Medical ai intelligent diagnosis and insurance claim risk hedging strategy optimization method

Abstract

This invention relates to the technical field of data optimization processing, and discloses a method, apparatus, equipment, and medium for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies. The method includes: acquiring input patient historical medical data and insurance data and integrating the data to form a basic analysis dataset; establishing patient-disease-insurance relationships based on graph neural networks, generating dynamic risk feature vectors, and constructing a multimodal medical data fusion and dynamic diagnosis model based on the dynamic risk feature vectors; based on the multimodal medical data fusion and dynamic diagnosis model, using AI intelligent diagnosis of real-time patient medical data, and mapping the patient disease probability and confidence level output by the AI ​​intelligent diagnosis results into insurance claim risk values. This application can be applied to the development of business systems in fintech, healthcare, and elderly care, improving the accuracy of medical diagnosis and risk assessment through deep coupling and a two-way feedback optimization mechanism between medical data and insurance risk control.

Description

Technical Field

[0001] This invention relates to the technical field of data optimization processing, and in particular to a method, apparatus, equipment, and medium for optimizing medical AI intelligent diagnosis and insurance claims risk hedging strategies. Background Technology

[0002] In the development of business systems for fintech, healthcare, and elderly care, most current medical AI systems only perform a specific function, such as image recognition, and do not incorporate insurance data such as user health records and claims history. Therefore, they lack a coherent risk prediction model for real-time updates and struggle to adapt to the patient's current condition and claims requirements. Furthermore, real-time updates lack precise data optimization. Current technologies are generally linear (diagnosis-claims), failing to create a positive cycle (improved diagnostic accuracy → reduced insurance risk) or a negative cycle (AI misdiagnosis → losses from claims cannot be used to optimize for a more accurate diagnostic model), thus wasting valuable resources. Summary of the Invention

[0003] This invention provides a method, apparatus, computer equipment, and medium for optimizing hedging strategies for medical AI intelligent diagnosis and insurance claims risks, in order to solve the technical problems in the prior art where there is no coherent risk prediction model for dynamic updates, and where there is a lack of accurate data optimization during real-time updates.

[0004] Firstly, this paper provides an optimization method for medical AI-based intelligent diagnosis and insurance claims risk hedging strategies, including: Acquire and integrate the input patient's historical medical and insurance data to form a basic analysis dataset; Based on graph neural networks, a patient-disease-insurance relationship is established, a dynamic risk feature vector is generated, and a multimodal medical data fusion and dynamic diagnostic model is constructed based on the dynamic risk feature vector; Based on multimodal medical data fusion and dynamic diagnostic model, this method uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy to construct a real-time risk scoring model driven by diagnostic results. Establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on a real-time risk scoring model; Based on the optimized scheme, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated, and the real-time risk scoring model is optimized.

[0005] Secondly, a device for optimizing medical AI-based intelligent diagnosis and insurance claims risk hedging strategies is provided, including: The integration module is used to acquire input patient historical medical data and insurance data and integrate the data to form a basic analysis dataset. The diagnostic module is used to establish patient-disease-insurance relationships based on graph neural networks, generate dynamic risk feature vectors, and construct a multimodal medical data fusion and dynamic diagnostic model based on the dynamic risk feature vectors. The scoring module is used to generate risk distribution curves based on multimodal medical data fusion and dynamic diagnostic models. It uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves and combines option pricing theory to derive the optimal hedging strategy. Based on the optimal hedging strategy, it adjusts the predicted premium or predicted sum insured in real time to build a real-time risk scoring model driven by diagnostic results. The feedback module is used to establish a two-way feedback optimization mechanism, generating positive or negative feedback optimization schemes based on the real-time risk scoring model. The optimization module is used to recalculate the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme, thereby optimizing the real-time risk scoring model.

[0006] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-mentioned method for optimizing the hedging strategy of medical AI intelligent diagnosis and insurance claims risk.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, it implements the steps of the above-mentioned method for optimizing the hedging strategy of medical AI intelligent diagnosis and insurance claims risk.

[0008] The aforementioned solution, implemented by the method, device, computer equipment, and storage medium for optimizing hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks, involves: acquiring and integrating historical medical and insurance data from patients to form a basic analysis dataset; establishing patient-disease-insurance relationships based on graph neural networks to generate dynamic risk feature vectors; constructing a multimodal medical data fusion and dynamic diagnostic model based on these dynamic risk feature vectors; using this model, mapping the patient's disease probability and confidence level output by AI-based intelligent diagnosis to insurance payout risk values ​​based on real-time patient medical data from AI-based intelligent diagnosis; generating risk distribution curves using Monte Carlo simulation; deriving the optimal hedging strategy using option pricing theory; adjusting the predicted premium or predicted sum insured in real-time based on the optimal hedging strategy to construct a real-time risk scoring model driven by diagnostic results; establishing a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on the real-time risk scoring model; and recalculating the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimized schemes to optimize the real-time risk scoring model. In this invention, the multimodal medical data fusion and dynamic diagnostic model continuously corrects the weight allocation of medical features through positive feedback and monetary difference optimization, making the prediction of disease probability and confidence level more in line with actual compensation patterns. It realizes the quantification of the probability of AI misdiagnosis into the price of insurance options to reduce the possibility of sudden large claims in insurance. By using closed-loop feedback, repetitive examinations and redundant claims processes are eliminated, which is expected to save 15%-20% of operating and management costs, improve the accuracy of medical diagnosis and risk assessment, reduce decision-making bias, and optimize insurance operation efficiency through a two-way feedback optimization mechanism, balancing risk and cost. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an application environment for an optimization method of medical AI intelligent diagnosis and insurance claim risk hedging strategy in one embodiment of the present invention.

[0011] Figure 2 This is a flowchart illustrating a method for optimizing the hedging strategy of medical AI intelligent diagnosis and insurance claim risk in one embodiment of the present invention.

[0012] Figure 3 This is a schematic diagram of a medical AI intelligent diagnosis and insurance claim risk hedging strategy optimization device according to an embodiment of the present invention.

[0013] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention.

[0014] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0015] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The medical AI intelligent diagnosis and insurance claim risk hedging strategy optimization method provided in this invention can be applied to, for example... Figure 1In this application environment, the client acquires and integrates historical medical and insurance data from patients to form a basic analysis dataset. Based on graph neural networks, a patient-disease-insurance relationship is established, generating dynamic risk feature vectors. A multimodal medical data fusion and dynamic diagnostic model is constructed based on these dynamic risk feature vectors. Based on this model, real-time patient medical data is analyzed using AI-powered intelligent diagnosis. The patient disease probability and confidence level output by the AI-powered intelligent diagnosis results are mapped into insurance payout risk values. A risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived using option pricing theory. Based on the optimal hedging strategy, the predicted premium or predicted sum insured is adjusted in real-time, constructing a diagnosis-driven real-time risk scoring model. A two-way feedback optimization mechanism is established to generate positive or negative feedback optimization schemes based on the real-time risk scoring model. Based on the optimization schemes, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated, thus optimizing the real-time risk scoring model. In this invention, the multimodal medical data fusion and dynamic diagnostic model continuously corrects the weight allocation of medical features through positive feedback and monetary difference optimization. This makes the predictions of disease probability and confidence level more consistent with actual compensation patterns, quantifying the probability of AI misdiagnosis into the price of insurance options to reduce the possibility of sudden large claims in insurance. Utilizing closed-loop feedback, repetitive examinations and redundant claims processes are eliminated, potentially saving 15%-20% in operational management costs, improving the accuracy of medical diagnosis and risk assessment, reducing decision-making bias, and optimizing insurance operational efficiency through a two-way feedback optimization mechanism, balancing risk and cost. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0017] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the method for optimizing medical AI-based intelligent diagnosis and insurance claims risk hedging strategies provided in this embodiment of the invention includes the following steps: S10: Obtain the input patient's historical medical data and insurance data, and integrate the data to form a basic analysis dataset; Patient historical medical data typically encompasses multiple dimensions, such as the patient's basic identification information (e.g., name, gender, age, ID number, etc., used to uniquely identify an individual), past medical history (including names of previous illnesses, diagnosis dates, treatment methods, etc.), medical records (e.g., number of outpatient / inpatient visits, departments visited, prescribed medications, examination and test results such as blood routine tests, imaging reports, etc.), and health management-related data (e.g., physical examination reports, vaccination records, etc.). This data may be scattered across different medical information systems, such as the hospital's electronic medical record system, laboratory information management system, and physical examination center database.

[0018] Insurance data primarily revolves around a patient's insurance participation, including basic information about the insurance product (such as type of insurance, application date, coverage period, sum insured, and premium payment records), claims records (such as claim application date, reason for claim, payout amount, claim approval status, and review comments), and relevant information provided by the policyholder during the application process (such as health declaration details, occupational information, etc., some of which may overlap with medical data). This type of data is typically stored in the insurance company's business systems, claims management systems, and other databases.

[0019] After data collection, data cleaning and preprocessing are necessary, primarily to eliminate errors, redundancy, and inconsistencies. For medical data, potential problems include: inconsistent terminology for the same disease across different hospitals (e.g., "myocardial infarction" vs. "heart attack"), inconsistent units for test results (e.g., blood pressure measured in mmHg and kPa), and missing values ​​in some records (e.g., a medical record omitting the consultation time). Handling methods include standardizing disease terminology (referencing the International Classification of Diseases (ICD) coding standards), converting data units to a unified standard, and filling in missing values ​​using appropriate methods (e.g., estimating based on average data from patients of the same age and with the same disease, or deleting severely missing and meaningless records).

[0020] Common issues with insurance data may include: inconsistent date formats in premium payment records (e.g., some using "year / month / day," others "month / day / year"), vague descriptions of claim reasons (e.g., only stating "illness" without specific details), and duplicate data (e.g., the same claim being entered multiple times). Processing requires standardizing date formats, supplementing and improving key information such as claim reasons, and removing duplicate records to ensure the accuracy and uniqueness of each insurance data entry.

[0021] The next step is data association and integration. The core of this step is to establish the correspondence between medical data and insurance data, linking the two types of data belonging to the same patient. Typically, the patient's unique identifier is used as the basis for association, such as their ID card number (which must be anonymized to protect privacy). If other cross-verifiable identifiers exist (such as medical insurance card numbers, policy numbers, and patient identity binding information), they can also be used as supplementary association criteria to avoid association biases caused by missing or incorrect identifiers.

[0022] Based on the association, data fields are integrated and structured. For example, basic patient information such as "age" and "gender" are correlated with insurance information such as "insurance type" and "coverage amount"; "diabetes in past medical history" is linked with "claims records related to diabetes"; and medical data such as "annual outpatient visits" and "total hospitalization costs" are integrated with insurance data such as "annual premium payment amount" and "total claim amount" into the same data record. Simultaneously, the integrated data fields are standardized with clear naming conventions, defining the definition and data type of each field (e.g., numeric, character, date), ensuring a clear data structure and facilitating subsequent reading and processing by analytical tools.

[0023] Finally, a basic analysis dataset is formed. After the above steps, the resulting dataset contains the patient's key medical and insurance information, comprehensively reflecting the patient's health status, insurance participation, and claims situation. Throughout the process, data privacy protection is emphasized, and de-identification, encryption, and other methods are used to ensure that the patient's personal information and sensitive data are not leaked, complying with relevant laws and regulations (such as the Medical Data Privacy Protection Ordinance and the Insurance Data Security Specification).

[0024] S20: Establish patient-disease-insurance relationship based on graph neural network, generate dynamic risk feature vector, and construct multimodal medical data fusion and dynamic diagnosis model based on dynamic risk feature vector; The core advantage of graph neural networks is their ability to model "entity-relationship" structures—they can transform discrete entities and their relationships into computable graph structures, and mine implicit relationships between entities by aggregating neighbor information. First, entities and relationships are defined, and then the association features of entities are learned through GNN.

[0025] First, the definition of entities and relationships refers to the nodes in the graph, which include three core types of objects: Patient entities: covering the patient's basic attributes (age, gender, history of underlying diseases), dynamic health status (such as current symptoms, recent test results), etc.; Disease entities: covering the medical attributes of the disease (such as ICD code, pathological features, risk of complications), treatment plans (such as medication type, surgical method), etc.; Insurance entities: covering insurance product attributes (coverage scope, deductible, reimbursement ratio), patient's insurance status (insurance duration, claims history, renewal status), etc. Relationships refer to the edges that connect entities, used to express interactions or dependencies between entities: Patient-Disease Relationship: such as "a patient has had a certain disease" (including the time of illness and treatment outcome), "a patient is currently suspected of having a certain disease" (including symptom matching degree), "a patient incurs medical expenses due to a certain disease" (including the amount of expenses and payment method), etc.; Patient-Insurance Relationship: such as "a patient has taken out an insurance policy" (including the insurance period and coverage amount), "a patient applies for insurance claims due to a certain disease" (including whether the claim is approved and the amount of compensation), "the insurance policy's coverage restrictions for a certain disease" (such as the exclusion period and the maximum compensation limit), etc.; Disease-Insurance Relationship: such as "a certain disease is covered by insurance" (including compensation conditions), "the treatment cost of a certain disease is linked to the insurance premium" (such as high-risk diseases correspond to high premiums), etc.

[0026] The construction of graph structures and GNNs mainly involves transforming the aforementioned entities and relationships into a graph structure (nodes are entities, edges are relationships, and edge weights can be set according to the strength of the association, such as "a patient was recently diagnosed with a certain disease" having a higher weight than "a patient had a certain disease five years ago"). GNNs are then used for feature learning: GNNs use a "message passing mechanism" to allow each node (entity) to aggregate information from its neighboring nodes. For example, a patient node aggregates features from its associated disease nodes (such as disease type and severity) and insurance nodes (such as claims records and coverage). Simultaneously, disease nodes aggregate health data (such as treatment response) and insurance payout rules from associated patients. Ultimately, each node's features not only include its own attributes but also incorporate key information from associated entities. This aggregation process can capture implicit associations: for example, through the patient-disease-insurance triangular relationship, GNNs can identify deep-seated patterns such as "when a certain type of patient (such as someone with a history of hypertension) suffers from a certain disease (such as stroke), the probability of payout for a certain insurance product (such as critical illness insurance) is significantly higher than other combinations." Traditional tabular data or linear models struggle to capture such cross-entity associations.

[0027] Dynamic risk feature vectors refer to feature sets that can be updated in real time with changes in time and entity states. They include both static attributes of entities and temporal change information, accurately reflecting the dynamic changes in a patient's health risk and insurance payout risk. Generation primarily relies on the dynamic updates of graph structures and the temporal learning capabilities of GNNs. The dynamism stems mainly from changes in two types of information: updates to entity states (e.g., new patient test results, such as elevated blood sugar, disease progression, such as tumor stage upgrades), and changes in insurance status (e.g., failed renewal, new claim records); and temporal changes in relationships (e.g., the relationship between the patient and the disease changes from "suspected" to "confirmed" over time, and the relationship between the patient and insurance changes from "normal insurance application" to "claim pending").

[0028] In the process of extracting dynamic features, firstly, the graph structure needs to be dynamically adjusted as the entity state is updated: for example, when a patient develops new symptoms, a new edge for "patient-suspected disease" needs to be added; when the insurance claim result is confirmed, the weight of the "patient-insurance" edge needs to be updated (if the claim is approved, the weight increases, reflecting a stronger association). Secondly, temporal graph neural networks (such as TGCN, Gated GNN) are used to process the dynamic graph: these models add a time dimension to the traditional GNN, memorize historical information through gating mechanisms (such as GRU / LSTM), and aggregate the neighbor features of the current graph structure. For example, for the patient node, the model combines its blood pressure data from 3 months ago, its electrocardiogram results from 1 week ago (historical information), and its current dizziness symptoms (new information) to generate features that include "recent cardiovascular risk change trends"; for the insurance node, it combines the patient's historical claim frequency and the severity of the current disease to generate a "current dynamic value of claim risk". The resulting dynamic risk feature vector includes, for patients, "current health risk score, disease progression rate, and matching degree with insurance coverage"; for diseases, "incidence rate changes in this patient group and the strength of correlation with insurance payouts"; and for insurance, "real-time payout probability for this patient and risk fluctuations affected by the patient's health status".

[0029] Multimodal medical data refers to medical information from various sources or in different formats (such as text medical records, image data, and test results). Dynamic diagnostic models mainly combine these multi-source data with the aforementioned dynamic risk features to achieve real-time and accurate diagnostic decisions. Multimodal data is characterized by its heterogeneous format. Features are extracted separately for each data type: Text data, such as symptom descriptions and doctor's diagnoses in medical records, is converted into semantic vectors using a pre-trained language model (such as BERT) to capture implicit information such as the semantic association between the patient's complaints of "chest tightness" and "angina"; Image data, such as CT and MRI images, is extracted using a convolutional neural network (CNN) to extract lesion features (such as tumor size, location, and density) and converted into visual feature vectors; Numerical data, such as blood routine indicators (white blood cell count) and vital signs (heart rate, blood pressure), is converted into numerical features through standardization (such as Z-score) to retain their quantitative relationships; Time-series data, such as patients' blood glucose monitoring curves and body temperature trends, is extracted using a time-series model (such as LSTM) to extract trend features (such as "sustained rise in blood glucose" and "sudden drop in body temperature").

[0030] The core of the fusion is to allow multi-source data and dynamic risk features generated by graph neural networks to "complement and corroborate" each other, avoiding information bias from a single modality. This is achieved through the following methods: the model automatically learns the weights of different modalities of data. For example, when an image shows "lung nodules," the weight of image features increases; when a medical record mentions "long-term smoking history," the weight of text features increases; and dynamic risk features (such as "the patient's insurance record shows a previous claim for lung disease") act as a "bridge," strengthening the connection between images and text (e.g., "smoking history + nodule image + previous claims" all point to "lung cancer risk"). After concatenating the modal features with the dynamic risk features, deep interaction is performed through fully connected layers or a Transformer encoder to generate fused features. For example, combining "image lesion features," "text symptom features," and "dynamic association features between patient, disease, and insurance" outputs a "comprehensive disease probability distribution" (e.g., lung cancer 70%, pneumonia 20%, other 10%).

[0031] The dynamic diagnostic model is not a static output; instead, it adjusts the diagnostic results in real time based on new data. The core output of the model is the "dynamic diagnostic probability": combining the current fusion features, it predicts the probability of a patient having various diseases, with dynamic confidence (e.g., "Based on the current data, the probability of lung cancer is 70%; if a new biopsy result is added, the confidence can be increased to 90%"). The real-time update mechanism is as follows: when a patient generates new medical data (e.g., new examination reports, follow-up visit records), the model re-extracts the features of that modality, updates the patient-disease relationship in the graph structure (e.g., "biopsy confirms lung cancer" strengthens the edge weights of the patient-lung cancer relationship), and then regenerates dynamic risk features through the GNN, ultimately updating the diagnostic probability. For example, if "genetic testing shows an EGFR mutation," the model will combine "the strong association between this mutation and lung cancer" and "insurance reimbursement rules for targeted drugs," increasing the lung cancer diagnosis probability from 70% to 95%, while simultaneously indicating "meets the conditions for targeted therapy insurance reimbursement."

[0032] This step breaks through the limitations of traditional "isolated data" by using graph neural networks, capturing real-time changes with dynamic features, and then integrating comprehensive information through multimodal fusion. Ultimately, it achieves "dynamic diagnosis and risk assessment based on the patient's full life cycle health data, disease medical characteristics, and insurance constraints," which not only assists doctors in making accurate decisions but also provides data support for insurance risk control.

[0033] After step S20, the method further includes: inputting insurance claim risk weights into the multimodal medical data fusion and dynamic diagnosis model for adversarial training, and optimizing the multimodal medical data fusion and dynamic diagnosis model based on the training results.

[0034] Insurance payout risk weights are core indicators for quantifying the impact of different medical factors on insurance payout risk. Their numerical values ​​reflect the importance of a specific factor in risk assessment. For example, high treatment costs for certain chronic diseases, morbidity rates in specific age groups, and high-frequency diseases in past claims records are all assigned corresponding weights. Higher weights indicate a greater impact of the factor on insurance payout risk. These weights are input as "constraint signals" into a multimodal medical data fusion and dynamic diagnostic model. At this point, the model's learning objective is not only to accurately diagnose diseases (such as identifying disease types and predicting disease progression), but also to incorporate these risk weights into the decision-making process to ensure that the diagnostic results align with the assessment logic of insurance payout risk (for example, for diseases corresponding to high-risk weights, the model needs to more carefully identify early symptoms to avoid underestimating subsequent payout risks due to missed diagnoses). Adversarial training forces models to improve robustness in complex scenarios through "contradictory confrontation." Specifically, it involves building a dynamic balance between "diagnostic tasks" and "risk adaptation" during training. On one hand, the model needs to focus on the multimodal fusion of medical data and diagnostic logic. For example, it deeply integrates different modalities of information, such as medical record text (describing symptoms), imaging data (e.g., lesions in CT scans), and laboratory indicators (e.g., blood routine tests, biochemical data), to extract key medical features (e.g., lesion size, symptom duration, and degree of abnormality), thereby outputting accurate diagnostic results (e.g., disease probability, severity). On the other hand, it introduces "adversarial challenges": by setting an implicit "adversarial objective" (or simulating dynamic fluctuations in risk weights), it deliberately creates conflict between "diagnostic accuracy" and "risk adaptation" in the model. For example, when the risk weight for insurance payout in a certain type of case suddenly increases, the adversarial mechanism will "interfere" with the model's regular diagnostic logic, examining whether it will overlook key medical characteristics due to an overemphasis on risk weights (e.g., misdiagnosing low-risk cases with similar symptoms in an attempt to reduce the misdiagnosis rate of high-risk cases); or whether it will ignore the constraints of risk weights due to an emphasis on medical diagnosis (e.g., missing early-stage diseases corresponding to high-risk weights, leading to ineffective risk assessment). Through the continuous confrontation between "diagnostic ability" and "risk adaptation ability," the model will gradually learn to find a balance between the two objectives: neither deviating from the essence of medicine (ensuring diagnostic accuracy) nor failing to respond sensitively to changes in risk weights (ensuring the reference value of diagnostic results for risk assessment).

[0035] Adversarial training exposes the model's weaknesses in fusing multimodal data and responding to risk weights (such as underestimating the correlation between a certain type of data and risk weights, or lagging response to changes in risk weights during dynamic diagnosis). These weaknesses become the core basis for model optimization: if the model ignores the diagnostic correlation between a certain type of data (such as microlesion features in images) and high-risk weight cases, the weight allocation of the modality fusion layer will be adjusted to enhance the influence of that modality feature in the fusion process, ensuring that key medical signals are not obscured by the "interference" of risk weights. If the model is slow to respond to dynamic changes in risk weights (such as the model still using the old logic for diagnosis after risk weight adjustment), the real-time learning capability of the dynamic diagnosis module will be optimized. For example, the adaptation period of the model to new risk weights may be shortened, or a feature re-extraction step may be added when risk weights change, allowing the model to quickly focus on medical features related to new risks.

[0036] By incorporating an "adversarial loss term" related to risk weights into the model's loss function, this term amplifies the error when the model exhibits diagnostic bias in high-risk cases, forcing the model to focus on learning the characteristics of such cases in subsequent training. Conversely, if the model maintains diagnostic stability amidst fluctuations in risk weights, its corresponding capabilities are strengthened through the reward mechanism of the loss term. Through these optimizations, the model ultimately meets the dual requirements of "accurate diagnosis" and "risk adaptation," possessing both powerful medical data fusion and dynamic diagnostic capabilities, and adaptability to real-world scenarios of insurance payout risk assessment.

[0037] S30: Based on multimodal medical data fusion and dynamic diagnostic model, it uses AI intelligent diagnosis to diagnose real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy to build a real-time risk scoring model driven by diagnostic results. The multimodal medical data fusion and dynamic diagnostic model continuously receives and integrates patients' real-time medical data, which covers multi-dimensional information: including both structured data (such as real-time monitored physiological indicators such as heart rate and blood sugar, and the latest blood routine and biochemical test values) and unstructured data (such as symptom descriptions recorded by doctors in real time, newly generated image reports (such as CT and ultrasound results), and voice-to-text information from real-time consultations).

[0038] The model dynamically integrates these multimodal data to update its assessment of the patient's health status in real time. For example, when a patient experiences new real-time data such as "persistent chest pain + a sudden increase in troponin levels," the model combines this with information on "coronary artery disease history" from the patient's medical records to quickly adjust the diagnostic direction, outputting the probability of "acute myocardial infarction" (e.g., 85%) and the confidence level of this diagnosis (e.g., 92%, reflecting the model's reliability of the result). This dynamic diagnosis is not a one-time result but continuously iterates with the input of new data (such as subsequent electrocardiogram results and interventional surgery records), ensuring that the assessment of the patient's current health status remains timely and accurate.

[0039] The "disease probability" and "confidence level" output by AI-powered diagnostic results need to be converted into "payout risk values" that can be directly used in the insurance field. This mapping process combines the logic of medical characteristics and insurance terms: the conversion of disease probability links "the likelihood of a patient having a certain disease" with "the corresponding insurance payout cost." For example, if AI diagnoses that "the patient has a 70% probability of having diabetes," then it needs to consider the average treatment period for diabetes, annual medical expenses, and the insurance product's reimbursement ratio for diabetes (e.g., 50% for outpatient reimbursement and 80% for inpatient reimbursement) to calculate "the expected annual payout amount that the insurance company might generate if the patient is diagnosed" (e.g., 70% × 12,000 yuan per year = 8,400 yuan). The confidence level reflects the reliability of the diagnostic results and will adjust the risk value accordingly. If the confidence level of a 70% probability of disease is only 60% (meaning the model has low confidence in the outcome), it indicates a 30% probability of misdiagnosis or misjudgment of the condition. In this case, the risk value needs to be lowered (e.g., 0.84 million yuan × 60% = 0.504 million yuan). If the confidence level reaches 95%, the risk value is closer to the actual expected payout (e.g., 0.84 million yuan × 95% = 0.798 million yuan). The final "insurance payout risk value" is a comprehensive indicator that reflects both the potential impact of the patient's current health status on insurance payouts and the degree of uncertainty surrounding that impact.

[0040] Insurance payout risk is not a fixed value, but is affected by a variety of uncertain factors (such as whether the patient develops complications, whether treatment costs exceed expectations, and whether the policy is surrendered early). Monte Carlo simulation presents a comprehensive picture of risk through "random scenario generation + extensive repeated calculations": During the simulation, a series of variables affecting the risk and their fluctuation ranges are set based on the generated basic payout risk value. For example, whether the patient develops complications (incidence rate 10%-40%), the length of the treatment period (3 months-2 years), the regional medical cost inflation rate (3%-8%), and further fluctuations in drug prices (±10%). Through random sampling, tens of thousands or even hundreds of thousands of variable combination scenarios are generated, such as "complication incidence rate 30% + treatment period 1 year + medical inflation 5% + drug price increase 8%" and "no complications + treatment period 6 months + medical inflation 3% + drug price decrease 5%". After calculating the corresponding payout risk value for each scenario, the probability distribution of all results is statistically analyzed to form a risk distribution curve. The horizontal axis of the curve represents the possible payout risk value (e.g., 5,000-50,000 yuan), and the vertical axis represents the probability of that risk value occurring. It can intuitively display the "most likely payout range" (e.g., 10,000-20,000 yuan, with a probability of 60%), the "medium risk range" (e.g., 20,000-30,000 yuan, with a probability of 25%), and the "extremely high risk range" (e.g., 30,000-50,000 yuan, with a probability of 5%), allowing insurance companies to clearly grasp the overall fluctuation characteristics of risk and avoid decision-making biases caused by relying solely on a single risk value.

[0041] Insurance is essentially a hedging tool for "future payout risk," and the core idea of ​​option pricing theory (such as the Black-Scholes model), "offsetting risk by constructing hedging portfolios," can be applied to insurance risk management: "Potential future payouts for patients" can be likened to "the exercise risk of an option," with the insurance company essentially "selling a put option." When a patient experiences a medical event covered by the insurance (equivalent to "the exercise condition being triggered"), the insurance company must pay out a payout (equivalent to "the exercise amount"). Derivation of optimal hedging strategies: For example, for a high-probability, medium-risk range (10,000-20,000 yuan), coverage can be achieved by reserving a special risk reserve; for a low-probability, extremely high-risk range (30,000-50,000 yuan), pressure can be dispersed by purchasing reinsurance (transferring some risk to other insurance companies); for highly volatile variables (such as drug prices), financial derivatives linked to drug prices can be introduced for hedging.

[0042] Based on the risk distribution curve, the expected loss under different payout scenarios is calculated. Then, combining this with the "delta hedging" logic in option pricing, a hedging strategy is designed. For example, if the probability of a high-risk scenario (payout of 15,000 yuan) is 5%, a corresponding risk reserve (e.g., 15,000 yuan × 5% = 750 yuan) needs to be reserved. If medium-risk scenarios (8,000 yuan) account for 60%, the ratio of premium income to investment returns is adjusted to ensure sufficient funds cover this portion of the expected payout. The core of the optimal hedging strategy is to find a balance between "hedging costs" and "risk coverage," avoiding both excessive hedging increasing operating costs and insufficient hedging exposing solvency risks. The optimal hedging strategy directly guides real-time adjustments to premiums or coverage amounts, ensuring that insurance product pricing matches risk levels. The specific adjustment logic varies depending on the direction of risk changes: When the risk value increases (e.g., AI diagnosis shows a patient's disease probability rises from 70% to 85%, shifting the overall risk distribution curve to the right), revenue is increased by raising the predicted premium (e.g., from 2,000 yuan to 2,500 yuan per year), or the predicted coverage amount is reduced (e.g., from 500,000 yuan to 400,000 yuan) to reduce the potential payout limit and offset the increased risk; when the risk value decreases (e.g., subsequent diagnosis excludes critical illnesses, shifting the risk distribution curve to the left), premiums are reduced (e.g., from 2,000 yuan to 1,800 yuan) to enhance product attractiveness, or the coverage amount is increased (e.g., from 500,000 yuan to 600,000 yuan) to improve user experience while maintaining reasonable hedging costs; because AI diagnostic results change dynamically with patient data updates (e.g., receiving new monitoring data every 6 hours), premium or coverage amount adjustments are also triggered synchronously, avoiding losses for insurance companies or damage to user rights due to the lag in risk and pricing.

[0043] Ultimately, information from all the above stages is integrated into a "real-time risk scoring model." Using AI-powered diagnostic results as the core input, it outputs a comprehensive score (e.g., 0-100 points) that intuitively reflects the patient's current overall risk level relative to the insurance system. The scoring dimensions cover disease risk (probability of disease based on AI diagnosis), payout uncertainty (fluctuation range based on confidence level and risk distribution curve), hedging effectiveness (risk coverage level based on optimal strategy), and adjustment response speed (timeliness of premium / coverage adjustments). In the dynamic update mechanism, whenever the patient generates new medical data (e.g., follow-up visit records, examination results), the AI ​​diagnostic results are updated, and the risk value, distribution curve, hedging strategy, and premium / coverage are adjusted accordingly. The real-time risk score is also updated synchronously (e.g., from 65 to 78 points, indicating increased risk). This score can be directly used for insurance underwriting (e.g., high-scoring users require additional health disclosures), claims review (e.g., simplified processes for low-scoring users), and product recommendation (e.g., matching flexible adjustment insurance to users with medium scores), realizing a shift from "post-event payout" to "real-time risk prediction and management." In summary, this step, through the deep integration of medical diagnosis and insurance risk control, closely links the pricing and coverage of insurance products with the patient's real-time health status. This not only improves the accuracy of risk management but also makes insurance services more aligned with users' actual medical needs, ultimately achieving a closed-loop management model of "diagnosis driving risk and risk guiding decision-making."

[0044] Step S30 includes: based on multimodal medical data fusion and dynamic diagnostic model, real-time patient medical data is diagnosed using AI intelligent diagnosis; real-time patient disease probability and confidence level are obtained based on AI intelligent diagnosis, and the patient disease probability, confidence level, insurance clause payout coefficient, and dynamic adjustment factor are mapped and converted into insurance payout risk value, wherein the dynamic adjustment factor includes at least regional incidence rate and drug price fluctuation; a risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived by combining option pricing theory; the predicted premium or predicted sum insured is adjusted in real time based on the optimal hedging strategy, and a real-time risk scoring model driven by diagnostic results is constructed.

[0045] It should be noted that the disease probability and confidence level output by AI need to be combined with multiple factors to be converted into a claim risk value in the insurance field, where the claim risk value R = , Let i be the probability of disease i. To correspond to the reimbursement coefficient in the insurance policy, These are dynamic regulating factors (such as regional incidence rates and drug price fluctuations).

[0046] This process needs to balance medical characteristics with the actual operational logic of insurance: First, it uses disease probability and confidence level as the basic framework. Disease probability directly relates to the likelihood of potential compensation. For example, "the probability of a patient having chronic kidney disease is 60%" means there is a 60% chance of triggering the corresponding insurance liability. Confidence level is used to adjust the reliability of this probability. If the confidence level is 80%, it means the model has a high degree of confidence in judging the "60% probability," and the risk value calculation will rely more on this data. If the confidence level is only 50%, its weight needs to be weakened to avoid misjudgment of risk due to diagnostic uncertainty. Second, it introduces the insurance clause's compensation coefficient. This coefficient is determined by the specific terms of the insurance product. For example, a critical illness insurance policy might have a compensation coefficient of 100% (full compensation of the insured amount) for "end-stage chronic kidney disease" and a compensation coefficient of 30% (proportional compensation) for "early-stage chronic kidney disease." The level of the coefficient directly affects the magnitude of the risk value, ensuring that the risk assessment and the scope of insurance liability are strictly matched. Finally, regional incidence rates and drug price fluctuations are incorporated as dynamic adjustment factors. Regional incidence rates reflect the prevalence of a specific disease in the patient's region. For example, if the incidence rate of diabetes in a certain region is 20% higher than the national average, the risk value for patients in that region to develop diabetes needs to be increased by 20% to adapt to localized disease risk characteristics. Drug price fluctuations directly affect treatment costs. If the price of a targeted drug that a patient may use has recently increased by 15%, the corresponding payout risk value needs to be adjusted upwards to reflect changes in actual medical costs. Through these adjustments, the final insurance payout risk value can reflect both the individual patient's health status and adapt to the dynamic changes in the macro environment and the healthcare market.

[0047] Information from all the above stages is used to construct a diagnostic result-driven real-time risk scoring model. This model outputs a comprehensive score (e.g., 0-100 points) covering dimensions such as disease risk (based on AI diagnostic probability), payout uncertainty (based on risk distribution curves), hedging effectiveness (based on optimal strategies), and regional and market factors (based on dynamic adjustment factors). The score is updated synchronously with the patient's real-time medical data. For example, if a patient adds a record of "blood sugar control achieved," the score may drop from 70 to 60 (risk reduction), directly reflecting the current risk level. This score can be directly applied to insurance underwriting, claims review, and product recommendation scenarios, realizing a shift from "passive payout" to "proactive, real-time risk management," allowing insurance services to more accurately match the patient's health status and actual needs.

[0048] S40: Establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on a real-time risk scoring model; A two-way feedback optimization mechanism is established, primarily by using the output of the real-time risk scoring model to influence the front-end medical data processing and diagnostic models, the mid-end risk transformation logic, and the back-end insurance strategies, forming a closed loop of "scoring feedback - solution optimization - effect verification." Its core is to generate targeted positive or negative feedback solutions based on the direction (increase or decrease) and magnitude of fluctuations in the real-time risk score, driving the entire system to dynamically adapt to actual risk changes.

[0049] First, the comprehensive score (e.g., 0-100 points) output by the real-time risk scoring model serves as the "signal center" of the feedback mechanism. Its changes contain two key pieces of information: the score trend (increasing / decreasing / stabilizing) and the degree of abnormality (whether the fluctuation exceeds a preset threshold). The real-time risk scoring mechanism uses these two types of information to determine "whether the current system needs adjustment" and "the direction of adjustment": When the score increases significantly (e.g., from 60 to 85 points, exceeding the "normal fluctuation threshold" by 15 points), it indicates that the patient's risk is rising faster than expected. This may be due to the AI ​​diagnostic model's failure to capture new medical data in a timely manner, the risk value conversion logic underestimating dynamic adjustment factors (e.g., a sudden increase in regional incidence), or the hedging strategy failing to cover new risks. In this case, negative feedback needs to be triggered to suppress further risk expansion by optimizing the model or strategy; when the score decreases significantly (e.g., from 70 points to...), the risk assessment mechanism should be adjusted accordingly. A score of 45 (with a fluctuation range of 25 points) indicates a risk decrease exceeding expectations. This could be due to overly conservative AI diagnosis (e.g., insufficient weighting of medical data related to "improved condition") or overly stringent premium / coverage adjustments. In this case, positive feedback should be triggered to optimize and release system flexibility, avoiding a decline in user experience or wasted costs due to over-prevention. When the score fluctuates within a normal threshold (e.g., 60 ± 5 points), the current system is considered well-suited, and no feedback scheme is triggered. Instead, data changes are continuously monitored to ensure that feedback actions match the actual needs of risk changes, avoiding frequent adjustments that could lead to system instability. The main purpose of positive feedback is to "release redundant prevention and control resources and improve system efficiency," and it is suitable for scenarios where the score has significantly decreased and the risk decline is indeed benign (excluding "false score reductions," such as falsely low risk due to data collection errors). The positive feedback scheme optimizes three aspects simultaneously: medical diagnosis, risk transformation, and hedging strategies. For the positive optimization of the AI ​​intelligent diagnostic model, if the score reduction stems from a substantial improvement in the patient's health (e.g., chronic disease indicators reaching target levels, complication regression), but the AI ​​model doesn't adequately weight "positive data" (e.g., the feature "sustained normal blood sugar" in the latest physical examination report isn't fully captured), the scheme adjusts the model's feature extraction logic: increasing the weight of "positive medical data" (e.g., increasing the weight of "blood pressure reaching target levels for 3 consecutive months" from 0.2 to 0.4), allowing diagnostic results to more promptly reflect the improvement in the condition; and relaxing the exclusion threshold for "suspected diseases" (e.g., increasing the reduction in the "probability of disease" for medical texts with "symptom relief" (e.g., "headache disappeared") from 10% to 20%), avoiding excessive retention of "high-risk predictions." For the positive optimization of the risk value transformation logic... If the score reduction is related to positive changes in dynamic adjustment factors (such as a decline in regional incidence rates or a decrease in drug prices), but the adjustment factors in the original conversion logic are "insufficiently sensitive" (such as a 10% reduction in drug prices only lowering the risk value by 3%), the positive optimization scheme will modify the conversion formula: increase the weight coefficient of the dynamic adjustment factors (such as adjusting the influence coefficient of regional incidence rates from 0.15 to 0.25) to make the risk value more sensitive to the improvement of the macro environment; reduce the "conservative correction term" (such as adjusting the confidence reduction coefficient from 0.8 to 0.9) to reduce the overestimation of the risk value due to excessive doubt about the reliability of the diagnosis. For positive adjustments to hedging strategies and premiums / coverage, if the original hedging strategy (such as high risk reserves) or premiums / coverage (such as excessively high premiums) appears redundant after the score decreases, the plan will increase control efforts: lowering the special risk reserve ratio (e.g., from 120% to 100% of expected payouts) to improve fund utilization efficiency; proactively reducing premiums (e.g., from RMB 1800 / year to RMB 1500 / year) or increasing the coverage amount (e.g., from RMB 300,000 to RMB 400,000), and simultaneously simplifying the underwriting process (e.g., reducing repeated health declarations for "low-risk users") to improve user experience. Positive feedback mainly reduces unnecessary control costs under the premise of controllable risk, but it needs to be verified through "secondary verification" to ensure the stability of the optimized score (e.g., the score remains at a low level for three consecutive rounds of monitoring) to avoid over-optimization due to misjudgment caused by short-term fluctuations. Negative feedback aims to "strengthen risk capture and control capabilities and curb risk spread." It is applicable to scenarios where scores rise significantly and the risk is indeed substantially increased (excluding "false increases in scores," such as false high risks caused by temporary data anomalies). The negative feedback scheme focuses on "filling gaps": for negative optimization of AI dynamic diagnostic models, if the score rises due to the AI ​​model missing "risk signals" (such as failing to promptly identify "ST segment elevation on ECG," a precursor to myocardial infarction), the negative feedback scheme will strengthen the model's sensitivity to high-risk features: increasing the monitoring frequency of "high-risk medical data" (such as shortening the data collection interval for "tumor markers" from 7 days to 3 days) to ensure the model obtains key information in a timely manner; and increasing the association weight of "suspected high-risk symptoms" (such as increasing the weight of the combined feature of "chest pain + dyspnea" from 0.3 to 0.5) to avoid diagnostic delays caused by feature dispersion. For negative optimization of the risk value conversion logic, if the increase in score is related to the deterioration of dynamic adjustment factors (such as the outbreak of regional infectious diseases or the inclusion of expensive and effective drugs in the treatment plan), but the original conversion logic underestimates its impact (such as a 20% increase in drug prices only raising the risk value by 5%), the solution will revise the conversion rules: introduce an "emergency adjustment coefficient" (such as adding a 1.5 times coefficient to the incidence rate in areas with public health emergencies) to forcibly increase the risk value's response to extreme events; add a "confidence level safety net clause" (such as when the disease probability is >70%, even if the confidence level is 80%, the risk value will no longer be reduced, avoiding underestimation of high-probability risks due to "over-reliance on confidence level"). For negative adjustments to hedging strategies and premiums / coverage, if an increase in the score indicates insufficient coverage by the original hedging strategy (such as a shortfall in claims reserves during extreme risk periods), the solution will strengthen prevention and control: supplement extreme risk hedging tools (such as purchasing "excess payout reinsurance" for high-risk users) to cover risk scenarios not covered by the original strategy; adjust premiums / coverage in stages (such as a 20% increase in premiums for users with scores above 80, while reducing the coverage by 10%), and add "risk disclosure" (such as sending users a notification that "current health status needs to be monitored more closely"), using price leverage to guide users to pay attention to risk control. The core of negative feedback is "precisely plugging risk loopholes," but a "one-size-fits-all" approach should be avoided. For example, for "an increase in the score due to short-term data fluctuations (such as occasional increases in blood pressure), only temporary strengthening of monitoring should be implemented instead of immediately adjusting the long-term strategy to prevent excessive prevention and control from affecting the user experience. The two-way feedback scheme includes a closed loop of "implementation-monitoring-re-optimization". After the two-way feedback scheme is implemented, the real-time risk scoring model will continuously track the score changes (such as daily updates) to determine whether the optimization measures are effective (such as whether the score has fallen back to a reasonable range after the implementation of the negative scheme). If the score does not meet expectations (such as the score continuing to decrease after the implementation of the positive scheme, which may be due to over-optimization), "secondary feedback" will be initiated to fine-tune the scheme parameters (such as reverting some feature weights). In the long run, all feedback cases will be included in the "feedback strategy library". By analyzing the scheme effects in different scenarios, the mechanism parameters such as "score fluctuation threshold" and "optimization weight of each link" will be continuously optimized to improve the accuracy of feedback. Specifically, step S40 includes: when it is positive feedback, obtaining the real-time compensation amount, comparing the difference between the real-time compensation amount and the predicted premium, and optimizing the multimodal medical data fusion and dynamic diagnosis model based on the difference; when it is negative feedback, when the accuracy of AI intelligent diagnosis is improved to a preset optimization threshold, triggering the optimization of the hedging ratio of the optimal hedging strategy in the real-time risk scoring model.

[0050] When the feedback is positive, the actual real-time claim amount is obtained and compared with the previously predicted premium to determine the difference. This difference is then used to optimize the multimodal medical data fusion and dynamic diagnostic model. For example, if the real-time claim amount is significantly lower than the predicted premium, it may indicate that the model's risk assessment is too conservative, requiring adjustments to the model's weights on relevant medical data to more accurately reflect actual risk and make the predicted premium more reasonable.

[0051] Positive feedback typically indicates that the output of the real-time risk scoring model tends to be robust (e.g., the score stabilizes in the low-risk range). In this case, by comparing the difference between the real-time payout amount and the predicted premium, the bias in risk prediction by the multimodal diagnostic model can be identified, enabling it to more accurately capture the correlation between medical data and actual payouts. The real-time payout amount is the actual payout incurred during the performance of the insurance contract (e.g., insurance money received by a patient due to a diagnosed disease), while the predicted premium is the premium amount calculated based on the risk value output by the multimodal diagnostic model (e.g., "high risk corresponds to high premium"). The difference between the two can be expressed in three dimensions: absolute difference, which is the difference between the real-time payout and the predicted premium (e.g., a payout of 10,000 yuan and a premium of 8,000 yuan result in a difference of -2,000 yuan, indicating a deficit; or a payout of 5,000 yuan and a premium of 8,000 yuan result in a difference of +3,000 yuan, indicating excess return); difference rate, which is the proportion of the difference to the predicted premium (e.g., +3,000 yuan / 8,000 yuan = 37.5%), used to standardize the degree of difference under different premium sizes; and difference trend, which is the change in difference over several consecutive periods (e.g., the difference rate increases from 10% to 30%), used to determine whether the bias is an occasional fluctuation or a systemic problem. For example, if the real-time payout for the "diabetic patient group" is 20% lower than the predicted premium for three consecutive months, it indicates that the multimodal model may have overestimated the risk of disease progression in this group, resulting in an overestimated predicted premium, and this bias needs to be corrected by optimizing the model. Optimization of multimodal diagnostic models based on differences involves adjusting the model's feature extraction and weight allocation for medical data to make diagnostic results (disease probability, confidence level) more closely reflect actual payout patterns. If the payout amount is significantly lower than the predicted premium (high positive difference rate), it indicates that the model's risk assessment is conservative, and the weight of "non-critical risk features" should be reduced. For example, in diagnosing "diabetic patients," if the model overemphasizes "occasional fluctuations in blood sugar" (a non-persistent feature) and overestimates risk, the weight of this feature should be reduced (e.g., from 0.2 to 0.1), while the weight of "long-term achievement of glycated hemoglobin targets" (a stability feature) should be increased (e.g., from 0.3 to 0.4) to make the disease probability prediction closer to actual payout needs. If the payout amount is significantly higher than the predicted premium (high negative difference rate), it indicates that the model underestimates potential risks, and the capture of "high-payout-related features" should be strengthened.For example, the model may overlook the feature of "hypertensive patients with proteinuria" which is strongly correlated with "chronic kidney disease payout". Therefore, the weight of this feature needs to be increased during multimodal fusion (e.g., from 0.1 to 0.3), and its memory period in the time-series model should be extended (e.g., from 3 months to 6 months) to ensure that risk signals are not missed. In the dynamic correction of confidence calculation logic, if the discrepancy stems from the contradiction between "high-confidence diagnostic results and low payouts" (e.g., the model predicts "high risk" with 90% confidence, but the actual payout is low), the confidence generation mechanism needs to be adjusted. For example, "historical payout matching degree" can be introduced as a correction term (e.g., if the historical payout rate for a certain type of diagnosis is only 60%, then its confidence level needs to be reduced by 60%) to avoid falsely high confidence levels misleading premium calculations.

[0052] When the feedback mechanism is reversed, if the accuracy of the AI-powered intelligent diagnosis improves to a preset optimization threshold, it will trigger the optimization of the hedging ratio of the optimal hedging strategy in the real-time risk scoring model. By adjusting this ratio, the reverse feedback indicates that the real-time risk scoring model may have insufficient risk coverage (such as a persistently high score), and the improvement in the accuracy of the AI-powered intelligent diagnosis provides a more reliable basis for optimizing the hedging strategy. At this point, the hedging ratio is adjusted through the trigger mechanism to make the hedging strategy more suitable for the current accuracy of the AI-powered intelligent diagnosis, ensuring that capital utilization efficiency is improved while effectively controlling risk.

[0053] The accuracy of AI-powered intelligent diagnosis is measured by the "match between diagnostic results and actual medical outcomes." Specific indicators include: disease prediction accuracy (the degree of agreement between the disease probability predicted by the real-time risk scoring model and the patient's final diagnosis, e.g., in cases predicting an "80% probability of lung cancer," the actual diagnosis rate is over 75%); confidence reliability (whether the accuracy of high-confidence diagnoses (e.g., ≥80%) is significantly higher than that of low-confidence diagnoses (e.g., ≤50%), reflecting the model's ability to grasp its own judgment); and missed / misdiagnosed rate (the proportion of missed diagnoses of critical diseases (e.g., myocardial infarction, cancer) or the proportion of non-diseases misdiagnosed as diseases), directly related to serious payout risks. Preset optimization thresholds need to be set in conjunction with business risk tolerance (e.g., "disease prediction accuracy ≥85%, missed diagnosis rate ≤5%)." When AI diagnosis reaches or exceeds this threshold through iteration, it indicates that the reliability of its output risk signals (disease probability, confidence) has significantly improved, meeting the conditions for optimizing hedging ratios. The optimal hedging strategy's hedging ratio (e.g., "60% risk reserve coverage + 30% reinsurance coverage + 10% retention") needs to be adjusted based on more accurate risk predictions. The core is to reduce the cost waste caused by "over-hedging" while ensuring coverage of high-risk scenarios: Increasing the retention ratio, with improved AI diagnostic accuracy, makes predictions of "low-risk scenarios" more reliable, allowing for an appropriate increase in the retention ratio (e.g., from 10% to 20%), reducing the use of reinsurance and other tools (e.g., reducing the reinsurance ratio from 30% to 20%), and lowering hedging costs; focusing on extreme risks. The increased coverage allows for the redirection of hedging resources (such as reduced reinsurance expenditures) to extreme risk scenarios (e.g., "million-dollar high payouts"), increasing the hedging ratio for corresponding risk ranges (e.g., from 5% to 15%) to ensure adequate protection even in low-probability, high-risk events not fully covered by AI-powered diagnostics. Dynamic risk distribution matching, combined with the optimized risk distribution curves from AI diagnostics (e.g., revising the probability of high-risk ranges from 10% to 7%), lowers the hedging intensity for those ranges (e.g., reducing risk reserves from 60% to 50%), making the hedging ratio more closely match the actual risk probability. For example, when the AI's accuracy in diagnosing "stroke" increases from 70% to 90%, its prediction of "high-risk patients" becomes more precise, allowing the reinsurance ratio for this group to decrease from 40% to 30%. Simultaneously, the saved costs can be used to purchase "excess payout insurance for rare complications," improving overall hedging efficiency. The two types of feedback complement each other: positive feedback corrects the "medical data → risk prediction" link through monetary differences, while negative feedback optimizes the execution of "risk prediction → hedging strategy" by improving diagnostic accuracy. Ultimately, this drives the entire "medical diagnosis - insurance risk control" system to continuously evolve in a dynamic balance, ensuring both risk controllability and improved operational efficiency.

[0054] S50: Based on the optimization scheme, recalculate the hedging ratio of the optimal hedging strategy in the real-time risk scoring model and optimize the real-time risk scoring model.

[0055] Recalculating the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimized scheme, and further optimizing the real-time risk scoring model, is a key iterative step after the implementation of the two-way feedback mechanism. Its core is to enable the hedging strategy and the risk scoring model to more accurately match the optimized system state, forming a deep linkage between "optimized scheme - hedging ratio - scoring model".

[0056] The hedging ratio refers to the proportion of various risk management tools (such as risk reserves, reinsurance, and financial derivatives) in the total risk exposure within the optimal hedging strategy (e.g., "risk reserves cover 60% + reinsurance covers 30% + retention 10%"). The recalculation process needs to be closely integrated with positive or negative feedback optimization schemes, and the ratio structure should be adjusted accordingly to ensure that hedging efficiency matches the current risk characteristics. The positive approach reduces the system's conservatism by releasing redundant risk control resources; therefore, the hedging ratio needs to be tilted towards "reducing excessive hedging." If the positive approach "reduces the proportion of special risk reserves" (e.g., from 120% to 100%), then the proportion of risk reserves in the hedging portfolio needs to be reduced simultaneously (e.g., from 50% to 40%). At the same time, the "self-retained risk ratio" can be appropriately increased (e.g., from 10% to 20%), and the released funds can be used to improve services or investments. If the plan "increases the sum insured" (e.g., from 300,000 yuan to 400,000 yuan), the new risk exposure brought about by the increase in the sum insured needs to be assessed. However, since the overall risk score is reduced, there is no need to increase the hedging intensity in proportion. Only the reinsurance ratio can be slightly adjusted (e.g., from 30% to 35%) to balance costs and coverage needs. The adjustment needs to refer to the "difference between real-time claims and predicted premiums" (e.g., claims are much lower than premiums). If the difference shows that the risk is significantly overestimated, then the total hedging scale should be further reduced (e.g., from the original coverage of 90% to 80%) to avoid idle funds. Negative risk mitigation strategies expose potential risk vulnerabilities by enhancing risk capture capabilities; therefore, the hedging ratio needs to be tilted towards "strengthening high-risk coverage." If the negative risk mitigation strategy includes "supplementing extreme risk hedging tools" (such as introducing excess-payout reinsurance), a specific ratio needs to be allocated to the new tools (such as adding "excess reinsurance coverage of 15%)," while reducing the proportion of regular risk reserves (such as reducing it from 50% to 40%) to focus on extreme scenarios. If the strategy includes "tiered premium increases" (such as a 20% premium increase for users with a score of 80 or above), the increase in premium income can partially offset this. While there is risk, the overall hedging strength needs to be increased due to the higher risk score (e.g., from 90% to 95% coverage). The hedging ratio for high-risk areas (e.g., reinsurance for payouts of 30,000-50,000 yuan) should be prioritized for improvement. When the accuracy of AI diagnosis reaches the target and triggers optimization (e.g., a preset threshold of 90%), it indicates that the reliability of risk prediction has improved. Based on a more accurate risk distribution curve (e.g., the probability of extreme risk is revised from 5% to 7%), the hedging ratio for the corresponding range can be increased (e.g., from 10% to 15%) to avoid underestimating risk due to a "reduced misjudgment rate". The real-time risk scoring model is optimized based on the "actual risk coverage effect after adjusting the hedging ratio," correcting the weights and calculation logic of the scoring dimensions to improve the accuracy of risk prediction. If the "self-retained risk ratio increases" after a positive solution and the actual payout does not exceed expectations, it indicates that the "risk-bearing capacity" is enhanced. The weight of the "hedging effectiveness" dimension in the model can be appropriately reduced (e.g., from 20% to 15%), while the weight of the "patient health improvement trend" can be increased (e.g., from 30% to 35%), making the scoring more focused on actual risk changes. If the "extreme risk hedging ratio increases" after a negative solution, effectively reducing the incidence of large payouts, it indicates that the "extreme risk response" capability is enhanced. The weight of the "probability of high-risk scenarios" in the model needs to be reduced (e.g., from 25% to 20%), while the weight of the "hedging tool suitability" dimension (e.g., the matching degree between reinsurance clauses and risk characteristics) needs to be added or increased (e.g., set to 15%) to reflect the value of strategy optimization. If the "secondary verification" after a positive strategy shows stable scores (e.g., maintaining a low level for three consecutive rounds), the strategy is effective. The model needs to reduce the "impact coefficient of short-term fluctuations on the score" (e.g., from 0.3 to 0.2) to avoid inflated scores due to temporary data disturbances. If the score falls back to a reasonable range after a negative strategy (e.g., from 85 to 70), risk control is effective. The model needs to strengthen the "weight decay logic of negative optimization measures" (e.g., the impact of the "emergency adjustment coefficient" gradually decreases over time to avoid long-term overestimation of risk). In the long term, the "risk coverage deviation after hedging ratio adjustment" (e.g., the difference between actual payouts and hedging expectations) needs to be included in the model training data. The scoring formula should be optimized through iterative learning (e.g., for cases of "excess payouts due to insufficient hedging," the score contribution of the corresponding risk characteristics should be increased).

[0057] After recalculating the hedging ratio and optimizing the scoring model, the effectiveness is verified through a closed-loop process of "real-time monitoring - deviation analysis - readjustment": The actual cost and risk coverage effect of the adjusted hedging strategy are compared (e.g., whether extreme risk payouts are reduced, and whether capital utilization efficiency is improved). If costs decrease but coverage is not weakened, the ratio is reasonable. The optimized scoring model is then examined to see if it can more accurately predict risks (e.g., whether the correlation between the score and subsequent actual payouts increases from 0.7 to 0.8). If the deviation narrows, the model iteration is effective. If "the score fluctuates abnormally after the hedging ratio adjustment" (e.g., the score suddenly increases after a positive adjustment), the optimization plan and ratio calculation logic need to be reviewed to correct potential conflicts (e.g., the increase in the sum assured exceeds the rate of risk reduction). Through this process, the hedging ratio of the hedging strategy and the real-time risk scoring model are dynamically adapted, ensuring that hedging measures accurately respond to optimized risk characteristics and that the scoring model achieves more efficient and accurate risk management.

[0058] The method in this application also includes: comparing the reduction of the payout cost ratio of traditional actuarial schemes under two different dataset models in the test; when the payout cost ratio deviates significantly, the real-time risk scoring model will be triggered to return to the previous historical optimal state, the incremental learning function will be enabled, and the defects of the real-time risk scoring model will be repaired based on the incremental learning.

