A multi-dimensional fusion-based cost optimization method and system for DRG-DIP

By synchronously accessing and standardizing multi-source heterogeneous medical data in time and space, constructing a multi-dimensional feature space and dynamically weighting and fusing it, and combining hybrid incremental prediction and multi-objective optimization, the problems of grouping errors and real-time response in DRG payment management are solved, and an optimization scheme that balances the interests of all parties is achieved.

CN123117450APending Publication Date: 2026-11-17SHANGHAI TONGREN HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611339669.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies fail to effectively integrate multi-dimensional features in DRG payment management, which makes optimization schemes prone to grouping errors, under- or over-grouping, and the predictive model has insufficient generalization ability, failing to meet the needs of real-time clinical response. The optimization logic is rigid and has a single objective, making it difficult to achieve a balance of interests among multiple parties.

Method used

By synchronously accessing and standardizing multi-source heterogeneous medical data in time and space, a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space are constructed. A hybrid incremental prediction model and dynamic constraint multi-objective optimization are adopted to balance medical compliance, DRG cost-effectiveness, and enrollment stability.

Benefits of technology

It enables real-time guidance for DRG payment management, reduces the risk of overspending and violations, and improves the level of refined DRG payment management and overall operational efficiency of medical institutions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117450A_ABST
    Figure CN123117450A_ABST
Patent Text Reader

Abstract

This invention provides a multi-dimensional fusion-based DRG-DIP cost optimization method. The method first performs spatiotemporal synchronous access and standardization processing of multi-source heterogeneous medical data to construct a real-time updated DRG rule vector library. Then, it constructs a three-dimensional feature space encompassing medical compliance, DRG benefits, and enrollment risk, and dynamically weights and fuses these features to obtain a fusion feature tensor. A hybrid incremental prediction model is used to complete DRG grouping and cost inference. Multi-objective Pareto optimization is performed based on dynamic constraint relaxation to optimize candidate treatment items temporally and smooth costs, outputting an execution plan. The execution plan is then validated and optimized for compliance using DIP billing rules, resulting in an optimized treatment execution plan. This invention achieves real-time response at the doctor's order level, balances the interests of all three parties, and reduces the risk of overspending and violations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical big data analysis technology, and in particular to a DRG-DIP cost optimization method and system based on multi-dimensional fusion. Background Technology

[0002] With the nationwide implementation of DRG (Diagnosis Related Groups) payment reform, medical insurance settlement has shifted from fee-for-service to bundled payment based on disease groups. DRG grouping rules rely on multiple dimensions such as primary diagnosis, surgical procedure, comorbidities and complications, age, and length of hospital stay, resulting in a large number of combinations and complex rules.

[0003] Currently, existing technical solutions for DRG payment management are mainly divided into two categories: one is a basic information management system, which mainly realizes functions such as DRG group query, cost statistics, and settlement report generation. The data relies on manual input and can only perform post-event statistical analysis, and cannot provide real-time guidance during the diagnosis and treatment process; the other is a preliminary application solution that introduces artificial intelligence technology. By building simple machine learning models or rule engines, it can realize DRG group prediction and total cost prediction. Some systems attempt to filter optimization items based on fixed rules. However, existing technologies still have the following shortcomings: 1. Most solutions only integrate the two dimensions of medical compliance and cost-effectiveness, without quantifying the impact of treatment items on enrollment risks, which makes it easy for optimization solutions to cause medical insurance violations such as incorrect grouping, under- or over-grouping; 2. The predictive model has insufficient generalization ability and real-time performance, and the incremental calculation mechanism is crude. Each new medical order requires a full recalculation, which cannot meet the needs of real-time clinical response; 3. The optimization logic is rigid and has a single objective. It uses fixed constraints and does not dynamically adjust the strictness of constraints according to the severity of the patient's condition, resulting in poor clinical feasibility of the optimization solution. It ignores the saving of medical insurance funds and the control of patients' out-of-pocket expenses, and it is difficult to achieve a balance of interests among all parties. Summary of the Invention

[0004] The purpose of this invention is to provide a DRG-DIP cost optimization method and system based on multi-dimensional fusion. It balances medical compliance, DRG cost-effectiveness and enrollment stability through dynamic weighted fusion of three-dimensional features, hybrid incremental prediction and dynamic constraint multi-objective optimization.

[0005] To achieve the above objectives, the present invention provides the following solution: A cost optimization method for DRG-DIP based on multi-dimensional fusion includes the following steps: Multi-source heterogeneous medical data is spatiotemporally synchronized and standardized to obtain standardized data; the standardized data includes: patient basic feature matrix, time series sequence of diagnosis and treatment items, and cost consumption trajectory curve; A rule vector library is constructed based on DRG grouping rules and medical insurance compliance rules; Based on the rule vector library and standardized data, a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space are constructed and dynamically weighted and fused to obtain a fused feature tensor. The fusion feature tensor is inferred based on the hybrid incremental prediction model to obtain the inference results. The inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability. A multi-objective optimization search is performed on the inference results based on the dynamic constraint relaxation to obtain a set of candidate treatment item combinations; The candidate treatment item combination set is optimized in time and smoothed in cost to obtain the treatment item time sequence execution plan; By using DIP billing rules to perform compliance checks and optimizations on the timing execution plan of medical services, an optimized medical service execution plan is obtained.

[0006] Optionally, multi-source heterogeneous medical data can be spatiotemporally synchronized and standardized to obtain standardized data, including: Patient data is extracted in real time from HIS, EMR, LIS, PACS and medical insurance settlement systems via standard HL7 interface and API adapter; patient data includes: ICD diagnosis code, surgical procedure code, age, gender, comorbidities and complications, performed medical services, incurred costs, laboratory test results and imaging report summaries; By standardizing and converting ICD diagnostic codes, surgical operation codes, and treatment item codes using a unified coding standard, a patient basic characteristic matrix is ​​obtained. Arrange the treatment items in chronological order of execution to obtain the treatment item time sequence; Using calendar days as the statistical unit, calculate the daily cumulative cost and the percentage of each itemized cost to obtain the cost consumption trajectory curve.

[0007] Optionally, a rule vector library is constructed based on DRG grouping rules and medical insurance compliance rules, including: We collect DRG payment policy documents, adjustment notices, interpretive documents, and violation case databases published on the official websites of the National Healthcare Security Administration and provincial healthcare security bureaus. The DRG payment policy document was semantically segmented using the BERT model to identify key information, including: DRG grouping rules, payment standards, list of restricted medical insurance items, definition of violations and penalty standards. Key information is converted into rule vectors, and a rule vector library is generated. The rule vector includes: rule ID, rule type, applicable disease group scope, effective time, constraints, execution action, and confidence level.

[0008] Optionally, the medical compliance feature space includes: the matching degree of necessary examination items, the matching degree of necessary treatment items, the compliance degree of clinical guidelines, the level of medical insurance restrictions, the level of medical safety risks, and the completion degree of informed consent; the DRG benefit feature space includes: the contribution of treatment items to costs, marginal cost, surplus contribution value, potential for improvement of disease group weight, medical insurance payment ratio, and cost recovery period; the enrollment risk feature space includes: the impact of diagnosis on grouping, the impact of surgical procedures on grouping, the impact of comorbidities and complications on CC / MCC determination, the probability of grouping error, the probability of low grouping risk, and the probability of high grouping risk.

[0009] