A disease-specific risk financing optimization method and device based on a double-layer value feedback loop

Abstract

The present application belongs to the field of medical information technology and intelligent insurance pricing, and relates to a disease-specific risk financing optimization method and device based on a double-layer value feedback loop, which comprises the following steps: collecting real-time multi-modal clinical data of patients, extracting time series derivatives, embedding semantics and fusing features to construct a global benchmark feature tensor; inputting a deep neural network to derive a high-dimensional clinical risk vector through cross-attention; calculating the actual settlement amount through logarithmic probability transformation, hyperbolic tangent boundary mapping and historical load negative feedback; packaging as a double-domain encrypted certificate and binding with the inpatient medical record number atomization to realize event-level precise mapping; triggering model sandbox retraining and hot loading with residual variance to form the first feedback loop of clinical counter-benefiting; reading the economic relief measurement value to dynamically adjust the message queue, cache preloading and container computing power to form the second feedback loop of operation scheduling. Cross-subsidies can be eliminated, model self-evolution and mathematical-level risk control can be realized, and the method is suitable for commercial health insurance and other scenarios.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and intelligent insurance pricing technology, and in particular to a method and apparatus for optimizing disease-specific risk financing based on a two-layer value feedback loop. Background Technology

[0002] In existing technologies, disease-specific risk financing optimization methods are generally divided into two independent technical routes: One approach is clinical risk prediction based on electronic medical records (EMR). This method typically involves collecting patients' electronic medical record data and using static models such as random forests, logistic regression, or deep neural networks to output the probability of postoperative complication risks (e.g., a score between 0 and 1) for clinical decision support (e.g., whether to perform prophylactic stoma). A typical example is the MySurgeryRisk system.

[0003] The second approach is dynamic insurance pricing based on health data. This method typically involves collecting customers' annual physical examination reports, lifestyle questionnaires, and wearable device data (steps, heart rate, etc.), and using AI models to calculate a comprehensive health score, which is then mapped to a premium discount rate. A typical example is Ping An Health Insurance's dynamic premium system.

[0004] The shortcomings of existing technologies include: Coarse pricing granularity: User-year pricing cannot match the economic impact of a single hospitalization event, resulting in distorted cross-subsidies such as over-subsidization of low-risk hospitalizations and under-subsidization of high-risk hospitalizations.

[0005] The data dimension is shallow: it relies on physical examination reports and wearable devices, and is not connected to the HIS / LIS / PACS system, so it cannot obtain core clinical data that determines the risk of complications, such as the dynamic flow of preoperative coagulation indicators and intraoperative imaging characteristics.

[0006] No postoperative outcome feedback: The feedback cycle is as long as 12 months, the model cannot capture acute complication events within 31 days after surgery, and it cannot self-evolve to adapt to clinical environment drift.

[0007] Prediction results are inconclusive: the output of risk probability stops at the information level and does not form a closed loop with the patient's economic burden. High-risk patients know that the risk is high but are unable to bear it. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a disease-specific risk financing optimization method based on a two-layer value feedback loop, employing the following technical solution, including the following steps: Real-time collection of patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history; and construction of a global benchmark feature tensor through time-series derivative feature extraction, semantic embedding, and feature fusion. The global baseline feature tensor is input into a deep neural network, and the temporal features and baseline features are fused through a cross-attention mechanism. The objective probability of occurrence of multiple complications is inferred through an independent Sigmoid output layer and combined to form a high-dimensional clinical risk vector. Based on the probabilities of various complications in the high-dimensional clinical risk vector, logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and negative feedback adjustment of historical economic load are performed in sequence, and the actual voucher settlement amount is calculated through deterministic analytical formulas. The actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number are encapsulated into a dual-domain encrypted fundraising voucher, and atomically bound to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher. Asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a measure of model drift, and trigger model retraining in a safe isolation sandbox when the threshold is exceeded. The online lossless update of model parameters is completed by hot loading the memory pointer in place, forming the first feedback loop of clinical feedback. The economic burden reduction metric is read from the financing certificate, the service routing priority weight parameter is calculated, and the service quality level of the distributed message queue is dynamically adjusted accordingly. The asynchronous cache preloading of distributed storage and the dynamic quota allocation of container computing resources are triggered to form a second feedback loop for operation resource scheduling.

[0009] Preferably, the step of constructing a global baseline feature tensor by real-time acquisition of patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history, followed by time-series derivative feature extraction, semantic embedding, and feature fusion, specifically includes: First and second derivatives were extracted and standardized from the time series of continuous clinical laboratory indicators to construct a time-series evolution feature sub-tensor. Using a pre-trained medical language model, low-dimensional semantic embedding is performed on unstructured electronic medical record text to generate semantic feature vectors; The temporal evolution feature sub-tensor, semantic feature vector, and discrete static medical history features are concatenated at the channel level and encapsulated to form a global baseline feature tensor.

[0010] Preferably, the step of inputting the global baseline feature tensor into a deep neural network, fusing temporal features and baseline features through a cross-attention mechanism, inferring the objective probability of occurrence of multiple complications through independent Sigmoid output layers, and combining them to form a high-dimensional clinical risk vector specifically includes: The temporal features in the global baseline feature tensor are fused with the baseline features through a cross-attention mechanism, and the objective probability of occurrence of multiple complications is inferred through an independent Sigmoid output layer, and combined to form a high-dimensional clinical risk vector. Based on the gradient-weighted class activation mapping algorithm, the contribution of each complication risk prediction score to the internal feature map of the network is calculated, and risk interpretation anchor points are generated on the informed consent form interface.

[0011] Preferably, the step of calculating the actual voucher settlement amount using a deterministic analytical formula by sequentially performing logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and negative feedback adjustment of historical economic load based on the probabilities of various complications in the high-dimensional clinical risk vector specifically includes: Log-odds transformation is performed on the probabilities of each complication in the high-dimensional clinical risk vector to map the probability space to a linear measure space; The transformed linear risk measure is weighted and summed by risk correction coefficients, then input into the hyperbolic tangent boundary constraint operator, and combined with the negative feedback adjustment factor of the historical economic load reduction cumulative value, to calculate the actual voucher settlement amount. The historical economic load reduction cumulative value is updated on a rolling basis using an index-weighted moving average formula, serving as the basis for adjusting the fairness of cross-hospitalization pricing.

[0012] Preferably, the step of encapsulating the actual voucher settlement amount, the high-dimensional clinical risk vector, and the hash fingerprint of the inpatient medical record number into a dual-domain encrypted fundraising voucher, and atomically binding it to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher specifically includes: The actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number are encapsulated into a dual-domain encrypted fundraising voucher containing a main data field and an extended data field. By using a distributed two-phase commit protocol, the fundraising voucher is atomically associated and bound to the inpatient medical record number in the hospital information system, ensuring a one-to-one mapping between a single hospitalization event and a single voucher.

[0013] Preferably, the steps of asynchronously monitoring objective outcome events during a preset postoperative observation period, calculating the residual variance between the actual outcome and the predicted probability as a model drift metric, triggering model retraining in a safe isolation sandbox when a threshold is exceeded, and completing online lossless updates of model parameters through in-situ hot loading of memory pointers to form the first feedback loop of clinical feedback specifically include: Asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the single-point residual between the actual outcome value and the preoperative predicted probability, and maintain a fixed-capacity circular buffer to calculate the residual variance as a measure of model drift. The residual variance is compared with a preset robustness threshold. When the threshold is exceeded, model retraining is triggered in a safe isolation sandbox. After retraining, the memory pointer is hot-loaded in place during the production inference process via remote invocation, enabling online lossless updates of model parameters.

[0014] Preferably, the step of reading the economic burden reduction metric from the financing certificate, calculating the service routing priority weight parameter, and dynamically adjusting the service quality level of the distributed message queue, triggering asynchronous cache preloading of distributed storage, and dynamically allocating container computing resources to form a second feedback loop for operational resource scheduling specifically includes: Read the economic burden reduction benchmark value from the extended data field of the financing certificate, and calculate the service routing priority weight parameter through a nonlinear mapping function; Based on the service routing priority weight parameters, the service quality level and consumption priority in the distributed message queue are dynamically adjusted. Based on the service routing priority weight parameters, the asynchronous cache preloading of distributed storage and the dynamic quota adjustment of container computing resources are triggered.

[0015] To address the aforementioned technical problems, this invention also provides a disease-specific risk financing optimization device based on a two-layer value feedback loop, employing the following technical solution, including: The module is used to collect patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history in real time. Through time-series derivative feature extraction, semantic embedding, and feature fusion, a global benchmark feature tensor is constructed. The combination module is used to input the global baseline feature tensor into a deep neural network, fuse temporal features and baseline features through a cross-attention mechanism, infer the objective probability of occurrence of multiple complications through an independent Sigmoid output layer, and combine them to form a high-dimensional clinical risk vector. The execution module is used to sequentially perform logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and historical economic load negative feedback adjustment based on the probability of each complication in the high-dimensional clinical risk vector, and calculate the actual voucher settlement amount through a deterministic analytical formula. The encapsulation module is used to encapsulate the actual voucher settlement amount, high-dimensional clinical risk vector and hash fingerprint of inpatient medical record number into a dual-domain encrypted fundraising voucher, and atomically bind it to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher. The calculation module is used to asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a measure of model drift, and trigger model retraining in a safe isolation sandbox when the threshold is exceeded. The model parameters are updated online without loss through in-situ hot loading of memory pointers, forming the first feedback loop of clinical feedback. The adjustment module is used to read the economic burden reduction metric value from the financing certificate, calculate the service routing priority weight parameter, and dynamically adjust the service quality level of the distributed message queue, trigger the asynchronous cache preloading of distributed storage, and dynamically allocate the quota of container computing resources, forming a second feedback loop for operation resource scheduling.

[0016] To address the aforementioned technical problems, the present invention also provides a computer device that employs the technical solution described below, comprising a memory and a processor. The memory stores computer-readable instructions, and the processor, when executing the computer-readable instructions, implements the steps of the aforementioned disease-specific risk financing optimization method based on a two-layer value feedback loop.

[0017] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium, which employs the technical solution described below. The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the aforementioned disease-specific risk financing optimization method based on a two-layer value feedback loop.

[0018] Compared with the prior art, the present invention has the following main advantages: (1) Event-level precise pricing can be achieved to eliminate cross-subsidization: By compressing the pricing granularity to a single hospitalization event and atomically binding the Token with the hospitalization medical record number, precise mutual assistance can be achieved.

[0019] (2) It can be driven by core clinical data and has more sensitive risk identification: By directly accessing real-time clinical data from HIS / LIS / PACS systems, it can extract time-series first / second derivative features and provide early warning of acute risks 24-48 hours in advance.

[0020] (3) Model self-evolution driven by postoperative outcome: Through residual variance monitoring and sandbox hot loading, the model evolution cycle is compressed from several months to hours.

[0021] (4) Mathematical architecture-level risk control: By embedding the Tanh boundary constraint operator in the pricing formula, the settlement amount is always bounded, preventing the penetration of the fund pool from the underlying mathematical level.

[0022] (5) It has a positive flywheel effect of economy and technology: the economic burden reduction measure is directly mapped to resource priority through the second feedback loop, forming an adaptive incentive mechanism in which the more burden is reduced, the better the service is.

[0023] (6) The entire process is explainable and auditable: Combining Grad-CAM attribution and Logit linear pricing, the risk basis of each voucher can be traced, meeting regulatory compliance requirements.

[0024] (7) It can be widely applied to medical financing scenarios such as commercial health insurance, inclusive insurance, and special disease pool, and has significant technological progress and commercial value. Attached Figure Description

[0025] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0026] Figure 1 This is a flowchart of an embodiment of the disease-specific risk financing optimization method based on a two-layer value feedback loop of the present invention; Figure 2 This is a schematic diagram of an embodiment of the disease-specific risk financing optimization device based on a two-layer value feedback loop of the present invention; Figure 3 This is a schematic diagram of the structure of an embodiment of the computer device of the present invention. Detailed Implementation

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects and not to describe a particular order.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0030] It should be noted that the disease-specific risk financing optimization method based on a two-layer value feedback loop provided in the embodiments of the present invention is generally executed by a server / terminal device, and correspondingly, the disease-specific risk financing optimization device based on a two-layer value feedback loop is generally installed in the server / terminal device.

[0031] It should be understood that the number of terminal devices, networks, and servers is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be used.

[0032] Example 1 Please refer to Figure 1 The flowchart illustrates an embodiment of the disease-specific risk financing optimization method based on a two-layer value feedback loop according to the present invention. The disease-specific risk financing optimization method based on a two-layer value feedback loop includes the following steps: Step S1: Real-time collection of patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history; and construction of a global baseline feature tensor through time-series derivative feature extraction, semantic embedding, and feature fusion.

[0033] In this embodiment, the electronic device (e.g., a server / terminal device) on which the disease-specific risk financing optimization method based on the two-layer value feedback loop runs can receive disease-specific risk financing optimization requests based on the two-layer value feedback loop via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future-developed wireless connection methods.

[0034] In this embodiment, step S1 may specifically include the following steps: S11 extracts and standardizes the first and second derivatives of the time series of continuous clinical laboratory indicators to construct the time series evolution feature sub-tensor.

[0035] The feature encapsulation module retrieves the patient's continuous monitoring sequence within time period t from the Laboratory Information System (LIS) via an internal secure API. The first derivative (rate of change) and the second derivative (acceleration) are calculated using a difference window: .

[0036] in: This represents the original value of the test index at time point t (such as D-dimer concentration, APTT seconds). The sampling interval (24 hours in one specific embodiment); The rate of change of the index at time t (first derivative), with units related to the original index / time; The acceleration of change (second derivative) reflects how fast a trend of deterioration or improvement changes.

[0037] Z-score standardization was then performed on the continuous clinical endpoint x: .

[0038] in: The mean of the entire historical sample (obtained from statistical analysis of laboratory indicators of hospitalized patients with the same disease in the past 12 months on the platform); The standard deviation is denoted as .

[0039] Finally, the standardized original time series values, first derivatives, and second derivatives are concatenated at the channel level to construct the temporal evolution feature subtensor. .

[0040] The first derivative captures abrupt changes in indicators (such as rapid deterioration of coagulation function), while the second derivative captures the acceleration of this deterioration, providing early warning of acute risks 24–48 hours earlier than single-point values. For example, when the second derivative of D-dimer is consistently greater than 2σ, it indicates an accelerating risk of thrombosis, and the system can adjust the risk vector upwards during the surgical registration phase. This method overcomes the limitation of existing technologies that rely solely on static snapshots (such as annual physical examinations) and cannot detect acute progression trends.

[0041] S12 utilizes a pre-trained medical language model to perform low-dimensional semantic embedding on unstructured electronic medical record text, generating semantic feature vectors.

[0042] The feature encapsulation module extracts free text information such as the chief complaint and present medical history from the electronic medical record (EMR) system, and performs feature vectorization using a pre-trained medical language model (Med-BERT-base in one specific implementation). The model maps variable-length text sequences into fixed-dimensional low-dimensional dense vectors. .

[0043] in: This represents the encoder of a pre-trained medical BERT model; The embedding dimension (128 in one specific implementation).

[0044] Unstructured medical records contain a wealth of information that structured anatomical indicators cannot express, such as "abdominal pain with bloody stools for 3 days" and "a history of aspirin resistance." Through the attention mechanism of Med-BERT, the model can capture long-range dependencies between words and relationships between medical entities, transforming them into numerical vectors that can participate in gradient descent. This method addresses the problem of significant information loss caused by existing techniques that only use structured fields (such as ICD encoding).

[0045] S13 concatenates the temporal evolution feature sub-tensor, semantic feature vector, and discrete static medical history features at the channel level to encapsulate them into a global baseline feature tensor.

[0046] The feature fusion unit performs channel-level concatenation of the following three types of features: Temporal Evolution Feature Tensor (Dimension: Number of time points × 3 channels); Semantic Embedding vector (dimension: 1×d); Discrete static medical history (dimension: k, encoded by one-hot encoding).

[0047] k represents the total dimension of the feature vector formed by one-hot encoding of discrete static medical history features.

[0048] Specifically, a patient's static medical history includes multiple categorical variables, such as: gender (male / female), blood type (A / B / O / AB), history of hypertension (yes / no), history of diabetes (yes / no), smoking history (yes / no / quit smoking), and allergy to specific drugs (yes / no). Each categorical variable, when one-hot encoded, is expanded into a binary vector of length equal to the number of possible values ​​for that variable, with only one dimension set to 1 and the rest to 0.

[0049] Therefore, k is the sum of the dimensions of all static medical history category variables included in the system after encoding. For example, if the system collects 10 binary medical histories (each requiring 1 dimension) and 2 four-category variables (each requiring 4 dimensions), then k = 10 × 1 + 2 × 4 = 18. The specific value of k is determined by the platform's preset medical history feature set, which ensures that all discrete information can be concatenated with continuous temporal features and semantic embedding features in numerical form to form a complete global baseline feature tensor, avoiding information loss.

[0050] The final global baseline feature tensor is formed as follows: .

[0051] In one specific implementation: the number of time points t=7 (7 days before surgery to 1 day before surgery, Δ=24h); 3 channels: raw value, first derivative, second derivative; d=128 (semantic embedding dimension); k is the discrete medical history unique thermo coding dimension; Step S13 fuses data from different modalities, sampling rates, and physical meanings into a compact tensor, enabling subsequent deep neural networks to simultaneously utilize temporal trends, semantic information, and static risk factors for joint inference. Compared to existing technologies that only use annual health check reports or single test values, this sub-step increases the information dimensionality by approximately two orders of magnitude.

[0052] The purpose of step S1 is to construct the global baseline feature tensor X. base This approach unifies heterogeneous raw clinical data (time series test values, unstructured text, discrete medical history) into standardized numerical tensors that can be processed by machine learning models, providing inputs with complete information and clear physical meaning for subsequent risk extrapolation.

[0053] Step S2: Input the global baseline feature tensor into the deep neural network, fuse the temporal features and baseline features through the cross-attention mechanism, and infer the objective probability of occurrence of multiple complications through the independent Sigmoid output layer, and combine them to form a high-dimensional clinical risk vector.

[0054] In this embodiment, step S2 may specifically include the following steps: S21 integrates the temporal features in the global baseline feature tensor with the baseline features through a cross-attention mechanism, and then uses an independent Sigmoid output layer to extrapolate the objective probability of multiple complications, combining them to form a high-dimensional clinical risk vector.

[0055] The clinical risk probability calculation network consists of a feature projection layer, a cross-fusion layer, and a residual output layer. First, the temporal evolution feature matrix is... With baseline feature vector Projecting to a unified dimension : .

[0056] in: These are the query, key, and value matrices, respectively. , The projection matrix is ​​learnable; To unify the projection dimension (64 in one implementation); d tm : Represents the temporal evolution characteristic matrix M time The feature dimension of each time step. In specific implementation, each time step contains 3 channels (original value, first derivative, second derivative), which, after projection, d tm This is actually equal to the number of features at that time step (i.e., 3, which increases accordingly if other engineering features are also included). M time The shape is [t, d] tm ], where t is the number of time steps (e.g., 7 days before surgery).

[0057] d bs : Represents the baseline feature vector V base The feature dimension. V baseIt consists of static features from the non-temporal portion of the global baseline feature tensor, including semantic embedding vectors (dimension d) and discrete one-hot codes for medical history (dimension k) concatenated. Therefore, d bs = d + k. V base The shape is [1, d]. bs ] represents the patient's static, non-temporal baseline information.

[0058] d tm and d bs They are heterogeneous and complementary. tm It describes the dynamic change pattern of the patient over time, while d bs This describes the patient's inherent, time-invariant, static background. In the cross-attention mechanism, this is achieved through the projection matrix W. Q W K W V Projecting these two feature spaces, which have different origins and different physical meanings, onto the same dimension d. k (e.g., 64-dimensional), thus enabling the calculation of their similarity (QK). T This design enables the interaction and fusion of dynamic temporal features and static baseline features. This allows the model to more accurately interpret the dynamic changes in clinical indicators based on the patient's individual baseline conditions.

[0059] Then perform cross-head attention computation: .

[0060] in: Calculate the similarity matrix between the time-series features and the baseline features; This is a scaling factor to prevent the dot product from becoming too large and causing gradient saturation. Softmax normalizes similarity into attention weights.

[0061] The final output layer is configured with m independent neurons, using the Sigmoid activation function: .

[0062] in: This indicates a category c complication; Let be the weight vector of the c-th neuron; This is the Sigmoid function, which outputs independent probabilities (non-exclusivity).

[0063] The final combination results in a high-dimensional clinical risk vector: .

[0064] The cross-attention mechanism allows for bidirectional information interaction between temporal features (such as the rate of change of coagulation parameters) and static baseline features (such as age and chronic disease history), better capturing the clinical logic that "a certain indicator is more dangerous to worsen under specific baseline conditions" than simple feature concatenation. The independent sigmoid output adapts to the real-world situation of multiple coexisting complications (patients may experience readmission and massive bleeding simultaneously), outperforming the exclusive softmax approach. This network design addresses the limitation of existing models in simultaneously predicting multiple concurrent outcomes.

[0065] S22, based on the gradient-weighted class activation mapping algorithm, calculates the contribution of each complication risk prediction score to the internal feature map of the network, and generates risk interpretation anchor points on the informed consent form interface.

[0066] The system uses the Grad-CAM algorithm to calculate risk prediction scores for specific complications. Relative to the network's internal feature map Partial derivatives: .

[0067] in: The predicted probability (scalar) of complication class c; The feature map of the k-th layer of the network (dimension H×W); For the prediction of the feature map of the k-th layer Contribution weight; This represents the total number of pixels in the feature map.

[0068] Then, using coordinate mapping technology, high-contribution input features (such as advanced age, specific medical history, or a sudden increase in a certain test indicator) are rendered in situ next to the corresponding risk clauses on the informed consent form's front-end interface to generate risk interpretation anchor points.

[0069] Step S22 enables white-box risk prediction. Patients can directly see "Your risk of anastomotic leakage is high, 42% from preoperative albumin levels and 31% from postoperative drainage fluid change rate," enhancing trust and compliance. It also meets regulatory requirements for algorithm auditability, resolving the "black box" problem of existing systems knowing the risk is high but not knowing why.

[0070] The purpose of step S2 is to convert the global baseline feature tensor X output in step S1 into a variable value. base Inputting the data into a deep neural network, the system fuses temporal and baseline features through a cross-attention mechanism, outputting m independent complication risk probabilities to form a high-dimensional clinical risk vector V. risk Grad-CAM is used to generate interpretable risk anchors.

[0071] Step S3: Based on the probabilities of various complications in the high-dimensional clinical risk vector, logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and negative feedback adjustment of historical economic load are performed in sequence, and the actual voucher settlement amount is calculated through deterministic analytical formula.

[0072] In this embodiment, step S3 may specifically include the following steps: S31 performs a log-odds transformation on the probabilities of various complications in the high-dimensional clinical risk vector, mapping the probability space to a linear measure space.

[0073] Processor Extraction The probability of various complications Run the logarithmic probability transformation: .

[0074] in: The probability of occurrence of the i-th type of complication (0≤P) i ≤1); For odds.

[0075] The natural logarithm maps probability to Linear space.

[0076] In probability space, when P approaches 0 or 1, the probability value is extremely insensitive to changes in the input (saturation region). For example, P changing from 0.01 to 0.02 only changes by 0.01, but the actual risk doubles. The Logit transformation removes this nonlinear compression, making the pricing formula uniformly sensitive to both low-to-medium risk and high-risk ranges, eliminating black-box distortion.

[0077] S32, after weighting and summing the transformed linear risk measure with risk correction coefficients, inputs it into the hyperbolic tangent boundary constraint operator, and combines it with the negative feedback adjustment factor of the historical economic load reduction cumulative value to calculate the actual voucher settlement amount.

[0078] The pricing engine retrieves the preset fixed benchmark pooling limit B and the risk adjustment coefficients for various complications. Call the piecewise hyperbolic tangent control function Perform deterministic analytical computation: .

[0079] in: The actual settlement amount on the voucher (unit: yuan); The base co-payment limit is 50,000 yuan after PCI and 80,000 yuan after colorectal cancer surgery. This is the risk correction coefficient for the i-th type of complication (initialized to 1.0, updated online with the model). For boundary constraint operators; The historical load feedback strength (0.2 in one implementation); Reduce the cumulative value of the user's historical economic load; Sigmoid is a sigmoid function: .

[0080] Boundary constraint operators Expand: .

[0081] Historical economic load reduction cumulative value (denoted as ΔE) hist This is a smoothed metric that represents the total economic relief a patient receives across all hospitalizations. Its core function is to act as a negative feedback adjustment factor, dynamically adjusting the settlement amount for the current hospitalization in the pricing formula of step S3, in order to achieve long-term equity across hospitalizations.

[0082] Its calculation method is Exponentially Weighted Moving Average (EWMA): .

[0083] Wherein, the burden reduction value ΔE(t) for each hospitalization is B - F actual (t) (base amount minus actual settlement amount) represents how much economic burden was saved for the patient during this hospitalization; λ is a decay factor (e.g., 0.9), which causes the impact of historical burden reduction to decay exponentially over time.

[0084] This is similar to a cumulative points system, but with a forgetting function. If a patient receives significant deductibles multiple times during hospitalizations (i.e., actual payments are far below the base amount), their ΔE... hist It will accumulate and rise. During the next hospitalization pricing, this high value will be affected by the Sigmoid function (ΔE). hist This generates a discount factor (1 - γ·Sigmoid(...)) to slightly increase the current settlement amount, preventing patients from continuously enjoying excessive subsidies. Conversely, if a patient has a limited history of relief payments, pricing tends to be closer to the benchmark amount. This mechanism ensures a more balanced allocation of pooled funds among different patients and different hospitalization events, and is a core regulatory parameter for achieving fairness in the system.

[0085] Logit summation: linearly adds up the risks of multiple complications, which is consistent with the clinical mechanism of multiple factors acting together.

[0086] Tanh boundary: when hour, , ;when hour, , .therefore Strictly limited to Within the specified range, even if the AI ​​model outputs abnormally or the input data drifts extremely, the single payout amount cannot exceed twice the benchmark amount, thus blocking the risk of fund pool penetration from a mathematical perspective.

[0087] Historical load feedback: If a user has historically accumulated significant economic burden reduction (ΔE) hist If the value is large, then the Sigmoid function approaches 1, and the overall multiplication term... Appropriately reduce the current pricing to achieve fair adjustment of cross-hospitalization.

[0088] S33 uses an index-weighted moving average formula to continuously update the cumulative value of historical economic load reduction, serving as the basis for adjusting the fairness of cross-hospitalization pricing.

[0089] The system retrieves the user's current cumulative economic load reduction value. Updated on a rolling basis according to the Exponentially Weighted Moving Average (EWMA) formula: .

[0090] in: This represents the cumulative burden reduction value updated after the t-th hospitalization; The historical decay factor (in one implementation, it is 0.9); That is, the actual amount of financial relief for this hospitalization (the base amount minus the actual settlement amount). New users' first hospitalization .

[0091] Then, after being smoothly scaled by Sigmoid, it is used as an elastic discount variable and substituted into the formula in step S32.

[0092] EWMA ensures that the impact of historical burden reductions decays exponentially over time, guaranteeing long-term fairness (users who receive multiple high burden reductions will have their subsequent pricing slightly increased) while preventing a single extreme event from permanently distorting pricing. This mechanism gives the system memory without excessive sensitivity, and is a form of negative feedback regulation in classical control theory.

[0093] The purpose of step S3 is to: convert the risk vector V output in step S2 into a variable value. risk The actual voucher settlement amount F is mapped to a deterministic analytical formula with boundary constraints. actual It enables precise conversion from clinical risk to economic compensation, and has built-in mathematical-level security boundaries to prevent the penetration of the fund pool.

[0094] Step S4: The actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number are encapsulated into a dual-domain encrypted fundraising voucher. The voucher is then atomically bound to the inpatient medical record number through a distributed two-phase commit protocol, thereby achieving a one-to-one mapping between a single hospitalization event and a single voucher.

[0095] In this embodiment, step S4 may specifically include the following steps: S41 encapsulates the actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of inpatient medical record number into a dual-domain encrypted fundraising voucher containing a main data field and an extended data field.

[0096] The flexible fundraising token generator generates an independent encrypted fundraising token in memory, which is divided into a main data field and an extended data field. Master data field: stores encrypted data. , and the hash fingerprint of the Case ID ; Extended data field: Stores the benchmark metric for economic burden reduction during this hospitalization. and the cumulative value of historical economic load reduction .

[0097] The hash function used is SHA-256, which ensures that the Case ID is irreversible but can be quickly compared.

[0098] The dual-domain design physically isolates core settlement data from operational optimization data. The main data domain is used for financial settlement (immutable), while the extended data domain is used for resource scheduling in step six (read and write). This separated structure ensures security while providing a data interface for the second feedback loop.

[0099] S42 uses a distributed two-phase commit protocol to atomically associate and bind fundraising vouchers with inpatient medical record numbers in the hospital information system, ensuring a one-to-one mapping between a single hospitalization event and a single voucher.

[0100] The system utilizes an atomic association module to execute a distributed transaction processing mechanism, and ensures through a two-phase commit (2PC) protocol that the fundraising token and the inpatient case ID generated within the hospital's HIS system are synchronously activated or rolled back in the globally distributed logic: Phase 1 (Preparation): The coordinator sends a ready-to-commit request to all participating nodes (Token generator, HIS interface), each node executes local transactions and locks resources, and returns ready or failed.

[0101] Phase 2 (Commit / Rollback): If all nodes return ready, the coordinator sends a global commit command, and the Token is permanently bound to the Case ID; otherwise, a global rollback is sent, the Token is invalidated, and the Case ID is restored to an unbound state; This mechanism achieves non-intrusive one-to-one binding at the underlying logic level, preventing data competition risks under high concurrency (e.g., when the same patient initiates two hospitalization fundraising requests simultaneously, the token will not be mismatched to the wrong Case ID). Compared to the loose association between annual premiums and user IDs in existing technologies, this sub-step achieves event-level atomic binding, which is a key technical guarantee for eliminating cross-subsidies.

[0102] The purpose of step S4 is to: convert the actual settlement amount F generated in step S3 into a fixed amount. actual By atomically hashing and binding the inpatient medical record number (Case ID), a dual-domain encrypted fundraising token is generated, enabling precise matching of a single hospitalization event, a single token, and a single settlement, thus completely eliminating cross-subsidies in traditional annual insurance policies.

[0103] Step S5: Asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a measure of model drift, and trigger model retraining in a safe isolation sandbox when the threshold is exceeded. The online lossless update of model parameters is completed by hot-loading the memory pointer in place, forming the first feedback loop of clinical feedback.

[0104] In this embodiment, step S5 may specifically include the following steps: S51 asynchronously monitors objective outcome events during the preset postoperative observation period, calculates the single-point residual between the actual outcome value and the preoperative predicted probability, and maintains a fixed-capacity circular buffer to calculate the residual variance as a measure of model drift.

[0105] After the patient is discharged, the event monitoring unit asynchronously monitors objective outcome events within a preset postoperative observation period T (31 days in one implementation) via a distributed message stream subscriber (such as Kafka). Upon detecting outcome confirmation (such as unplanned readmission or massive hemorrhage) or financial settlement transactions, the actual outcome status value is obtained. (0 indicates no occurrence, 1 indicates occurrence). The residual analysis unit retrieves the preoperative predicted probability from memory using the hash pointer in the token. Calculate the single-point residual: .

[0106] in: Let be the prediction residual for the i-th event, ranging from [-1, 1]. The true ending (0 or 1); This represents the preoperative prediction probability (between 0 and 1).

[0107] The system maintains a fixed-capacity, FIFO circular buffer. When a new residual is pushed in, the oldest historical data is automatically popped. A hardware multiply-accumulator (MAC) iterates through the N currently retained residual samples in the buffer and calculates the sample variance S as a monitoring statistic. .

[0108] in: The average residual; The capacity of the circular buffer (one implementation is 1000, corresponding to approximately 2 months of surgical volume).

[0109] The residual variance S quantifies the degree of fluctuation in the model's prediction error. When the clinical environment is stable, the residuals fluctuate randomly around 0, and S is small. When the introduction of new surgical procedures, new drugs, or new guidelines causes changes in the probability of basic complications, the model systematically overestimates or underestimates the probability, and S increases significantly. Compared to traditional methods that only monitor the absolute value of accuracy, variance is more sensitive to distribution drift and is not affected by class imbalance. This sub-step addresses the blind spot in existing technologies where it is impossible to know when the model fails.

[0110] S52 compares the residual variance with a preset robustness threshold. When the threshold is exceeded, model retraining is triggered in a safe isolation sandbox.

[0111] The system compares the statistic S with a preset robustness threshold θ in real time (in one implementation, θ = 0.05, obtained based on the 95th quantile of the historical residual distribution): like If the data is determined to be normal clinical noise, the desensitized data will be stored in the encrypted storage area without triggering retraining.

[0112] like If a shift in the clinical environment is detected, a retraining instruction is immediately triggered.

[0113] Retraining is performed within a fully isolated, secure sandbox (physically isolated from the production environment in terms of network and memory), calling historical encrypted feature packets (stored in a ciphertext storage area) and running the backpropagation algorithm to update the model weights. Hot reloading is then performed upon completion.

[0114] Sandbox isolation ensures that anomalies during retraining (such as gradient explosion or memory overflow) do not affect running production inference services. Hot reloading allows new parameters to take effect without restarting the service, compressing model evolution from months to hours. This mechanism solves the problem of continuously declining prediction accuracy caused by fixed model deployment until manual retraining in existing technologies.

[0115] After retraining, S53 performs in-situ hot loading of memory pointers during the production inference process via remote call instructions, enabling online lossless updates of model parameters.

[0116] After the weight parameters are updated, the sandbox retraining engine sends an RPC remote call instruction to the production inference process. The real-time inference process, without disconnecting any existing network connections, atomically stores the weight matrix address of the current network connection in the server's memory. Switch in place to the new address of the updated weight file This operation ensures atomicity through the CAS (Compare-And-Swap) instruction, and the switching process takes less than 1 microsecond.

[0117] In-situ switching avoids service interruptions in the traditional stop-update-start mode. For continuously running fundraising systems, this means that even during high-frequency surgical periods (such as 200 PCI surgeries per day), model updates will not cause any pricing requests to be blocked or timed out. Combining drift awareness in step S51 and sandbox retraining in step S52 forms a complete automated closed loop of awareness-computation-switching.

[0118] The purpose of step S5 is to asynchronously monitor objective outcome events during the postoperative observation period, calculate the predicted residual variance as a measure of model drift, trigger model retraining and hot loading in the safe isolation sandbox, and realize that each postoperative outcome is a dense feedback of training signals, thereby compressing the model evolution cycle from several months to hours.

[0119] Step S6: Read the economic burden reduction metric from the fundraising certificate, calculate the service routing priority weight parameter, and dynamically adjust the service quality level of the distributed message queue, trigger the asynchronous cache preloading of distributed storage, and dynamically allocate the quota of container computing resources accordingly, forming the second feedback loop of operation resource scheduling.

[0120] In this embodiment, step S6 may specifically include the following steps: S61 reads the economic burden reduction benchmark metric from the extended data field of the financing certificate and calculates the service routing priority weight parameter through a nonlinear mapping function.

[0121] While the claims settlement is asynchronous, the resource optimization and scheduling module captures the benchmark value of economic burden reduction for patients due to settlement using encrypted fundraising vouchers in real time through the internal data bus. This metric is used as a continuous control variable, and a preset nonlinear mapping function for resource routing allocation is input: .

[0122] in: For service routing priority weight parameters (dimensionless, ≥0); The scaling factor (in one implementation, it is set to 0.02 yuan) is the scale scaling factor. -1 ); It is a very small positive number (to prevent the weight from being zero); ReLU is a linear rectified function: .

[0123] β (scale scaling factor, taken as 0.02) -1 The value of ) is based on: Map the economic burden reduction amount ΔE (in yuan) to the service routing weight W. route The reasonable range of values ​​for W. route This is then used to drive resource scheduling (such as queue priority and computing power quota), and its value should match the level of business importance.

[0124] Assuming a typical high-cost-reduction scenario is ΔE = 30,000 yuan (e.g., significant cost savings from complex surgery), then β × ΔE = 0.02 × 30000 = 600. This value (600) is sufficient in a resource scheduling system to be transformed into a significantly higher priority level than ordinary users (with a default weight of 1) through logarithmic or linear mapping, thus creating a difference in service quality. If β is too large (e.g., 0.1), the weight will balloon (3000), potentially leading to system resources being monopolized by an extremely small number of users; if β is too small (e.g., 0.001), even users with high cost reduction will only have a weight of about 30, insufficient differentiation, and unable to effectively trigger resource preferences. Therefore, 0.02 is an empirical value based on a balance between typical business volume (cost reduction of tens of thousands of yuan) and system scheduling granularity (weight multiples in the hundreds).

[0125] The basis for the value of (the smallest positive number): Ensure W route It is greater than zero in all cases. Because W route If a value of 0 is used as a divisor or weighting factor in subsequent queue priority and load balancing calculations, it may lead to anomalies such as being divided by zero or having invalid weights.

[0126] Typically, a positive number much smaller than the minimum effective value of the normal weight is chosen, such as 1e. -6 or 1e -8 Its sole function is as a mathematical safety net, with negligible impact on actual scheduling results (contributing only a tiny amount of base weight when ΔE=0). Its specific value is determined by the system's numerical precision tolerance, and should be as small as possible while ensuring no floating-point underflow occurs, to guarantee that the weight distribution of normal business logic is not affected.

[0127] β and The value of is an engineering choice made in the system design after comprehensive consideration of business volume, scheduling granularity and numerical stability. It can be calibrated through simulation in the early stage of system operation and remain stable in long-term operation.

[0128] Step S61 maps the economic burden reduction (in yuan) to a technology priority (dimensionless). For example, if the actual burden reduction for a hospitalization is ΔE = 30,000 yuan, then... ≈0.02 × 30000 = 600, meaning this user will receive significantly higher service rewards than ordinary users (W). route The resource priority is approximately 1). This establishes an economic incentive mechanism where the greater the contribution to the mutual aid, the faster the service response.

[0129] S62 dynamically adjusts the service quality level and consumption priority in the distributed message queue based on the service routing priority weight parameter.

[0130] Real-time extraction of network transport layer When the client initiates a remote follow-up data interaction request or online follow-up audio / video stream signaling interaction with the server, the microservice gateway automatically... The proportion of this adjustment increases the QoS (Quality of Service) level and consumption priority of the interaction request in the distributed message queue (such as Kafka / RabbitMQ). Specifically, this is achieved by setting the queue's `consumer_priority` field to a specific value. This ensures that requests from users with high W_route are consumed first.

[0131] In high-concurrency scenarios (such as peak evening follow-up visits), messages from ordinary users may queue for several seconds or even tens of seconds, while medical signaling from high-W_route users (such as real-time vital sign uploads and video consultations) can be prioritized. Experimental data shows that this method reduces the average waiting time for remote follow-up requests from high-cost users in the Kafka queue from 1200ms to 80ms, significantly improving user experience and medical security.

[0132] S63, based on the service routing priority weight parameter, triggers asynchronous cache preloading of distributed storage and dynamic quota adjustment of container computing resources.

[0133] In practice, step S63 further includes the step of parallel operation: S631 performs distributed storage pre-loading (Cache Pre-warm).

[0134] The storage management module monitored Upon changes, an asynchronous preheating command is automatically sent to the storage resource layer. The system asynchronously preloads the user's historical chronic disease prescription details and drug-specific pharmacokinetic monitoring dataset, which are likely to be accessed by the user subsequently, from the underlying cold storage database (such as object storage step S3) to the Redis in-memory distributed cache cluster. The preloading is triggered by the following conditions: The preloaded data range includes chronic disease prescription records and drug concentration monitoring curves from the most recent 6 months.

[0135] S632 performs fine-tuning of the container computing power allocation algorithm.

[0136] The system will The parameters are injected into the load balancing algorithm of the microservice gateway (such as the Weighted Round-Robin algorithm) to dynamically increase the container computing resources (CPU / memory quota) allocated to the user's dedicated continuous health tracking process. Specifically, the container orchestration platform (Kubernetes) adjusts the cpu_limit and mem_limit parameters of the Pod to give downstream service instances of users with high W_route a higher resource allocation. In one implementation, when W_route=600, the CPU limit is increased from 0.5 cores to 2.0 cores.

[0137] Cache preloading reduces the I / O response time of the underlying system data from seconds (reading from object storage) to milliseconds (reading from Redis) when a user makes a subsequent request for continuous health services. Container computing power adjustment ensures that high-priority users' data analysis tasks (such as real-time ECG waveform analysis) are not delayed due to resource contention. Together, these measures achieve a positive feedback loop of economic reduction → technological advantages.

[0138] The purpose of step S6 is to read the economic burden reduction metric ΔE from the Token extended data domain, dynamically adjust service routing priority, message queue QoS, cache preloading and container computing power allocation, and transform the economic effect into technical parameters to achieve a positive flywheel effect where the more burden is reduced, the better the service.

[0139] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0140] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0141] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0142] Example 2 Further reference Figure 2 As a response to the above Figure 1 The present invention provides an embodiment of a disease-specific risk financing optimization device based on a two-layer value feedback loop, which is similar to the implementation of the method shown. Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0143] like Figure 2As shown, the disease-specific risk financing optimization device 70 based on a two-layer value feedback loop described in this embodiment includes: a construction module 71, a combination module 72, an execution module 73, an encapsulation module 74, a calculation module 75, and an adjustment module 76. Wherein: Module 71 is used to collect patients' time-series test indicators, unstructured electronic medical record text and discrete static medical history in real time. Through time-series derivative feature extraction, semantic embedding and feature fusion, a global benchmark feature tensor is constructed. Combination module 72 is used to input the global baseline feature tensor into a deep neural network, fuse temporal features and baseline features through a cross-attention mechanism, infer the objective probability of occurrence of multiple complications through an independent Sigmoid output layer, and combine them to form a high-dimensional clinical risk vector. Execution module 73 is used to sequentially perform logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and historical economic load negative feedback adjustment based on the probability of each complication in the high-dimensional clinical risk vector, and calculate the actual voucher settlement amount through a deterministic analytical formula; Encapsulation module 74 is used to encapsulate the actual voucher settlement amount, high-dimensional clinical risk vector and hash fingerprint of inpatient medical record number into a dual-domain encrypted fundraising voucher, and atomically bind it with the inpatient medical record number through a distributed two-phase commit protocol to realize a one-to-one mapping between a single hospitalization event and a single voucher. The calculation module 75 is used to asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a model drift metric, trigger model retraining in a safe isolation sandbox when the threshold is exceeded, and complete the online lossless update of model parameters through in-situ hot loading of memory pointers, forming the first feedback loop of clinical feedback. The adjustment module 76 is used to read the economic burden reduction metric value from the financing certificate, calculate the service routing priority weight parameter, and dynamically adjust the service quality level of the distributed message queue, trigger the asynchronous cache preloading of distributed storage, and dynamically allocate the quota of container computing resources, forming a second feedback loop for operation resource scheduling.

[0144] Example 3 To address the aforementioned technical problems, embodiments of the present invention also provide a computer device. Please refer to [link / reference needed]. Figure 3 , Figure 3 This is a basic structural block diagram of the computer device in this embodiment.

[0145] The aforementioned computer device 8 includes a memory 81, a processor 82, and a network interface 83 that are interconnected via a system bus. It should be noted that only the computer device 8 with components 81, 82, and 83 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0146] The aforementioned computer devices can be desktop computers, laptops, handheld computers, and cloud servers, among other computing devices. These devices can facilitate human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0147] The aforementioned memory 81 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the aforementioned memory 81 may be an internal storage unit of the aforementioned computer device 8, such as the hard disk or memory of the computer device 8. In other embodiments, the aforementioned memory 81 may also be an external storage device of the aforementioned computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 8. Of course, the aforementioned memory 81 may also include both the internal storage unit and its external storage device of the aforementioned computer device 8. In this embodiment, the aforementioned memory 81 is typically used to store the operating system and various application software installed on the aforementioned computer device 8, such as computer-readable instructions for a disease-specific risk financing optimization method based on a two-layer value feedback loop. In addition, the aforementioned memory 81 can also be used to temporarily store various types of data that have been output or will be output.

[0148] In some embodiments, the processor 82 described above may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 82 is typically used to control the overall operation of the computer device 8. In this embodiment, the processor 82 is used to execute computer-readable instructions stored in the memory 81 or to process data, for example, to execute the computer-readable instructions of the disease-specific risk financing optimization method based on a two-layer value feedback loop.

[0149] The network interface 83 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 8 and other electronic devices.

[0150] Example 4 The present invention also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the above-described method for optimizing disease risk financing based on a two-layer value feedback loop.

[0151] The beneficial effects of implementing the above embodiments are as follows: (1) Event-level precise pricing can be achieved to eliminate cross-subsidization: By compressing the pricing granularity to a single hospitalization event and atomically binding the Token with the hospitalization medical record number, precise mutual assistance can be achieved.

[0152] (2) It can be driven by core clinical data and has more sensitive risk identification: By directly accessing real-time clinical data from HIS / LIS / PACS systems, it can extract time-series first / second derivative features and provide early warning of acute risks 24-48 hours in advance.

[0153] (3) Model self-evolution driven by postoperative outcome: Through residual variance monitoring and sandbox hot loading, the model evolution cycle is compressed from several months to hours.

[0154] (4) Mathematical architecture-level risk control: By embedding the Tanh boundary constraint operator in the pricing formula, the settlement amount is always bounded, preventing the penetration of the fund pool from the underlying mathematical level.

[0155] (5) It has a positive flywheel effect of economy and technology: the economic burden reduction measure is directly mapped to resource priority through the second feedback loop, forming an adaptive incentive mechanism in which the more burden is reduced, the better the service is.

[0156] (6) The entire process is explainable and auditable: Combining Grad-CAM attribution and Logit linear pricing, the risk basis of each voucher can be traced, meeting regulatory compliance requirements.

[0157] (7) It can be widely applied to medical financing scenarios such as commercial health insurance, inclusive insurance, and special disease pool, and has significant technological progress and commercial value.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0159] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A disease-specific risk financing optimization method based on a two-layer value feedback loop, characterized in that, Includes the following steps: Real-time collection of patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history; and construction of a global benchmark feature tensor through time-series derivative feature extraction, semantic embedding, and feature fusion. The global baseline feature tensor is input into a deep neural network, and the temporal features and baseline features are fused through a cross-attention mechanism. The objective probability of occurrence of multiple complications is inferred through an independent Sigmoid output layer and combined to form a high-dimensional clinical risk vector. Based on the probabilities of various complications in the high-dimensional clinical risk vector, logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and negative feedback adjustment of historical economic load are performed in sequence, and the actual voucher settlement amount is calculated through deterministic analytical formulas. The actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number are encapsulated into a dual-domain encrypted fundraising voucher, and atomically bound to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher. Asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a measure of model drift, and trigger model retraining in a safe isolation sandbox when the threshold is exceeded. The online lossless update of model parameters is completed by hot loading the memory pointer in place, forming the first feedback loop of clinical feedback. The economic burden reduction metric is read from the financing certificate, the service routing priority weight parameter is calculated, and the service quality level of the distributed message queue is dynamically adjusted accordingly. The asynchronous cache preloading of distributed storage and the dynamic quota allocation of container computing resources are triggered to form a second feedback loop for operation resource scheduling.

2. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to claim 1, characterized in that, The steps for constructing a global baseline feature tensor by real-time acquisition of patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history, followed by time-series derivative feature extraction, semantic embedding, and feature fusion, specifically include: First and second derivatives were extracted and standardized from the time series of continuous clinical laboratory indicators to construct a time-series evolution feature sub-tensor. Using a pre-trained medical language model, low-dimensional semantic embedding is performed on unstructured electronic medical record text to generate semantic feature vectors; The temporal evolution feature sub-tensor, semantic feature vector, and discrete static medical history features are concatenated at the channel level and encapsulated to form a global baseline feature tensor.

3. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to claim 1, characterized in that, The steps of inputting the global baseline feature tensor into a deep neural network, fusing temporal features and baseline features through a cross-attention mechanism, inferring the objective probability of occurrence of multiple complications through an independent Sigmoid output layer, and combining them to form a high-dimensional clinical risk vector specifically include: The temporal features in the global baseline feature tensor are fused with the baseline features through a cross-attention mechanism, and the objective probability of occurrence of multiple complications is inferred through an independent Sigmoid output layer, and combined to form a high-dimensional clinical risk vector. Based on the gradient-weighted class activation mapping algorithm, the contribution of each complication risk prediction score to the internal feature map of the network is calculated, and risk interpretation anchor points are generated on the informed consent form interface.

4. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to claim 1, characterized in that, The steps for calculating the actual voucher settlement amount using a deterministic analytical formula, based on the probabilities of various complications in the high-dimensional clinical risk vector, sequentially performing logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and negative feedback adjustment of historical economic load, specifically include: Log-odds transformation is performed on the probabilities of each complication in the high-dimensional clinical risk vector to map the probability space to a linear measure space; The transformed linear risk measure is weighted and summed by risk correction coefficients, then input into the hyperbolic tangent boundary constraint operator, and combined with the negative feedback adjustment factor of the historical economic load reduction cumulative value, to calculate the actual voucher settlement amount. The historical economic load reduction cumulative value is updated on a rolling basis using an index-weighted moving average formula, serving as the basis for adjusting the fairness of cross-hospitalization pricing.

5. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to claim 1, characterized in that, The steps of encapsulating the actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number into a dual-domain encrypted fundraising voucher, and atomically binding it to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher specifically include: The actual voucher settlement amount, high-dimensional clinical risk vector, and hash fingerprint of the inpatient medical record number are encapsulated into a dual-domain encrypted fundraising voucher containing a main data field and an extended data field. By using a distributed two-phase commit protocol, the fundraising voucher is atomically associated and bound to the inpatient medical record number in the hospital information system, ensuring a one-to-one mapping between a single hospitalization event and a single voucher.

6. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to claim 1, characterized in that, The steps of asynchronously monitoring objective outcome events during the preset postoperative observation period, calculating the residual variance between the actual outcome and the predicted probability as a model drift metric, triggering model retraining in a safe isolation sandbox when the threshold is exceeded, and completing online lossless updating of model parameters through in-situ hot loading of memory pointers to form the first feedback loop of clinical feedback specifically include: Asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the single-point residual between the actual outcome value and the preoperative predicted probability, and maintain a fixed-capacity circular buffer to calculate the residual variance as a measure of model drift. The residual variance is compared with a preset robustness threshold. When the threshold is exceeded, model retraining is triggered in a safe isolation sandbox. After retraining, the memory pointer is hot-loaded in place during the production inference process via remote invocation, enabling online lossless updates of model parameters.

7. The disease-specific risk financing optimization method based on a two-layer value feedback loop according to any one of claims 1 to 6, characterized in that, The steps of reading the economic burden reduction metric from the financing certificate, calculating the service routing priority weight parameter, and dynamically adjusting the service quality level of the distributed message queue, triggering asynchronous cache preloading of distributed storage, and dynamically allocating container computing resources to form a second feedback loop for operational resource scheduling specifically include: Read the economic burden reduction benchmark value from the extended data field of the financing certificate, and calculate the service routing priority weight parameter through a nonlinear mapping function; Based on the service routing priority weight parameters, the service quality level and consumption priority in the distributed message queue are dynamically adjusted. Based on the service routing priority weight parameters, the asynchronous cache preloading of distributed storage and the dynamic quota adjustment of container computing resources are triggered.

8. A disease-specific risk financing optimization device based on a two-layer value feedback loop, characterized in that, include: The module is used to collect patients' time-series test indicators, unstructured electronic medical record text, and discrete static medical history in real time. Through time-series derivative feature extraction, semantic embedding, and feature fusion, a global benchmark feature tensor is constructed. The combination module is used to input the global baseline feature tensor into a deep neural network, fuse temporal features and baseline features through a cross-attention mechanism, infer the objective probability of occurrence of multiple complications through an independent Sigmoid output layer, and combine them to form a high-dimensional clinical risk vector. The execution module is used to sequentially perform logarithmic odds transformation, hyperbolic tangent mapping with boundary constraints, and historical economic load negative feedback adjustment based on the probability of each complication in the high-dimensional clinical risk vector, and calculate the actual voucher settlement amount through a deterministic analytical formula. The encapsulation module is used to encapsulate the actual voucher settlement amount, high-dimensional clinical risk vector and hash fingerprint of inpatient medical record number into a dual-domain encrypted fundraising voucher, and atomically bind it to the inpatient medical record number through a distributed two-phase commit protocol to achieve a one-to-one mapping between a single hospitalization event and a single voucher. The calculation module is used to asynchronously monitor objective outcome events during the preset postoperative observation period, calculate the residual variance between the actual outcome and the predicted probability as a measure of model drift, and trigger model retraining in a safe isolation sandbox when the threshold is exceeded. The model parameters are updated online without loss through in-situ hot loading of memory pointers, forming the first feedback loop of clinical feedback. The adjustment module is used to read the economic burden reduction metric value from the financing certificate, calculate the service routing priority weight parameter, and dynamically adjust the service quality level of the distributed message queue, trigger the asynchronous cache preloading of distributed storage, and dynamically allocate the quota of container computing resources, forming a second feedback loop for operation resource scheduling.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the disease-specific risk financing optimization method based on a two-layer value feedback loop as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the disease-specific risk financing optimization method based on a two-layer value feedback loop as described in any one of claims 1 to 7.

Images