[0059] Testing the ratio of reduced payout costs under two different dataset models, as well as the model repair mechanism when the ratio deviates, is a key step in ensuring the robustness of the real-time risk scoring model. The core is to verify the cost optimization effect of the model through comparison and establish a rapid recovery and iteration mechanism after failure.

[0060] The payout cost ratio refers to the ratio of the actual payout cost after applying the model to the payout cost of the traditional actuarial approach (e.g., "new model payout cost / traditional approach payout cost = 80%", meaning a 20% reduction). Testing should be based on two significantly different dataset models (e.g., "dataset A focuses on chronic disease patients" and "dataset B covers all diseases"). A controlled experiment should be conducted to quantify the cost optimization effect and ensure the model's stability under different data distributions. The two datasets need to be differentiated in terms of core features (e.g., A contains 70% chronic disease data + 30% common disease data, B contains 30% chronic disease data + 70% common disease data), and both must include the traditional actuarial plan payout records and the real-time payout records after the application of the new model for the corresponding patients to ensure data comparability. For each model, calculate the "total payout cost reduction rate" ((total payout of the traditional plan - total payout of the new model) / total payout of the traditional plan × 100%), and further subdivide it into "reduction rate by disease dimension" (e.g., reduction rate of chronic disease payout, reduction rate of acute disease payout) and "reduction rate by risk interval" (e.g., cost optimization margin for low-risk users, high-risk users), to comprehensively evaluate the model's cost control capabilities in different scenarios. Use statistical methods (e.g., t-test) to verify whether the difference in the ratios between the two models is statistically significant (e.g., model A's reduction rate is 25% ± 3%, model B's reduction rate is 20% ± 4%, is the difference due to data distribution rather than random fluctuations), to avoid misjudging model performance due to data bias. If the overall reduction rate of both models remains stable within a preset reasonable range (e.g., 15%-30%), and the trends of the subdivision ratios of disease types and risk ranges are consistent (e.g., both are more effective at optimizing costs for high-risk users), it indicates that the models have universality under different data distributions and that the cost control logic is reliable. If the ratios of the two models differ significantly (e.g., Model A reduces costs by 30%, while Model B only reduces costs by 5%), it is necessary to trace the differences in data characteristics (e.g., a higher proportion of high-cost rare diseases in Data B), analyze the model's shortcomings in adapting to specific data types, and provide direction for subsequent optimization. If the claims cost ratio is found to deviate significantly from the preset threshold during testing or actual application (e.g., the normal range of 15%-30%, but actually drops to below 5% or rises to above 40%), it indicates that the risk scoring model may have misjudged the risk (e.g., excessively underestimating the payout leads to damage to user rights, or overestimating the risk leads to failure of cost control). The protection mechanism will be triggered immediately to ensure the stability of insurance operations.

[0061] When the payout ratio exceeds the preset upper or lower limits (e.g., below 10% or above 35%), it is judged as "significant deviation" regardless of short-term fluctuations or continuous trends; when the payout ratio drifts outside the threshold for three consecutive cycles (e.g., one week per cycle) (e.g., gradually rising from 25% to 38%), it is judged as "potential systemic deviation" even if it does not reach the absolute threshold; the trigger signal is automatically generated by the real-time risk scoring model, combined with manual review (e.g., checking for data collection errors, policy changes, and other external factors), and after confirming that the deviation is not caused by accidental factors, the next step of operation is initiated. The historical optimal state refers to the state in which the risk scoring model has achieved "stable payout ratio within a reasonable range, highest correlation between risk score and actual payout (e.g., correlation coefficient above 0.8), and lowest user complaint rate" in past applications. Its parameter configuration (e.g., hedging ratio, scoring dimension weights) is archived as the "baseline version". The current model's decision output is paused for a rollback operation. The archived baseline version parameters are called to restart the real-time risk scoring model, including restoring the historical hedging ratio (e.g., risk reserve ratio of 50% + reinsurance of 30%) and resetting the scoring dimension weights (e.g., the "disease probability" weight is restored to 30%), ensuring that the model temporarily returns to a verified robust state. After the rollback, the real-time data collection function of the current model is retained to accumulate samples for subsequent incremental learning. At the same time, the rollback effect is verified through "dual-track operation" (the baseline model outputs decisions, and the current model calculates and compares them simultaneously) to avoid the rollback itself causing new fluctuations. After reverting to the historical optimal state, incremental learning is used to specifically address model flaws that cause deviations in the payout cost ratio. This allows the risk scoring model to retain its historical advantages while absorbing new data, achieving a closed loop of "repair-iteration-relaunch." The root causes of payout cost ratio deviations are analyzed (e.g., the risk scoring model's "lagging cost predictions for new treatment technologies" or "failure to capture sudden changes in regional incidence rates in a timely manner"), identifying specific risk scoring modules that need repair (e.g., the dynamic adjustment factor calculation module and the risk value mapping logic). Collect abnormal data that caused deviations before rollback (such as claims for new treatments and patient data on sudden increases in regional incidence rates), and combine it with high-quality data from the best historical state (as a benchmark sample) to form a hybrid training set of "abnormal samples + benchmark samples" to ensure that learning focuses on the defective scenarios; set clear repair indicators (such as "reducing the prediction error of new treatment costs from 30% to 10%" and "shortening the response delay of regional incidence rates from 7 days to 2 days") to avoid incremental learning generalization without focus. For the identified defective modules, the parameters of other parts of the risk scoring model are frozen, and only the target modules (such as the weight coefficients of the dynamic adjustment factors and the generation logic of the risk distribution curve) are retrained. The parameters are optimized by gradient descent to reduce interference with the overall stability of the model. After each round of learning, the repair effect is tested using a validation set that was not used in the training (such as the latest payout data). If the indicators meet the target (such as the cost prediction error being reduced to the target value), the learning stops. If the indicators do not meet the target, more defective samples are added and the training is repeated. The repaired modules are integrated with the risk scoring model framework of the historical best state to generate a "repaired version". The payout cost ratio is verified to have returned to a reasonable range (such as 15%-30%) through "offline simulation" (simulating decision-making with new data during the rollback period), and the correlation between the risk score and the actual payout is better than the baseline version. The patched version is first applied to a small percentage of users (e.g., 10%) to compare its payout cost ratio and risk scoring accuracy with the baseline model. Once no anomalies are confirmed, the application scope is gradually expanded. After deployment, the fluctuation of the ratio and the recurrence of defective scenarios are continuously monitored. If deviations occur again, the "rollback-incremental learning" process is repeated to form a long-term mechanism for the risk scoring model to self-correct. Comparative testing verifies the risk scoring model's cost optimization capabilities under different data scenarios. It can also quickly roll back to a safe state when the model fails and accurately fix defects through incremental learning, ensuring that the real-time risk scoring model always maintains its core capabilities of "effective cost control and accurate risk prediction," supporting the sound operation of the insurance business.

[0062] The method further includes: adding homomorphic encryption to the transmission of diagnostic results and risk score gradients in the multimodal medical data fusion and dynamic diagnostic model and real-time risk scoring model; and adding noise to the transmission of diagnostic results and risk score gradients based on differential privacy to enhance the protection of patient privacy and information privacy in insurance contracts.

[0063] In multimodal medical data fusion and dynamic diagnostic models, as well as real-time risk scoring models, protecting the privacy of diagnostic results and risk scoring gradients during transmission through homomorphic encryption and differential privacy technologies is a key means of balancing data utilization and privacy security. These two technologies address this from two dimensions: "encrypted transmission" and "privacy processing," ensuring that sensitive patient medical information and insurance contract details are not leaked during transmission, while simultaneously not affecting the normal operation and decision-making of the model. The core advantage of homomorphic encryption is that it allows for direct computation on encrypted data without decryption. This characteristic enables it to support parameter transmission and collaborative computation between models while protecting data privacy, making it suitable for cross-node transmission scenarios of medical diagnostic results (such as disease probability and confidence levels) and risk scoring gradients (such as the update direction and magnitude of scoring model parameters).

[0064] Real-time diagnostic results generated by multimodal medical data fusion and dynamic diagnostic models (such as "85% probability of lung cancer, 90% confidence level") need to be transmitted to a real-time risk scoring model as the core input for risk assessment. With the introduction of homomorphic encryption: During the encryption process, at the output of the diagnostic model, additive or multiplicative homomorphic encryption algorithms (such as the Paillier algorithm and the BFV algorithm) are used to encrypt the diagnostic results. For example, "85% probability" is converted into ciphertext, ensuring that even if intercepted during transmission, attackers cannot decipher the original value. After receiving the ciphertext, the real-time risk scoring model can directly perform risk value mapping calculations based on the encrypted diagnostic results without decryption (such as generating encrypted risk values ​​by combining insurance policy payout coefficients). Because homomorphic encryption supports operations such as "ciphertext + plaintext = ciphertext" and "ciphertext × plaintext = ciphertext," the entire risk transformation process can be completed in an encrypted state, avoiding privacy leaks in intermediate stages. Only after the final risk score is generated can an authorized node (such as the core server of an insurance company) use a private key to decrypt and obtain the plaintext score that can be used for decision-making, strictly controlling decryption permissions. Parameter optimization in real-time risk scoring models relies on gradient information (e.g., "disease probability weights need to be increased by 5%" and "the sensitivity of dynamic adjustment factors needs to be decreased by 3%)." These gradients need to be fed back to the diagnostic model for bidirectional optimization. Homomorphic encryption is used in this process: the parameter adjustment gradients (essentially numerical vectors) calculated by the risk scoring model are first encrypted before being transmitted to the multimodal diagnostic model. For example, "disease probability weight gradient + 5%" is encrypted to ensure that the specific value and direction of the gradient are not obtained by unauthorized parties. After receiving the encrypted gradient, the diagnostic model directly updates its own parameters in the encrypted state (e.g., "original weights + encrypted gradient = new weight ciphertext"), completing model iteration without decryption. This avoids reverse engineering of the model structure due to gradient information leakage (e.g., attackers use gradients to infer training data features). A distributed key generation and management mechanism (e.g., multi-party secure computation collaborative key generation) is adopted to prevent the entire encryption system from failing due to single-node key leakage. It also supports key updates when nodes dynamically join or leave, adapting to distributed model deployment scenarios. Differential privacy prevents attackers from distinguishing individual information by observing transmitted data by "adding carefully designed noise to the data," while ensuring that the overall statistical properties of the data are not compromised. It is suitable for further strengthening the privacy protection of diagnostic results and risk gradients, and is especially effective in resisting "correlation analysis attacks" (such as combining publicly available information to infer patient identity). Diagnostic results (such as "70% probability of diabetes, 85% confidence level") contain sensitive information about the patient's health status. Adding differential privacy noise requires a balance between "privacy protection strength" and "data availability": select the noise distribution (such as Gaussian noise, Laplace noise) based on the numerical range of the diagnostic results. For example, for probability values ​​between 0 and 100%, Gaussian noise with a mean of 0 and a variance related to a preset privacy budget (ε value) is added (e.g., when ε=1, the variance is set to 0.01). This ensures that the encrypted result (e.g., 70% ± 1.5%) both masks the true value and remains within a reasonable medical logic range (e.g., the probability will not exceed 100% due to noise). For high-confidence diagnostic results (e.g., confidence above 90%), the noise intensity is appropriately reduced (e.g., variance 0.005) to avoid excessive interference with key medical decisions. For low-confidence results (e.g., confidence below 60%), the noise intensity can be increased (e.g., variance 0.02) to enhance privacy protection, as these results themselves have a weaker impact on risk assessment. If the transmitted results are batch diagnostic results (e.g., a diagnostic set of 100 patients in a certain region), an "aggregated noise" strategy is used to add noise to individuals while ensuring that the overall statistics (e.g., the average disease probability) are not significantly affected, thus meeting the actuarial needs for group risk analysis. The risk score gradient (e.g., "the weight of the regional incidence rate adjustment factor needs to be increased by 2%) reflects the model's sensitivity to individual risk. Adding noise can prevent attackers from tracking specific users' insurance contract information (e.g., coverage amount, past claims records) through gradients: First, the gradient values ​​are pruned (e.g., the maximum absolute value is limited to 5%) to avoid extreme gradients causing noise to run away; then, the noise intensity is dynamically adjusted according to the gradient size (e.g., when the absolute value of the gradient is 2%, ±0.3% noise is added; when the gradient is 5%, ±0.5% noise is added) to ensure that the noise is proportional to the gradient, both masking the specific values ​​and preserving the adjustment direction; the total privacy budget (e.g., ε=2) is allocated to different gradient dimensions (e.g., disease probability-related gradients are allocated ε=1, insurance clause-related gradients are allocated ε=1) to avoid excessive consumption of the privacy budget by a single dimension, leading to an imbalance in protection; for the same gradient transmitted in multiple consecutive periods (e.g., "drug price fluctuation factor gradient"), correlated noise (e.g., adjacent period noise is positively correlated) is used to reduce the interference of noise on the model's iterative stability and ensure that the long-term optimization direction is not affected. The dual protection of encryption and noise complements each other, covering different privacy risks during transmission. Homomorphic encryption primarily defends against eavesdropping on transmission channels (such as man-in-the-middle attacks), ensuring the complete transmission of data in an encrypted state. Differential privacy primarily defends against data analysis attacks (such as associating individual information through multiple transmission results), and even if the ciphertext is decrypted, the added noise can prevent the leakage of individual information. The parameters of the two technologies are dynamically adjusted according to the data sensitivity. For example, for diagnostic results directly related to patient medical records (such as HIV positive probability), a "high encryption strength (such as 2048-bit key) + high noise strength (ε=0.5)" is used; for risk gradients not directly related (such as the group average insurance amount adjustment gradient), a "normal encryption strength + low noise strength (ε=3)" is used, protecting core privacy while reducing the impact on model efficiency. The dual protection mechanism is more likely to meet the requirements of medical privacy regulations (such as HIPAA, GDPR) and insurance data security standards (such as the Personal Information Protection Act). For example, homomorphic encryption proves that "data transmission was not leaked throughout the entire process," and differential privacy proves that "individuals cannot be identified," providing technical endorsement for business compliance. By introducing homomorphic encryption and differential privacy into the transmission of diagnostic results and risk scoring gradients, multimodal medical data fusion models and real-time risk scoring models can fully utilize data-driven decision-making while blocking privacy leakage paths from the technical level, ultimately achieving the goals of "data usable but invisible" and "model optimization without leakage," thus building a solid privacy security defense for collaborative applications in healthcare and insurance.

