medical data encryption standards to build a structured individualized dry eye dataset.

[0062] In this embodiment, the multi-source data integration unit 100 includes a data acquisition module 110, a data integration module 120, and an encrypted storage module 130 connected in sequence, wherein:

[0063] The data acquisition module 110 is based on the medical data standardization protocol. It connects to the hospital's electronic medical record system through the HL7 FHIR interface to collect clinical medical history data (including history of systemic diseases, history of ocular surface surgery, and long-term medication records). It connects to ophthalmic examination equipment (tear secretion meter, meibomian gland imaging meter, corneal topography instrument) through the DICOM standard interface to collect objective ocular examination data. It also collects lifestyle data (including average daily screen time, environmental humidity exposure records, and contact lens wearing cycle) through patient-side H5 page forms.

[0064] Specifically, the HL7 FHIRR4 interface can be used, and after security certification by the hospital's information department, it can connect to the electronic medical record system to collect clinical medical history data such as "history of systemic diseases," "history of ocular surface surgery," and "long-term medication records" through preset field mapping rules. The interface has an automatic retry mechanism, retrying within 1 minute in case of network interruption, and generating an exception log and pushing it to the administrator after 3 failures; the collection cycle is automatically triggered on the day of the patient's visit to obtain the patient's electronic medical record data for the past 3 years to ensure timeliness.

[0065] Furthermore, the tear secretion meter, meibomian gland imaging meter, and corneal topography meter all connect via the DICOM 3.0 interface, requiring the corresponding drivers to be installed on the device and bound to the platform IP. After the check is completed, the device automatically pushes the DICOM file to the data acquisition module 110. The data acquisition module 110 verifies the file integrity; if successful, it marks it as "to be integrated"; otherwise, it sends a retransmission reminder to the medical workstation.

[0066] Furthermore, patients can access the H5 page through the embedded portal in the hospital's official WeChat account or the outpatient dynamic QR code. The form has clear instructions for filling it out (such as "average daily screen time" which requires the cumulative number of hours spent on electronic devices). Before submission, the required fields and format are verified, and errors are prompted in real time. After submission, the data is marked as "patient self-reported" and transmitted to the data acquisition module 110.

[0067] The data integration module 120 is used to perform structured processing on the multi-source data output by the data acquisition module 110. Specifically, it includes: extracting key information from unstructured electronic medical record text using natural language processing technology; converting DICOM image metadata output by the examination device into numerical indicators; and completing the field association and format unification of multi-source data through field definition specifications based on dry eye clinical diagnosis and treatment standards (covering four core fields: patient basic information, clinical characteristics, examination indicators, and behavioral characteristics) to form a structured dataset containing a unique patient identifier.

[0068] Specifically, the data integration module 120 transforms multi-source heterogeneous data into a unified structured dataset, including:

[0069] Electronic medical record text structuring: Through the process of "word segmentation-entity recognition-structured annotation", entities such as disease, surgery, and time are extracted from the electronic medical record text and filled into the corresponding "clinical history data" subfield to complete the transformation of unstructured text.

[0070] Numericalization of DICOM metadata: Extract metadata related to dry eye diagnosis and treatment (such as tear secretion, meibomian gland absence area, and corneal surface regularity index) from DICOM files and convert them into numerical indicators; if the same indicator is measured multiple times, the average value is taken to ensure uniqueness.

[0071] Data association and format unification: Using "Patient ID" as the core association field, multi-source data is mapped to four core fields: "Patient basic information, clinical characteristics, examination indicators, and behavioral characteristics"; data format is unified (e.g., date is YYYY-MM-DD, and values ​​are retained to one decimal place); duplicate data of the same patient on the same day are retained with the later record to remove duplicates; finally, a dataset containing "Patient ID-Collection time-Data type-Structured field set" is generated and marked "Integration complete".

[0072] The encrypted storage module 130 uses the AES-256 encryption algorithm to perform field-level encryption on the structured dataset. The encrypted data is stored in an encrypted database with access log auditing function according to a three-level index rule of "patient ID-data type-collection time". It also configures role-based access control permissions, which comply with medical data security standards.

[0073] Specifically, the encrypted storage module 130 processes data in accordance with medical data security standards, as follows:

[0074] Field-level AES-256 encryption: Sensitive fields such as "patient name, contact information, and complete electronic medical record text" are encrypted using AES-256. The key is managed by the hospital's information department and rotated every 3 months. Historical data is re-encrypted during the rotation. After encryption, a checksum is generated. The integrity is verified when reading. If the data does not match, access is prohibited and a log is recorded.

[0075] Encrypted database storage and indexing: Data is stored in a physically isolated MySQL encrypted database within the hospital's intranet. A three-level index is created based on "patient ID-data type-collection time" to improve query efficiency. Access log auditing is enabled for the database, recording the visitor, time, operation type, and result. The logs are retained for 6 months for review.

[0076] Role-based access control: The system is divided into four roles: system administrator, attending ophthalmologist, ophthalmology nurse, and patient. Administrators are only responsible for database maintenance, doctors can view patient data and supplement clinical records, nurses can view objective data from assisted examinations, and patients can only view their own lifestyle data and authorized examination summaries. Changes to permissions require approval from the hospital's information department, and operation records are entered into the log.

[0077] It should be added that after data integration is completed, the multi-source data integration unit 100 automatically counts the number of successful and failed data collections and generates a "Data Integration Verification Report" which is pushed to the ophthalmologist's workstation. If the failure rate exceeds 5%, the system sends an alert to the administrator. Daily data backups are initiated, backing up the encrypted database data to a remote server. Database writes are prohibited during backups to ensure data security and availability.

[0078] The AI ​​diagnosis and analysis unit 200, based on ophthalmic clinical feature engineering and statistical learning algorithms, processes the dataset output by the multi-source data integration unit 100. It uses variance inflation factor to select the core features most strongly correlated with dry eye disease and employs multivariate regression analysis to establish a quantitative correlation model between features. Based on clinically recognized classification standards, it combines the feature weight values ​​output by the correlation model to achieve subtype classification. Finally, it constructs a disease progression risk function using Kaplan-Meier analysis, employing corneal staining score and disease duration as independent variables to calculate the probability of progression at different time points, thus forming a risk assessment result.

[0079] Understandably, the structured dataset output by the multi-source data integration unit 100 is used as input. Relying on ophthalmic clinical feature engineering and statistical learning algorithms, the core feature screening, quantitative association model construction, subtype classification and risk assessment are completed, providing key analysis results for the personalized solution generation unit 300. This unit is deployed on a dedicated computing node in the hospital's intranet (compliant with medical data privacy protection standards and not connected to the public network). The feature screening module 210, association model construction module 220, and risk analysis module 230 are linked sequentially according to the logic of "data input - feature processing - model construction - risk output".

[0080] In this embodiment, the AI ​​diagnosis and analysis unit 200 includes a feature selection module 210, which is used to select core features from the structured dataset output by the multi-source data integration unit 100. The feature selection module 210 performs the core feature selection by the following steps:

[0081] S210.1 Calculate the variance inflation factor (VIF) of each feature in the dataset and set a VIF threshold. ;

[0082] S210.2, Compare the VIF of each feature with Excluding VIF High collinearity;

[0083] S210.3. Based on the importance weighting table of clinical characteristics of dry eye, select those with VIF not exceeding [a certain value]. Furthermore, the top 30% of the most important features form the core feature set. .

[0084] Specifically, the core feature selection steps performed by the feature selection module 210 are as follows:

[0085] First, data preprocessing is performed. Missing values ​​in the dataset are filled with the "median of the feature". Categorical features such as "contact lens wearing type" and "history of ocular surface surgery" are converted into numerical features through one-hot encoding to ensure that the subsequent variance inflation factor (VIF) calculation can be performed normally.

[0086] Subsequently, the variance inflation factor (VIF) calculation function from Python's statsmodels library was used to calculate the VIF value for each of the preprocessed features (including clinical history features, examination indicator features, and lifestyle behavior features). The formula for calculating VIF is as follows: ,in For the first The variance inflation factor of a feature is used to measure the degree of collinearity between that feature and other features. For the first The coefficient of determination of a linear regression model constructed with one feature as the dependent variable and all other features in the dataset as independent variables, with a value range of [0,1].

[0087] Next, determine the VIF threshold. Based on the research findings on collinearity of dry eye-related characteristics in the "Chinese Clinical Diagnosis and Treatment Guidelines for Dry Eye (2023)," The value was set at 10 (a commonly used clinical threshold for determining collinearity). (When there is strong collinearity among features), and compare the VIF of each feature with... Excluding VIF If the number of high collinearity features remaining after removal is less than 20, then the number should be appropriately reduced. Re-execute VIF calculation and removal at step 8 to avoid excessive loss of effective features. After removal, generate a "High Collinearity Feature Removal Log" to record relevant information.

[0088] Finally, the core feature set was selected based on the "importance weighting table of clinical features of dry eye". The weighting table was reviewed and determined by three chief ophthalmologists in conjunction with the "Chinese Clinical Diagnosis and Treatment Guidelines for Dry Eye (2023)". It includes three columns: "Feature Name", "Clinical Importance Score (1-5 points, 5 points being the most important)", and "Associated Dry Eye Pathological Mechanism". During the screening process, the following columns were initially retained: The features are then sorted in descending order by "clinical importance score," and the top 30% of features are used to form the core feature set. (The formula is expressed as) ,in for The total number of features, Represents the first in the dataset One characteristic, This function represents the floor function, used to determine the boundary of the number of features in the "top 30% of importance"; after the screening is completed, it outputs the "Core Feature Set Confirmation Report" and pushes it to the result cache area of ​​the AI ​​diagnosis and analysis unit for the association model building module 220 to call.

[0089] In this embodiment, the AI ​​diagnosis and analysis unit 200 further includes an association model construction module 220, which is used to construct a core feature set based on the feature selection module 210. A quantitative correlation model is established using multifactor regression analysis; the correlation model construction module 220 establishes the quantitative correlation model through the following steps:

[0090] S220.1, From the core feature set Continuous features were extracted and used as independent variables. ;

[0091] S220.2, Quantification of dry eye symptoms severity Using the least squares method to solve for the parameters of the regression model, we obtain the constant term. and the weight values ​​of each feature ;

[0092] S220.3 Output the weight values ​​of each feature. , The larger the absolute value, the greater the influence of the corresponding feature on the dry eye condition.

[0093] Specifically, the steps for establishing a quantitative correlation model using multifactor regression analysis include:

[0094] First, from the core feature set Continuous features were extracted as independent variables and standardized. These continuous features included "tear secretion" and "average daily screen time." If categorical features existed, their numerical encoding was retained. Standardization was performed using the Z-score formula. ,in For the first The original values ​​of the core features, This is the mean of the feature in the historical training set. Let the standard deviation of this feature be the value within the historical training set. These are the standardized eigenvalues, with a mean of 0 and a standard deviation of 1, to eliminate the influence of differences in eigenvalue dimensions.

[0095] Next, define the dependent variable. (Quantification of the severity of dry eye symptoms) The calculation method is as follows: ,in Score the frequency of dry eye, such as daily = 20 points, weekly = 10 points, monthly = 5 points; The degree of foreign body sensation is scored, such as severe = 30 points, moderate = 20 points, mild = 10 points; The impact of blurred vision is scored, such as significant impact = 30 points, slight impact = 20 points, no impact = 10 points; The score for nighttime dry eye is as follows: frequent attacks = 20 points, occasional attacks = 10 points, no attacks = 5 points. The value range is 0-100 points;

[0096] Subsequently, a multifactor linear regression model was constructed: ;in, For model constants, For the first ( The range of values ​​is The weight values ​​of the core features, For continuous feature quantity, The random error term follows a mean of 0 and a variance of . The model follows a normal distribution; the model parameters are solved using the least squares method, and the solution formula is as follows: ,in For the parameter estimation vector, Let be the matrix of independent variables, with dimension ? , The first column is a vector of all 1s, representing the number of samples in the training set. For matrix transpose, The inverse of the matrix. As the dependent variable vector, the 1000 historical structured datasets were divided into training and validation sets in a 7:3 ratio before solving. The parameters were then solved using the training set, and the model fit was verified using the coefficient of determination R². The calculation formula is , For the first Each sample value, For predicted values, for The mean of the sample; and a validation set is required. Otherwise, readjust the core feature set. Then solve it again;

[0097] Finally, output the weight values ​​of each feature. With "core feature name" and "standardized" The values ​​and the direction of influence of features (positive or negative correlation) are stored in a table format and simultaneously pushed to the model parameter library of the medical staff interaction module 510 and the AI ​​diagnosis and treatment analysis unit for the risk analysis module 230 to call.

[0098] In this embodiment, the AI ​​diagnosis and analysis unit 200 further includes a risk analysis module 230, which is used to perform subtype classification and disease progression risk assessment of dry eye disease. The risk analysis module 230 performs subtype classification and risk assessment, including the following steps:

[0099] S230.1. Based on clinically recognized classification criteria for dry eye disease, the classification criteria for dry eye disease include aqueous hypoplasia. Excessive evaporation type Hybrid Combined with the feature weight values ​​output by the association model construction module 220 Calculate the matching score for each subtype. ,in Feature weight values absolute value and characteristics and fractals Related indicator variables The sum of the products of (1 when associated, 0 otherwise);

[0100] S230.2 Compare the matching scores of each subtype. The highest-scoring subtype is determined to be the patient's dry eye disease subtype;

[0101] S230.3 Constructing a disease progression risk function using Kaplan-Meier analysis. Corneal staining score Duration of illness As the independent variable, according to a preset time interval calculate Probability of corneal epithelial damage progression at any given time ;

[0102] S230.4, according to The values ​​are used to classify risk levels, forming a system that includes risk level and corresponding time points. The risk assessment results.

[0103] Specifically, the steps for subtyping and risk assessment are as follows:

[0104] First, dry eye disease subtypes are classified according to clinically recognized dry eye disease classification criteria (aqueous deficiency type). Excessive evaporation type Hybrid ), combined with the feature weight values ​​output by the association model construction module 220 Calculate the matching score for each subtype. The scoring formula is as follows: ,in , The absolute value of the weights is used to eliminate positive and negative influences. When the indicator variable is associated with the trait, the trait is related to the typology. Otherwise, it is 0, where Related to "tear secretion" and "duration of illness". Correlate with "degree of meibomian gland absence" and "tear film breakup time". Related to "tear secretion volume", "degree of meibomian gland absence", and "corneal staining score";

[0105] Subsequently, comparison The highest-scoring subtype is used to determine the patient's dry eye disease subtype. The specific formula is as follows: The judgment result is output in the form of "subtype name + matching score details", and is transmitted in real time to the personalized solution generation unit 300 and displayed in the medical staff interaction module 510. It supports medical staff to manually review and record the reasons for adjustment.

[0106] Next, a disease progression risk function is constructed. This was achieved through Kaplan-Meier analysis; firstly, data from 1000 cases with complete follow-up records (including corneal fluorescein staining scores) were retrieved from the encrypted database of the multi-source data integration unit 100. "Duration of illness" "Progress of corneal epithelial damage", with preset time intervals. Months (based on the cycle of dry eye condition changes, time points) (months, 3 months, 6 months), and then the KaplanMeierFitter class from Python's lifelines library was used to calculate. Progress probability at time step The calculation function is: ,in For the first The timing of the occurrence of a "corneal epithelial damage progression event". for The number of cases that are progressing at any given time. for The number of cases that have not progressed before a certain point in time. The calculation is based on the patient's current condition. Value and The value determines the hierarchical group to which it belongs (e.g., by...). Divided into three groups: 0-5 points, 6-10 points, and 11-15 points, according to... Divided into three groups: <3 months, 3-6 months, and >6 months, data from the corresponding groups were retrieved for calculation. ;

[0107] Finally, risk levels are classified according to... The value is determined (e.g.) Low risk Medium risk. (High risk), forming a system containing "time nodes" "Probability of Progress" The risk assessment results of “risk level” are simultaneously transmitted to the personalized plan generation unit 300 and the medical staff interaction module 510 for medical staff to refer to the follow-up frequency.

[0108] It should be added that the AI ​​diagnosis and analysis unit 200 also has an operation verification and result transmission mechanism, which specifically includes: in terms of operation verification, each module automatically performs logical verification after outputting results, including checking whether the VIF calculation of the feature screening module 210 meets the requirements. Regression model in module 220 of the association model construction Risk analysis module 230 If the result falls within the 0-100% range, and the verification fails, the system returns to the previous step for reprocessing and generates an "abnormal log" which is pushed to the system administrator's terminal. Simultaneously, regarding result traceability, the calculation process of all modules (including parameter values, intermediate results, and formula application processes) is marked with a "generation timestamp" and a "data source identifier" (e.g., "based on patient structured data collected on 2024-05-10"), stored in an encrypted result library, and accessible for review by medical staff through an interactive module. Regarding result transmission, the final "dry eye disease subtype classification result + disease progression risk assessment result" from the AI ​​diagnosis and analysis unit 200 is transmitted in real-time to the personalized treatment plan generation unit 300 via a secure hospital intranet channel, providing core evidence for subsequent personalized treatment plan generation. A data interface is also reserved to allow the efficacy feedback unit 400 to call upon this result for model parameter correction, ensuring the closed-loop application of diagnosis and treatment data.

[0109] Personalized treatment plan generation unit 300 is based on a rule base constructed from the clinical diagnosis and treatment guidelines for dry eye disease. It combines the subtyping results and risk assessment output by the AI ​​diagnosis and treatment analysis unit 200, screens patients' allergy history and contraindications to exclude unsuitable treatment methods, optimizes the parameters of applicable treatment methods, and generates personalized treatment plans that include specific treatment measures, implementation guidelines and expected effect assessments.

[0110] In this embodiment, the personalized treatment plan generation unit 300 includes a treatment guideline structuring module 310. The treatment guideline structuring module 310 is used to convert clinical treatment guidelines for dry eye disease into computable structured data, specifically including:

[0111] Natural language processing technology was used to segment the guideline text, identify entities, and extract relationships to extract the correspondence between treatment methods and classifications, as well as the correlation threshold between treatment intensity and risk level.

[0112] The extracted relationships are converted into a triplet data structure and stored in the medical knowledge graph. The subject is the subtype or risk level, the object is the treatment method or parameter, and the relationship is a logical association of "adaptation" or "correspondence".

[0113] By comparing the structured data with more than 5 years of clinical diagnosis and treatment data, the matching degree between the structured data and actual diagnosis and treatment behavior is verified. When the matching degree reaches 90% or above, the calibration is completed, and a guideline knowledge base that can be directly called is formed.

[0114] Specifically, the structured module 310 of the clinical practice guidelines for dry eye disease is used to transform these guidelines into computable structured data, providing a knowledge base for subsequent protocol generation. It uses authoritative guidelines such as the "Chinese Clinical Practice Guidelines for Dry Eye (2023)" as the processing objects, employing natural language processing technology in a step-by-step process, specifically including:

[0115] First, the entire guideline was split into core terms such as “treatment methods”, “dry eye disease classification”, “risk level”, and “treatment parameter threshold” using the jieba word segmentation tool to ensure that no key information was omitted.

[0116] Subsequently, a BERT-based entity recognition model was used to identify key entities in the text, such as "aqueous dry eye", "intense pulsed light therapy", "moderate disease risk", and "daily frequency of artificial tears". Then, a relation extraction algorithm was used to extract the relationships between entities, such as "aqueous dry eye is suitable for artificial tears" and "moderate risk corresponds to moderate treatment intensity".

[0117] Next, the extracted relationships are transformed into a triplet (subject, relation, object) structure and stored in a medical knowledge graph constructed using the Neo4j graph database, where the subject is "dry eye disease classification (e.g.)". The object is either "aqueous deficiency type" or "disease risk level (e.g., medium risk)", or "treatment method (e.g., cyclosporine eye drops)" or "treatment parameters (e.g., 4 drops per day)", and the relationship is either "fitting" or "corresponding" in two logical associations.

[0118] Finally, using field matching and statistical comparison algorithms, the relationships between "classification-treatment methods" and "risk level-treatment intensity" in the knowledge graph are matched case by case with the hospital's clinical diagnosis and treatment data (including patient classification, actual treatment methods and parameters) over the past 5 years. The matching degree is calculated (matching degree = number of matched cases / total number of cases × 100%). When the matching degree reaches 90% or above, the structured data is considered to match the actual diagnosis and treatment behavior well, and the guideline knowledge base calibration is completed. If the matching degree is less than 90%, the entity recognition and relationship extraction rules are adjusted and the data is reprocessed and verified until the matching degree reaches the standard. Finally, a guideline knowledge base that can be directly called by the dynamic solution generation module 320 is formed.

[0119] In this embodiment, the personalized solution generation unit 300 further includes a dynamic solution generation module 320, which is used to generate personalized solutions by combining multi-source data, specifically including:

[0120] Contraindication screening: By comparing the allergy history records in the multi-source data integration unit 100 with the treatment components or operational characteristics in the guideline knowledge base through a field matching algorithm, contraindication options with a matching degree of 80% or higher are automatically removed.

[0121] Parameter optimization: Based on the classification results output by the AI ​​diagnosis and analysis unit 200, the baseline parameters of the corresponding classification in the guide knowledge base are called, and a parameter adjustment model is constructed by combining the patient's average daily eye use time and environmental humidity data.

[0122] Solution output: Integrate and optimize treatment measures, implementation guidelines, and expected effect evaluation indicators to generate a structured solution document with a timestamp, and link it to the guideline knowledge base reference nodes and parameter adjustment basis.

[0123] Specifically, the dynamic treatment plan generation module 320 combines patient data from the multi-source data integration unit 100 and the classification and risk results from the AI ​​diagnosis and treatment analysis unit 200 to generate personalized treatment plans, which specifically include:

[0124] First, contraindication screening is conducted. Using a field matching algorithm, the "allergy history records" of patients in the multi-source data integration unit 100 (such as allergies to preservatives or drug ingredients) are compared with the "ingredients or operational characteristics" of various treatment methods in the guideline knowledge base (such as artificial tears containing specific preservatives, and contraindications for intense pulsed light operation). A matching threshold of 80% is set. If the matching degree between a patient's allergy history and the contraindication characteristics of a certain treatment method is ≥80%, the treatment method is automatically removed (e.g., if a patient is allergic to "polyacrylic acid" preservatives and the matching degree with a certain artificial tear is 85%, then the artificial tear is marked as contraindicated and excluded).

[0125] Then, treatment parameters were optimized based on the classification results output by the AI ​​diagnostic analysis unit 200 (such as...). (Mixed type), retrieve the baseline treatment parameters for the corresponding subtype from the guideline knowledge base (such as... The baseline regimen is "artificial tears four times daily + intense pulsed light twice weekly").

[0126] Next, by combining the patient's "average daily eye use time" and "environmental humidity exposure record" from the multi-source data integration unit 100, the baseline parameters are adjusted through a set of preset rules developed by ophthalmology clinical experts based on guidelines and experience. For example, if the patient's average daily eye use time is >8 hours, the number of artificial tear drops is increased by 1 on the baseline; if the patient is in an environment with humidity <40% for a long time, the interval between intense pulsed light therapy is shortened by 0.5 weeks on the baseline.

[0127] Finally, by integrating "treatment methods retained after contraindication screening" and "treatment intensity after parameter optimization," a structured protocol document is generated, which includes treatment measures (such as specific drugs and procedures), implementation guidelines (such as instillation time and operating procedures), and expected effect evaluation indicators (such as symptom relief time and tear secretion improvement targets). The document is stored in PDF format and is linked to the guideline knowledge base reference nodes (marking the original source of the guideline for the treatment method) and the basis for parameter adjustment (such as "due to an average daily eye use time of 9 hours, the number of artificial tears is adjusted from 4 to 5 times"). It is then pushed to the medical staff interaction module 510 for doctors to review, and after approval, it can be transmitted to the patient interaction module 520.

[0128] It should be added that the personalized treatment plan generation unit 300 also includes a plan verification mechanism to ensure the rationality and compliance of the generated plan. Specifically, this includes: after the plan is generated, automatically comparing the treatment methods, parameters, and core relationships with the guidelines and knowledge base to check for obvious conflicts (such as using low-intensity treatment for high-risk patients without special instructions); simultaneously, logically verifying the basis for parameter adjustments to ensure that data such as "average daily eye use time" and "ambient humidity" are effectively incorporated into the adjustment process. If an anomaly is found during verification, the system prompts the dynamic plan generation module 320 to re-screen for contraindications or optimize parameters until the plan meets the verification requirements.

[0129] The efficacy feedback unit 400 is based on the physiological cycle pattern of dry eye disease monitoring and improved statistical analysis logic. Through multi-node objective physiological indicator collection, adaptive data distribution baseline difference algorithm, indicator-efficacy correlation driven dynamic weight algorithm, and cross-unit parameter iterative feedback, it realizes the quantitative evaluation and dynamic adjustment of dry eye disease treatment efficacy. It is used to correct the parameters of the correlation model and risk function of AI diagnosis and treatment analysis unit 200, and to provide objective data support and algorithm output basis for the optimization of treatment plan of personalized plan generation unit 300.

[0130] Understandably, the efficacy feedback unit 400, as the core unit of this platform to achieve the closed loop of "quantification of treatment effect - correction of model parameters - dynamic optimization of plan", is based on the physiological cycle of dry eye disease monitoring (clinical observation shows that dry eye disease changes significantly within 1-3 months after treatment, and there is a short adjustment cycle of about 2 weeks) and improved statistical analysis logic. Through multi-node data collection, adaptive difference analysis, dynamic weight evaluation and cross-unit feedback, it completes efficacy quantification and plan iteration support, and provides objective basis for model optimization of AI diagnosis and treatment analysis unit 200 and parameter adjustment of personalized plan generation unit 300. This unit is linked with multi-source data integration unit 100, AI diagnosis and treatment analysis unit 200 and personalized plan generation unit 300 through the hospital intranet security interface. Follow-up data collection module 410, adaptive baseline difference analysis module 420, dynamic weight evaluation module 430 and cross-unit parameter feedback module 440 are executed in sequence according to the logic of "data collection - difference analysis - efficacy calculation - parameter feedback".

[0131] In this embodiment, the efficacy feedback unit 400 includes a follow-up data acquisition module 410 and an adaptive baseline difference analysis module 420, wherein:

[0132] The follow-up data acquisition module 410 is used to collect tear secretion data by connecting to specialized ophthalmic examination equipment at preset time points after treatment initiation (which correspond to the physiological cycle of dry eye disease progression). Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four categories of objective physiological indicators; each collected objective physiological indicator is associated with a unique patient ID and a collection timestamp. The data is transmitted to the encrypted database via the standard interface of the multi-source data integration unit 100. Storage is compatible with the device interface and data storage specifications of the multi-source data integration unit 100;

[0133] Specifically, the follow-up data acquisition module 410 is used to connect to ophthalmic examination equipment at preset time points after treatment initiation to collect four types of objective physiological indicators, including:

[0134] Based on the physiological cycle of dry eye disease progression (short-term symptom adjustment cycle of approximately 2 weeks, and mid-term efficacy evaluation cycle of approximately 1 month), the preset time points were determined as "1 week, 2 weeks, 1 month, and 3 months after treatment initiation" to ensure coverage of both short-term adjustment and mid-term efficacy observation; subsequently, corresponding ophthalmology-specific examination equipment was used to measure tear secretion. Tear film breakup time was measured using a tear secretion meter according to the Schirmer I test (without anesthetic instillation). Data was collected using a tear film breakup timer and the fluorescein staining method (recording the time from blinking to tear film breakup). Corneal fluorescein staining score was determined. Tear fluid inflammatory factor concentrations were measured using corneal fluorescein staining according to the Oxford scoring system (0-15 points). Collected using a tear film inflammatory factor detection device (detecting dry eye-related inflammatory factors such as IL-6);

[0135] Furthermore, all collected objective physiological indicators are associated with the patient's unique identifier ID (consistent with the patient ID of the multi-source data integration unit 100) and the collection timestamp. (Accurate to the minute), transmitted to the encrypted database via the HL7FHIR interface or DICOM standard interface of the multi-source data integration unit 100. The storage and data format are fully compatible with the device interface and data storage specifications of the multi-source data integration unit 100, ensuring the uniformity of the format when retrieving data in the future.

[0136] The adaptive baseline difference analysis module 420 is used to execute an adaptive data distribution baseline difference algorithm, specifically including: baseline data of objective physiological indicators acquired at the beginning of treatment. For reference, the indicator data at each follow-up point of the same patient. Conduct normality assessment and within-group difference analysis; use test methods to assess the normality of the data and output the assessment results. According to the judgment result Select statistical methods If the data conforms to a normal distribution, then For paired t-tests, otherwise Wilcoxon signed-rank test; based on statistical methods Calculate follow-up data Compared with baseline data The within-group differences are calculated, and the statistical results of the differences are output. It generates a trend curve with time nodes on the horizontal axis and index values ​​on the vertical axis. .

[0137] Specifically, the adaptive baseline difference analysis module 420 is used to execute an adaptive data distribution baseline difference algorithm to treat the baseline data of the initially acquired objective physiological indicators. For reference, the indicator data at each follow-up point of the same patient. Conducting normality assessment and within-group difference analysis specifically includes:

[0138] First, determine the baseline data. The baseline data were obtained from four types of objective physiological indicators collected on the day the patient's treatment was initiated. ,Right now ,in Initial tear secretion, Initial tear film breakup time, For initial corneal fluorescein staining score, Initial tear film inflammatory factor concentrations; data from a specific follow-up point. (For example, data from one week's nodes);

[0139] Subsequently, the Shapiro-Wilk test was used to analyze the data of each group (such as the 1-week milestone for all patients). To determine the normality of the data, the test statistic formula is: ,in This is the Shapiro-Wilk test statistic (with a value range of (0,1], where the closer to 1 is, the more the data conforms to a normal distribution). This represents the patient sample size for the follow-up data in this group. The first number of samples sorted from smallest to largest One value, For the sake of sample size The determining constant (generated via Shapiro-Wilk distribution table or statistical software). For the original sample data, The mean of the sample data; calculated using statistical software. Value and Correspondence Value, if Then output the judgment result. (The data follows a normal distribution) If Then output (The data does not conform to a normal distribution), and based on the judgment result Select statistical methods The details are as follows:

[0140] like (like , Follow-up data usually conform to a normal distribution. For paired t-tests, the calculation formula is: ,in Here, is the t-test statistic, and is the mean of the differences between the paired data. , For the first For example, patient follow-up index values, Its baseline index value, The standard deviation of the differences between paired data is denoted as . , , The patient sample size; the t-value was calculated using statistical software. value;

[0141] like (like , Follow-up data may not conform to a normal distribution. The specific steps for the Wilcoxon signed-rank test are: calculate the paired difference (positive index) , for negative indicators , for ),neglect Pairing of 0, for the remainder Sort and assign ranks (same) (Take the average rank), and calculate the rank sum of the positive differences respectively. Rank sum of negative differences ,Pick As a test statistic, it is calculated using statistical software. Value and Value; based on the selected statistical method Calculate follow-up data Compared with baseline data The within-group differences are calculated, and the statistical results of the differences are output. (Including indicator name, follow-up node, difference value, statistic, Values, such as "Tear secretion S: Week 1 node, difference value +2mm / 5min", ( ), and generate a trend curve. (The horizontal axis represents the follow-up time points, and the vertical axis represents the indicator values ​​and corresponding units. Each indicator corresponds to a curve.)

[0142] It is understood that in the adaptive baseline difference analysis module 420 of this embodiment, The value is a core indicator for measuring the probability of a statistical hypothesis being true. In determining normality, it is calculated using the Shapiro-Wilk test. The value represents the probability (taking values ​​[0,1]) of observing the current or more extreme data, assuming the null hypothesis that "the data follows a normal distribution" holds true. As a standard, Then the data is determined to conform to a normal distribution. Conversely, it does not conform to ( Meanwhile, in the analysis of differences within groups, if Paired t-test The value represents the probability that the null hypothesis "the mean difference between follow-up and baseline data is 0" is true; if Wilcoxon signed-rank test The value represents the probability that the null hypothesis "the distribution of the difference between follow-up and baseline data is symmetrical to 0" holds true. When When the difference is statistically significant, the output difference statistics are... Includes This value provides a quantitative basis for subsequent efficacy evaluation.

[0143] In this embodiment, the therapeutic effect feedback unit 400 further includes a dynamic weight evaluation module 430 and a cross-unit parameter feedback module 440, wherein:

[0144] The dynamic weight assessment module 430 is used to execute a dynamic weight algorithm driven by the correlation between indicators and treatment efficacy, specifically including: historical clinical case data stored in the multi-source data integration unit 100. (Sample size ≥ 1000 cases) Tear secretion volume collected by follow-up data acquisition module 410 Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four types of objective physiological indicators were used to calculate the correlation coefficient between each objective physiological indicator and the degree of relief of dry eye disease using Pearson correlation analysis. ;by As a dynamic weight for each objective physiological indicator, combined with the improvement rate of each indicator. (including tear secretion) Tear film breakup time As a positive indicator, its For follow-up data Compared with baseline data Relative rate of change; corneal fluorescein staining score Concentration of inflammatory factors in tears As a negative indicator, its Baseline data With follow-up data (relative rate of change) to calculate the comprehensive efficacy index ;

[0145] Specifically, the dynamic weight evaluation module 430 is used to execute a dynamic weight algorithm driven by the correlation between indicators and efficacy to calculate the comprehensive efficacy index. Specifically, it includes:

[0146] First, from the encrypted database of the multi-source data integration unit 100 Data on dry eye disease treatment cases from the past 5 years were retrieved as historical clinical case data. The selection criteria were "complete pre-treatment baseline data, at least 3 follow-up data, and clear records of disease remission", ensuring a sample size of ≥1000 cases;

[0147] Subsequently, Pearson correlation analysis was used to calculate the correlation coefficients between four types of objective physiological indicators and the degree of relief of dry eye disease. The degree of relief of dry eye disease is quantified by the rate of change in OSDI (Ocular Surface Disease Index) scores before and after treatment, as shown in the following formula:

[0148] ;

[0149] The calculation will include the change in indicators for each case (e.g.) Pearson correlation analysis was performed between the indicators and the degree of disease remission, and the results for each indicator were output. value( The value ranges from [-1, 1], with a larger absolute value indicating a higher correlation between the indicator and disease remission. absolute value As dynamic weights for each indicator (eliminating the influence of positive and negative correlation directions on the weights);

[0150] Next, the improvement rate of each indicator is calculated. Among them, positive indicators (tear secretion) Tear film breakup time (The higher the indicator value, the better the condition.) Negative indicators (corneal fluorescein staining score) Concentration of inflammatory factors in tears (The lower the indicator value, the better the condition.) ;

[0151] Finally, the comprehensive efficacy index is calculated using the following formula. ( The value ranges from [0, 100%], with higher values ​​indicating better overall therapeutic effects.

[0152] ;

[0153] in, Indicates the amount of tear secretion The absolute value of the Pearson correlation coefficient with the degree of relief of dry eye disease is used to quantify the strength of the association between "changes in tear secretion" and "disease relief" (the larger the absolute value, the stronger the association). Indicates the amount of tear secretion The improvement rate, because tear secretion volume is a positive indicator that "the higher the index value, the better the condition," is calculated as follows: , The amount of tear secretion at the follow-up nodes. Baseline tear secretion on the day treatment was initiated;

[0154] Indicates tear film breakup time The absolute value of the Pearson correlation coefficient with the degree of remission of dry eye disease quantifies the strength of the association between "changes in tear film breakup time" and "disease remission"; Indicates tear film breakup time The improvement rate was assessed, with tear film breakup time as a positive indicator. The calculation method was as follows: , The tear film breakup time at the follow-up node. Baseline tear film breakup time;

[0155] Indicates corneal fluorescein staining score The absolute value of the Pearson correlation coefficient with the degree of remission of dry eye disease quantifies the strength of the association between "changes in corneal fluorescein staining score" and "disease remission"; Indicates corneal fluorescein staining score The improvement rate, because the corneal fluorescein staining score is a negative indicator that "the lower the index value, the better the condition," is calculated as follows: , For corneal fluorescence staining scores at follow-up nodes, Baseline corneal fluorescence staining score;

[0156] Indicates the concentration of inflammatory factors in tears The absolute value of the Pearson correlation coefficient with the degree of remission of dry eye disease quantifies the strength of the association between "changes in the concentration of inflammatory factors in tears" and "remission of disease". Indicates the concentration of inflammatory factors in tears The improvement rate, with tear inflammatory factor concentration as a negative indicator, is calculated as follows: ,in The concentration of inflammatory factors in tears at follow-up points. This represents the baseline concentration of inflammatory factors in the tear film.

[0157] This embodiment also provides the following example: For example, a patient ,but .

[0158] The cross-unit parameter feedback module 440 is used to implement cross-unit parameter iterative feedback, specifically including: the comprehensive efficacy index and the statistical results of the differences output by the adaptive baseline difference analysis module 420 The data is transmitted to the AI ​​diagnosis and analysis unit 200, where the feature weights and regression coefficients of the risk function in the correlation model are corrected using the gradient descent method; when two consecutive follow-up nodes... Below the dynamic threshold of the corresponding node ( For the follow-up nodes corresponding to 100 historical valid cases in the multi-source data integration unit When the 75th percentile (obtained based on historical data statistics) is reached, a scheme adjustment trigger signal is sent to the personalized scheme generation unit 300. The trigger signal includes the improvement rate of each indicator. and correlation coefficient This provides data to support the optimization of treatment parameters for personalized treatment plans.

[0159] Specifically, the cross-unit parameter feedback module 440 is used to implement cross-unit parameter iterative feedback, which includes:

[0160] First, the comprehensive efficacy index and the statistical results of the differences output by the adaptive baseline difference analysis module 420 The data is transmitted to the AI ​​diagnosis and analysis unit 200, where the feature weights and regression coefficients of the risk function of the associated model are corrected using the gradient descent method. The correction process is based on the correlation between the model's predicted efficacy and the actual efficacy. The objective is to minimize the bias, with a learning rate set to 0.01 (a commonly used learning rate for adjusting parameters in clinical models to avoid excessive parameter adjustments). The iteration count continues until the bias converges (the bias is less than a preset threshold of 0.05). For example, if the actual... The efficacy value was lower than that predicted by the association model, and Displays tear secretion volume If the improvement rate is low, then the corresponding improvement should be made in the correlation model. Feature weights ;

[0161] Subsequently, the dynamic threshold was determined. From the historical valid case data of the multi-source data integration unit 100, the data is grouped according to follow-up nodes (e.g., 1 week, 2 weeks, 1 month, 3 months), and the results for each group are calculated. The 75th percentile is used as the corresponding node. (e.g., historical valid cases at the 1-month mark) The quantiles are 15%, 18%, 20%, 22%, and 25%, and the 75th percentile is 22%. Therefore, this node... ); when two consecutive follow-up nodes Lower than the corresponding node (e.g., 1-month node) 2-month node If the signal is positive, a solution adjustment trigger signal is sent to the personalized solution generation unit 300. The trigger signal includes the improvement rate of each indicator. and correlation coefficient (e.g., tear film breakup time) Improvement rate Correlation coefficient This provides data support for the personalized treatment plan generation unit 300 to optimize treatment parameters in a targeted manner (such as increasing the frequency of treatments that promote tear film stability).

[0162] It should be added that the efficacy feedback unit 400 also has a data integrity verification and anomaly handling mechanism to ensure the reliability of efficacy evaluation and feedback, specifically including:

[0163] During the follow-up data collection phase, if data for a certain indicator is missing (e.g., the patient did not complete the tear inflammatory factor test on time), the system will automatically send a "data collection reminder" to the medical staff interaction module 510 and retain the collection entry for that node until the data is complete.

[0164] In difference analysis and During the calculation phase, if an error occurs in the selection of statistical methods (such as misusing the Wilcoxon test on data that conforms to a normal distribution) or Calculation anomalies (such as (If the value is negative), the system automatically backtracks to previous steps to check the normality determination result or The calculation logic was corrected and recalculated; all data (follow-up data, difference results, All parameters (including feedback parameters) are tagged with "generation timestamp", "data collector ID", and "calculation algorithm version" and stored in an encrypted database. The "Efficacy Feedback" sub-database supports medical staff to access and trace data through the doctor-patient interaction unit 500, which complies with the medical data traceability standard.

[0165] The doctor-patient interaction unit 500 provides an interactive interface that conforms to the medical information display standards. It provides medical staff with an operation entry point for data query, treatment plan editing, and efficacy analysis, and supports manual review and adjustment of AI analysis results. For patients, it provides functions such as visual display of treatment plans, medication reminders, follow-up appointments, and symptom recording, realizing two-way transmission of diagnosis and treatment information.

[0166] Understandably, the doctor-patient interaction unit 500, as the core carrier for the two-way transmission of medical information, is based on the hospital's intranet security architecture and medical information display standards. Through the collaboration of the medical staff interaction module 510 and the patient interaction module 520, it provides targeted operation entry points for medical staff and patients respectively, realizing a closed loop of "efficient management of medical data by medical staff + convenient access to treatment information by patients". This unit is linked with the multi-source data integration unit 100, the AI ​​diagnosis and treatment analysis unit 200, and the personalized solution generation unit 300 through standardized interfaces to ensure the consistency and security of data transmission.

[0167] In this embodiment, the doctor-patient interaction unit 500 includes a doctor-caregiver interaction module 510 and a patient-side interaction module 520, wherein:

[0168] The medical staff interaction module 510 is based on the patient's unique identifier ID and the encrypted database of the multi-source data integration unit 100. It uses a structured interface and operation log retention technology to enable multi-source data association query, manual review of AI analysis results, and operation traceability;

[0169] Specifically, the medical staff interaction module 510 is based on the patient's unique identifier ID and the encrypted database of the multi-source data integration unit 100. It uses a structured interface and operation log retention technology: After logging in, medical staff can enter the patient ID or scan the medical barcode to query multi-source data (clinical history, eye examination indicators) and AI analysis results (typing, risk assessment), which are displayed in sections on the interface; the AI ​​results can be "reviewed and confirmed" or "adjusted and modified". Modifications require filling in the reason and generating a record; all operations are logged (including the operator, time, and content). The logs are stored in an encrypted database and cannot be tampered with, supporting traceability.

[0170] The patient-side interaction module 520, based on the structured treatment plan of the personalized treatment plan generation unit 300 and the appointment interface of the hospital's diagnosis and treatment system, uses timeline visualization, multi-terminal push and standardized data collection technology to realize the display of treatment information, medication follow-up reminders and patient symptom data collection.

[0171] Specifically, the patient-side interaction module 520, based on the structured treatment plan of the personalized plan generation unit 300 and the hospital's medical system appointment interface, uses timeline visualization, multi-terminal push, and standardized data collection technologies: After logging in, the homepage timeline displays treatment information (medication, treatment, and follow-up nodes), and clicking on a node displays the original plan; one hour before medication / follow-up, a reminder is sent via WeChat and SMS, and clicking the reminder displays the medication guide; patients record symptoms through a standardized form (selecting symptom type, severity, and frequency), and after submission, the data is transmitted to the multi-source data integration unit 100; clicking the follow-up node redirects to the hospital's registration system for appointment, and upon successful appointment, the data is synchronized to the timeline and the medical staff's end.

[0172] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0173] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-assisted decision-making platform for personalized treatment of dry eye disease, characterized in that, include: Multi-source data integration unit (100), which is based on medical data standardization protocol, collects and integrates patients' clinical medical history data, objective eye examination data and lifestyle data by connecting to the hospital's electronic medical record system, ophthalmology-specific examination equipment and patient-end data collection interface, and stores it using medical data encryption standards to construct a structured individualized dry eye dataset; The AI ​​diagnosis and treatment analysis unit (200) processes the dataset output by the multi-source data integration unit (100) based on ophthalmic clinical feature engineering and statistical learning algorithms. It selects the core features with the highest correlation to dry eye disease through variance inflation factor and establishes a quantitative correlation model between features using multivariate regression analysis. Based on the clinically recognized classification standards, it achieves subtype classification by combining the feature weight values ​​output by the correlation model. It constructs a disease progression risk function through Kaplan-Meier analysis, uses corneal staining score and disease duration as independent variables, calculates the progression probability at different time points, and forms a risk assessment result. Personalized treatment plan generation unit (300) is based on the rule base constructed by the clinical diagnosis and treatment guidelines for dry eye disease. It combines the subtyping results and risk assessment output by the AI ​​diagnosis and treatment analysis unit (200), screens the patient's allergy history and contraindications to exclude unsuitable treatment methods, optimizes the parameters of the applicable treatment methods, and generates a personalized treatment plan that includes specific treatment measures, implementation specifications and expected effect assessment. The efficacy feedback unit (400) is based on the physiological cycle pattern of dry eye disease monitoring and improved statistical analysis logic. Through the collection of objective physiological indicators at multiple nodes, the baseline difference algorithm of adaptive data distribution, the dynamic weight algorithm driven by indicator-efficacy correlation, and cross-unit parameter iterative feedback, it realizes the quantitative evaluation and dynamic adjustment of the efficacy of dry eye disease treatment. It is used to correct the parameters of the correlation model and risk function of the AI ​​diagnosis and treatment analysis unit (200) and to provide objective data support and algorithm output basis for the optimization of treatment plan of the personalized plan generation unit (300). The doctor-patient interaction unit (500) is used to provide an interactive interface that conforms to the medical information display specifications, provide medical staff with an operation entry for data query, plan editing and efficacy analysis, and support manual review and adjustment of AI analysis results; and provide patients with functions such as treatment plan visualization, medication reminder, follow-up appointment and symptom recording, so as to realize two-way transmission of diagnosis and treatment information.

2. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 1, characterized in that, The multi-source data integration unit (100) includes a data acquisition module (110), a data integration module (120), and an encrypted storage module (130) connected in sequence, wherein: The data acquisition module (110) is based on the medical data standardization protocol. It connects to the hospital's electronic medical record system through the HL7 FHIR interface to collect clinical medical history data, connects to ophthalmology-specific examination equipment through the DICOM standard interface to collect objective eye examination data, and collects lifestyle behavior data through the patient-side H5 page form. The data integration module (120) is used to perform structured processing on the multi-source data output by the data acquisition module (110), specifically including: extracting key information from unstructured electronic medical record text through natural language processing technology, converting DICOM image metadata output by the examination device into numerical indicators, and completing the field association and format unification of multi-source data through the field definition specification based on the dry eye clinical diagnosis and treatment standard, forming a structured dataset containing the patient's unique identifier. The encrypted storage module (130) uses the AES-256 encryption algorithm to perform field-level encryption on the structured dataset. The encrypted data is stored in an encrypted database with access log auditing function according to the three-level indexing rule of "patient ID-data type-collection time". It also configures role-based access control permissions, which comply with medical data security standards.

3. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 2, characterized in that, The AI ​​diagnosis and analysis unit (200) includes a feature filtering module (210), which is used to filter the core features of the structured dataset output by the multi-source data integration unit (100). The feature filtering module (210) performs core feature filtering by the following steps: S210.1 Calculate the variance inflation factor (VIF) of each feature in the dataset and set a VIF threshold. ; S210.2, Compare the VIF of each feature with Excluding VIF High collinearity; S210.

3. Based on the importance weighting table of clinical characteristics of dry eye, select those with VIF not exceeding [a certain value]. Furthermore, the top 30% of the most important features form the core feature set. .

4. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 3, characterized in that, The AI ​​diagnosis and analysis unit (200) further includes an association model construction module (220), which is used to construct a core feature set based on the feature selection module (210). A quantitative correlation model is established using multifactor regression analysis; the correlation model construction module (220) establishes the quantitative correlation model by the following steps: S220.1, From the core feature set Continuous features were extracted and used as independent variables. ; S220.2, Quantification of dry eye symptoms severity Using the least squares method to solve for the parameters of the regression model, we obtain the constant term. and the weight values ​​of each feature ; S220.3 Output the weight values ​​of each feature. , The larger the absolute value, the greater the influence of the corresponding feature on the dry eye condition.

5. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 4, characterized in that, The AI ​​diagnosis and analysis unit (200) also includes a risk analysis module (230), which is used to perform subtype classification and disease progression risk assessment of dry eye disease; the risk analysis module (230) performs subtype classification and risk assessment by the following steps: S230.

1. Based on clinically recognized classification criteria for dry eye disease, the classification criteria for dry eye disease include aqueous hypoplasia. Excessive evaporation type Hybrid ; Combined with the feature weight values ​​output by the association model construction module (220) Calculate the matching score for each subtype. ,in Feature weight values absolute value and characteristics and fractals Related indicator variables The sum of the products; S230.2 Compare the matching scores of each subtype. The highest-scoring subtype is determined to be the patient's dry eye disease subtype; S230.3 Constructing a disease progression risk function using Kaplan-Meier analysis. Corneal staining score Duration of illness As the independent variable, according to a preset time interval calculate Probability of corneal epithelial damage progression at any given time ; S230.4, according to The values ​​are used to classify risk levels, forming a system that includes risk level and corresponding time points. The risk assessment results.

6. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 5, characterized in that, The personalized treatment plan generation unit (300) includes a treatment guideline structuring module (310), which is used to convert clinical treatment guidelines for dry eye disease into computable structured data, specifically including: Natural language processing technology was used to segment the guideline text, identify entities, and extract relationships to extract the correspondence between treatment methods and classifications, as well as the correlation threshold between treatment intensity and risk level. The extracted relationships are converted into a triplet data structure and stored in the medical knowledge graph. The subject is the subtype or risk level, the object is the treatment method or parameter, and the relationship is a logical association of "fit" or "correspondence". By comparing the structured data with more than 5 years of clinical diagnosis and treatment data, the matching degree between the structured data and actual diagnosis and treatment behavior is verified. When the matching degree reaches 90% or above, the calibration is completed, and a guideline knowledge base that can be directly called is formed.

7. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 6, characterized in that, The personalized solution generation unit (300) further includes a dynamic solution generation module (320), which is used to generate personalized solutions by combining multi-source data, specifically including: Contraindication screening: By comparing the allergy history records in the multi-source data integration unit (100) with the treatment components or operational characteristics in the guideline knowledge base through the field matching algorithm, contraindication options with a matching degree of 80% or higher are automatically removed; Parameter optimization: Based on the classification results output by the AI ​​diagnosis and treatment analysis unit (200), the baseline parameters of the corresponding classification in the guide knowledge base are called, and a parameter adjustment model is constructed by combining the patient's average daily eye use time and environmental humidity data. Solution output: Integrate and optimize treatment measures, implementation guidelines, and expected effect evaluation indicators to generate a structured solution document with a timestamp, and link it to the guideline knowledge base reference nodes and parameter adjustment basis.

8. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 7, characterized in that, The efficacy feedback unit (400) includes a follow-up data acquisition module (410) and an adaptive baseline difference analysis module (420), wherein: The follow-up data acquisition module (410) is used to connect to ophthalmic examination equipment at preset time points after treatment initiation to collect tear secretion volume. Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four categories of objective physiological indicators; each collected objective physiological indicator is associated with a unique patient ID and a collection timestamp. The data is transmitted to the encrypted database via the standard interface of the multi-source data integration unit (100). Storage, compatible with device interface and data storage specifications of the multi-source data integration unit (100); The adaptive baseline difference analysis module (420) is used to execute an adaptive data distribution baseline difference algorithm, specifically including: baseline data of objective physiological indicators collected at the beginning of treatment. For reference, the indicator data at each follow-up point of the same patient. Conduct normality assessment and within-group difference analysis; use test methods to assess the normality of the data and output the assessment results. According to the judgment result Select statistical methods If the data conforms to a normal distribution, then For paired t-tests, otherwise Wilcoxon signed-rank test; based on statistical methods Calculate follow-up data Compared with baseline data The within-group differences are calculated, and the statistical results of the differences are output. It generates a trend curve with time nodes on the horizontal axis and index values ​​on the vertical axis. .

9. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 8, characterized in that, The efficacy feedback unit (400) further includes a dynamic weight assessment module (430) and a cross-unit parameter feedback module (440), wherein: The dynamic weight evaluation module (430) is used to execute a dynamic weight algorithm driven by the indicator-efficacy correlation, specifically including: historical clinical case data stored by the multi-source data integration unit (100). The tear secretion volume collected by the follow-up data acquisition module (410) Tear film breakup time Corneal fluorescence staining score Concentration of inflammatory factors in tears Four types of objective physiological indicators were used to calculate the correlation coefficient between each objective physiological indicator and the degree of relief of dry eye disease using Pearson correlation analysis. ;by As a dynamic weight for each objective physiological indicator, combined with the improvement rate of each indicator. Calculate the comprehensive efficacy index ; The cross-unit parameter feedback module (440) is used to realize cross-unit parameter iterative feedback, specifically including: the comprehensive efficacy index and the statistical results of the differences output by the adaptive baseline difference analysis module (420) The data is transmitted to the AI ​​diagnosis and analysis unit (200), where the feature weights and regression coefficients of the risk function of the correlation model in the AI ​​diagnosis and analysis unit (200) are corrected using the gradient descent method; when two consecutive follow-up nodes... Below the dynamic threshold of the corresponding node At that time, a scheme adjustment trigger signal is sent to the personalized scheme generation unit (300), and the trigger signal includes the improvement rate of each indicator. and correlation coefficient This provides data to support the optimization of treatment parameters for personalized treatment plans.

10. The AI-assisted decision-making platform for personalized treatment of dry eye disease according to claim 9, characterized in that, The doctor-patient interaction unit (500) includes a doctor-nurse interaction module (510) and a patient-side interaction module (520), wherein: The medical staff interaction module (510) is based on the encrypted database of the patient's unique identifier ID and the multi-source data integration unit (100). It uses a structured interface and operation log retention technology to enable multi-source data association query, manual review of AI analysis results, and operation traceability; The patient-side interaction module (520) uses the structured treatment plan of the personalized treatment plan generation unit (300) and the hospital diagnosis and treatment system appointment interface, and uses timeline visualization, multi-terminal push and standardized collection technology to realize treatment information display, medication follow-up reminders and patient symptom data collection.

Images