[0065] The method further includes: receiving global data; training the global data for several rounds using a FedAvg-type distributed optimization algorithm with gradient descent based on local data to obtain updated local data; and uploading the updated local data to all terminal devices participating in the training.

[0066] Receiving global data, training FedAvg-like algorithms based on local data, and uploading updated data are the core processes in distributed machine learning that enable collaborative model optimization without data sharing. This is particularly suitable for data privacy-sensitive scenarios such as healthcare and finance. It protects data locality while enabling collaborative model evolution across multiple devices. Global data does not refer to "all raw data" (to avoid privacy leaks), but rather to "global model parameters" generated by a central server or coordinating node, serving as the initial baseline for local training on each device. Its core function is to provide a unified starting point for training processes distributed across different devices, ensuring that the optimization direction of each device is consistent. Global data typically exists in the form of a "model parameter set," specifically including: the model's basic weights and biases (such as the weight matrices of each layer of a neural network and the parameters of the convolutional kernel); training configuration information (such as the learning rate, the maximum number of iterations, and the type of loss function); and historical optimization records (such as the global gradient direction and parameter adjustment magnitude in the previous training rounds), used to assist in judging the optimization trend during local training. For example, in a medical diagnostic model, global data may include "the weights of the disease feature extraction layer," "the bias parameters of the classifier," and a configuration of "learning rate of 0.01 and a maximum of 5 local training rounds." This information together constitutes the initial training baseline for each hospital terminal. Terminal devices (such as hospital servers and user mobile phones) receive global data from the central node through an encrypted communication link (such as the TLS protocol) to ensure that parameters are not tampered with or leaked during transmission. After receiving the data, the terminal verifies the integrity of the global data (such as through hash verification). If the data is found to be corrupted or incomplete, it requests the central node to resend the data to avoid starting training based on incorrect parameters. For new terminals that are dynamically added to the training (such as newly added hospitals), the central node will synchronously send historical global parameter snapshots to enable them to quickly catch up with the overall training progress and avoid affecting the efficiency of global optimization due to "new terminal lag".

[0067] After receiving the global data, each terminal does not upload its original local data. Instead, it uses the global parameters as a starting point and performs multiple rounds of gradient descent training with the local data to generate updated parameters adapted to the characteristics of the local data. This step is the core manifestation of the "decentralized" characteristic of FedAvg-like algorithms. It avoids data leakage through localized computation and allows the model to learn the unique patterns of the local data. Local data refers to raw data unique to the terminal device (such as patient medical records from a hospital or a user's health monitoring records), and its characteristics may differ from those of other terminals (such as differences in disease distribution across different regions). During training, the terminal loads global parameters into the local model, using the model output corresponding to the global parameters as the initial prediction; it calculates the loss between the prediction result and the true label using local data (such as classification error rate and mean squared error), and solves for the gradient (i.e., the direction of parameter adjustment) through backpropagation; it iteratively updates the global parameters based on gradient descent algorithms (such as SGD and Adam), with each update relying solely on local data and without interaction with other terminals. For example, if a hospital terminal's local data shows a high proportion of cases with diabetes mellitus complicated by nephropathy, the training process will focus on optimizing the model's parameters for recognizing this complication, making the updated parameters more closely match the characteristics of local patients. The key training strategy of FedAvg-like algorithms differs from the traditional distributed training's "single-round gradient upload." FedAvg-like algorithms allow the terminal to perform multiple rounds (e.g., 5-10 rounds) of local training. For example, after initializing the model with global parameters, the terminal will continuously train on local data for 5 rounds. Each round continues to optimize based on the parameter update results of the previous round, allowing the parameters to learn more fully from the features of the local data and reducing the upload frequency to lower communication costs. To avoid excessive parameter update amplitude (e.g., gradient explosion) caused by extreme local data (e.g., many outliers), the algorithm will prune the gradient (e.g., limit the maximum norm) or use an adaptive learning rate (e.g., reduce the learning rate with each training round) to ensure that local parameter updates are both effective and stable. Some improved algorithms (e.g., FedProx) will add a "proximal regularization term with global parameters" to the local loss function to constrain the deviation of local parameters from the global benchmark, avoiding excessive "personalization" of parameters due to special local data distribution (e.g., Non-IID data), which would affect the subsequent global aggregation effect. After completing local training, the terminal does not upload the original data. Instead, it uploads the "updated local parameters" (i.e., model parameters optimized based on global parameters) to the central node. The central node then aggregates and distributes the data to all terminals participating in the training, thereby achieving parameter sharing and global model iteration. After multiple rounds of training, the parameter set is uploaded as "updated local data," including the locally optimized model weights and biases (adjustments relative to global parameters); local training metadata (such as training rounds, local data volume, and loss descent curve) for the central node to judge parameter quality; and optional gradient information (some algorithms upload gradients instead of complete parameters) to further reduce the amount of data uploaded. For example, the parameters uploaded by a terminal might include "the weights of the disease classification layer are increased by 12% compared to global parameters," "the parameter adjustment matrix of the convolution kernel for complication identification," and metadata such as "after 5 rounds of local training, the loss decreased from 0.8 to 0.3." After receiving updated parameters from all terminals, the central node performs a weighted average based on the "terminal data volume ratio" (e.g., if a terminal's data volume accounts for 20% of the total data volume, then its parameter weight is 20%) to generate new global parameters. For example, if three terminals upload updated parameters with data volume ratios of 30%, 50%, and 20%, then the new global parameters = 30% × Terminal 1 parameter + 50% × Terminal 2 parameter + 20% × Terminal 3 parameter. After the new global parameters are generated, the central node encrypts and distributes them to all terminals participating in training (including the terminal that uploaded the parameters) as the benchmark for the next round of local training. For example, the aggregated global parameters may incorporate optimizations from Hospital A for "geriatric diseases" and Hospital B for "chronic diseases," enabling all terminal models to learn features from multiple data sources simultaneously. For terminals that are temporarily offline (e.g., due to equipment failure or network interruption), the central node caches the latest global parameters and prioritizes synchronization upon their re-entry, ensuring that all terminals are always trained based on the latest global benchmark and avoiding fragmentation of the global model due to terminal offline. By enabling multi-terminal collaboration through parameter uploading and downloading, the privacy of local data is protected (the raw data never leaves the terminal), while allowing the model to continuously evolve by integrating features from multiple data sources. In the medical field, this means that hospitals can jointly optimize disease diagnosis models without sharing patient medical records; in the insurance field, terminals can collaborate to improve the accuracy of risk assessment, ultimately achieving a win-win situation for both "privacy protection" and "model performance."

[0068] As can be seen, in the above scheme, regarding the deep coupling of medical data and insurance risk control and the synergistic two-way feedback optimization mechanism, the following steps are taken: First, historical medical data and insurance data of the patients are acquired and integrated to form a basic analysis dataset. Then, a patient-disease-insurance relationship is established based on a graph neural network, generating a dynamic risk feature vector. Based on this dynamic risk feature vector, a multimodal medical data fusion and dynamic diagnostic model is constructed. Based on this model, real-time patient medical data is diagnosed using AI intelligent diagnosis, and the patient disease probability and confidence level output by the AI ​​intelligent diagnosis results are mapped into insurance payout risk values. A risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived using option pricing theory. Based on the optimal hedging strategy, the predicted premium or predicted sum insured is adjusted in real time to construct a real-time risk scoring model driven by the diagnostic results. A two-way feedback optimization mechanism is established to generate positive or negative feedback optimization schemes based on the real-time risk scoring model. Based on the optimization schemes, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated to optimize the real-time risk scoring model. In this invention, the multimodal medical data fusion and dynamic diagnostic model continuously corrects the weight allocation of medical features through positive feedback and monetary difference optimization, making the prediction of disease probability and confidence level more in line with actual compensation patterns. It realizes the quantification of the probability of AI misdiagnosis into the price of insurance options to reduce the possibility of sudden large claims in insurance. By using closed-loop feedback, repetitive examinations and redundant claims processes are eliminated, which is expected to save 15%-20% of operating and management costs, improve the accuracy of medical diagnosis and risk assessment, reduce decision-making bias, and optimize insurance operation efficiency through a two-way feedback optimization mechanism, balancing risk and cost.

[0069] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0070] In one embodiment, a device for optimizing hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks is provided. This device corresponds one-to-one with the method for optimizing hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks described in the above embodiments. Figure 3 As shown, this medical AI intelligent diagnosis and insurance claim risk hedging strategy optimization device includes an integration module 101, a diagnosis module 102, a scoring module 103, a feedback module 104, and an optimization module 105. Detailed descriptions of each functional module are as follows: Integration module 101 is used to acquire the input patient's historical medical data and insurance data and integrate the data to form a basic analysis dataset; Diagnostic module 102 is used to establish patient-disease-insurance association based on graph neural network, generate dynamic risk feature vector, and construct multimodal medical data fusion and dynamic diagnostic model based on dynamic risk feature vector; The scoring module 103 is used to construct a real-time risk scoring model driven by diagnosis results. Based on the multimodal medical data fusion and dynamic diagnostic model, it uses AI intelligent diagnosis to diagnose real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy. Feedback module 104 is used to establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on the real-time risk scoring model. The optimization module 105 is used to recalculate the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme, thereby optimizing the real-time risk scoring model.

[0071] In one embodiment, the scoring module 103 is specifically used for: Based on multimodal medical data fusion and dynamic diagnostic models, AI intelligently diagnoses real-time patient medical data; Based on AI-powered intelligent diagnosis, real-time patient disease probability and confidence level are obtained. Based on the patient disease probability, confidence level, insurance clause payout coefficient, and dynamic adjustment factors, these are mapped and transformed into insurance payout risk values. Among these, dynamic adjustment factors include at least regional incidence rate and drug price fluctuations. The risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived by combining it with option pricing theory. Based on the optimal hedging strategy, the predicted premium or predicted sum assured is adjusted in real time, and a real-time risk scoring model driven by diagnostic results is constructed.

[0072] In one embodiment, the feedback module 104 is specifically used for: When the feedback is positive, the real-time compensation amount is obtained, and the difference between the real-time compensation amount and the predicted premium is compared. Based on the difference, the multimodal medical data fusion and dynamic diagnosis model is optimized. When feedback occurs, if the accuracy of AI-powered intelligent diagnosis is improved to a preset optimization threshold, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model will be optimized.

[0073] In one embodiment, the apparatus further includes: The model is used to test and compare the reduction of the payout cost ratio of traditional actuarial schemes under two different dataset models. When the payout cost ratio deviates significantly, the real-time risk scoring model will be triggered to return to its previous historical best state, and the incremental learning function will be activated to fix the defects of the real-time risk scoring model based on the incremental learning.

[0074] In one embodiment, the apparatus further includes: An adversarial training model is used to input insurance claim risk weights into a multimodal medical data fusion and dynamic diagnosis model for adversarial training, and to optimize the multimodal medical data fusion and dynamic diagnosis model based on the training results.

[0075] In one embodiment, the apparatus further includes: The encryption model is used to add homomorphic encryption to the transmission of diagnostic results and risk score gradients in multimodal medical data fusion, dynamic diagnostic models, and real-time risk scoring models; based on differential privacy, noise is added to the transmission of diagnostic results and risk score gradients to enhance the protection of patient privacy and information privacy in insurance contracts.

[0076] In one embodiment, the apparatus further includes: The training model receives global data; based on local data, it trains the global data for several rounds using a FedAvg-like distributed optimization algorithm with gradient descent to obtain updated local data; and all terminal devices that upload the updated local data participate in the training.

[0077] This invention provides a device for optimizing hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks. It integrates historical patient medical data and insurance data to form a basic analysis dataset. Based on a graph neural network, it establishes a patient-disease-insurance relationship, generating a dynamic risk feature vector. A multimodal medical data fusion and dynamic diagnosis model is constructed based on this dynamic risk feature vector. Using this model, real-time patient medical data is analyzed via AI-based intelligent diagnosis. The patient's disease probability and confidence level output by the AI-based intelligent diagnosis are mapped into an insurance payout risk value. A risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived using option pricing theory. Based on this optimal hedging strategy, the predicted premium or predicted sum assured is adjusted in real-time, constructing a diagnosis-driven real-time risk scoring model. A bidirectional feedback optimization mechanism is established to generate positive or negative feedback optimization schemes based on the real-time risk scoring model. The hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated based on the optimized scheme, thus optimizing the real-time risk scoring model. In this invention, the multimodal medical data fusion and dynamic diagnostic model continuously corrects the weight allocation of medical features through positive feedback and monetary difference optimization, making the prediction of disease probability and confidence level more in line with actual compensation patterns. It realizes the quantification of the probability of AI misdiagnosis into the price of insurance options to reduce the possibility of sudden large claims in insurance. By using closed-loop feedback, repetitive examinations and redundant claims processes are eliminated, which is expected to save 15%-20% of operating and management costs, improve the accuracy of medical diagnosis and risk assessment, reduce decision-making bias, and optimize insurance operation efficiency through a two-way feedback optimization mechanism, balancing risk and cost.

[0078] Specific limitations regarding the optimization device for hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks can be found in the limitations of the optimization method for hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks described above, and will not be repeated here. Each module in the aforementioned optimization device for hedging strategies for medical AI-based intelligent diagnosis and insurance claims risks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0079] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements server-side functions or steps of a method for optimizing medical AI intelligent diagnosis and insurance claims risk hedging strategies.

[0080] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps for a method to optimize a medical AI intelligent diagnosis and insurance claims risk hedging strategy. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire and integrate the input patient's historical medical and insurance data to form a basic analysis dataset; Based on graph neural networks, a patient-disease-insurance relationship is established, a dynamic risk feature vector is generated, and a multimodal medical data fusion and dynamic diagnostic model is constructed based on the dynamic risk feature vector; Based on multimodal medical data fusion and dynamic diagnostic model, this method uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy to construct a real-time risk scoring model driven by diagnostic results. Establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on a real-time risk scoring model; Based on the optimized scheme, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated, and the real-time risk scoring model is optimized.

[0081] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire and integrate the input patient's historical medical and insurance data to form a basic analysis dataset; Based on graph neural networks, a patient-disease-insurance relationship is established, a dynamic risk feature vector is generated, and a multimodal medical data fusion and dynamic diagnostic model is constructed based on the dynamic risk feature vector; Based on multimodal medical data fusion and dynamic diagnostic model, this method uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy to construct a real-time risk scoring model driven by diagnostic results. Establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on a real-time risk scoring model; Based on the optimized scheme, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated, and the real-time risk scoring model is optimized.

[0082] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0083] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0085] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for optimizing medical AI-based intelligent diagnosis and insurance claims risk hedging strategies, characterized in that, include: Acquire and integrate the input patient's historical medical and insurance data to form a basic analysis dataset; Based on graph neural networks, a patient-disease-insurance relationship is established, a dynamic risk feature vector is generated, and a multimodal medical data fusion and dynamic diagnostic model is constructed based on the dynamic risk feature vector; Based on multimodal medical data fusion and dynamic diagnostic model, this method uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves, combines option pricing theory to derive the optimal hedging strategy, and adjusts the predicted premium or predicted sum insured in real time based on the optimal hedging strategy to construct a real-time risk scoring model driven by diagnostic results. Establish a two-way feedback optimization mechanism to generate positive or negative feedback optimization schemes based on a real-time risk scoring model; Based on the optimized scheme, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model is recalculated, and the real-time risk scoring model is optimized.

2. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, The steps of constructing a real-time risk scoring model driven by diagnostic results, based on the multimodal medical data fusion and dynamic diagnostic model, include: using AI-powered intelligent diagnosis of real-time patient medical data, mapping the patient's disease probability and confidence level output by the AI-powered intelligent diagnosis results into an insurance payout risk value, generating a risk distribution curve using Monte Carlo simulation, deriving the optimal hedging strategy using option pricing theory, and adjusting the predicted premium or predicted sum insured in real time based on the optimal hedging strategy. Based on multimodal medical data fusion and dynamic diagnostic models, AI intelligently diagnoses real-time patient medical data; Based on AI-powered intelligent diagnosis, real-time patient disease probability and confidence level are obtained. Based on the patient disease probability, confidence level, insurance clause payout coefficient, and dynamic adjustment factors, these are mapped and transformed into insurance payout risk values. Among these, dynamic adjustment factors include at least regional incidence rate and drug price fluctuations. The risk distribution curve is generated using Monte Carlo simulation, and the optimal hedging strategy is derived by combining it with option pricing theory. Based on the optimal hedging strategy, the predicted premium or predicted sum assured is adjusted in real time, and a real-time risk scoring model driven by diagnostic results is constructed.

3. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, The steps for establishing a two-way feedback optimization mechanism and generating positive or negative feedback optimization schemes based on a real-time risk scoring model include: When the feedback is positive, the real-time compensation amount is obtained, and the difference between the real-time compensation amount and the predicted premium is compared. Based on the difference, the multimodal medical data fusion and dynamic diagnosis model is optimized. When feedback occurs, if the accuracy of AI-powered intelligent diagnosis is improved to a preset optimization threshold, the hedging ratio of the optimal hedging strategy in the real-time risk scoring model will be optimized.

4. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, After the step of recalculating the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme and optimizing the real-time risk scoring model, the method further includes: The test compares the ratio of reduced payout costs for traditional actuarial solutions under two different dataset models. When the payout cost ratio deviates significantly, the real-time risk scoring model will be triggered to return to its previous historical optimal state, and the incremental learning function will be activated to fix the defects in the real-time risk scoring model based on the incremental learning.

5. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, After the steps of establishing patient-disease-insurance relationships based on graph neural networks, generating dynamic risk feature vectors, and constructing a multimodal medical data fusion and dynamic diagnostic model based on the dynamic risk feature vectors, the method further includes: Input the insurance claim risk weights into the multimodal medical data fusion and dynamic diagnosis model for adversarial training, and optimize the multimodal medical data fusion and dynamic diagnosis model based on the training results.

6. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, After the step of recalculating the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme and optimizing the real-time risk scoring model, the method further includes: In multimodal medical data fusion and dynamic diagnostic models and real-time risk scoring models, homomorphic encryption is added to the transmission of diagnostic results and risk scoring gradients; Based on differential privacy, noise is added during the transmission of diagnostic results and risk score gradients to enhance the protection of patient privacy and information privacy in insurance contracts.

7. The method for optimizing medical AI intelligent diagnosis and insurance claim risk hedging strategies as described in claim 1, characterized in that, After the step of recalculating the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme and optimizing the real-time risk scoring model, the method further includes: Receive global data; Based on local data, the global data is trained several times using a FedAvg-type distributed optimization algorithm with gradient descent to obtain updated local data; Upload updated local data and all terminal devices participating in the training.

8. A device for optimizing medical AI-based intelligent diagnosis and insurance claim risk hedging strategies, characterized in that, include: The integration module is used to acquire input patient historical medical data and insurance data and integrate the data to form a basic analysis dataset. The diagnostic module is used to establish patient-disease-insurance relationships based on graph neural networks, generate dynamic risk feature vectors, and construct a multimodal medical data fusion and dynamic diagnostic model based on the dynamic risk feature vectors. The scoring module is used to generate risk distribution curves based on multimodal medical data fusion and dynamic diagnostic models. It uses AI intelligent diagnosis to analyze real-time patient medical data and maps the patient disease probability and confidence level output by AI intelligent diagnosis results into insurance payout risk values. It uses Monte Carlo simulation to generate risk distribution curves and combines option pricing theory to derive the optimal hedging strategy. Based on the optimal hedging strategy, it adjusts the predicted premium or predicted sum insured in real time to build a real-time risk scoring model driven by diagnostic results. The feedback module is used to establish a two-way feedback optimization mechanism, generating positive or negative feedback optimization schemes based on the real-time risk scoring model. The optimization module is used to recalculate the hedging ratio of the optimal hedging strategy in the real-time risk scoring model based on the optimization scheme, thereby optimizing the real-time risk scoring model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for optimizing the medical AI intelligent diagnosis and insurance claim risk hedging strategy as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the medical AI intelligent diagnosis and insurance claim risk hedging strategy optimization method as described in any one of claims 1 to 7.

Images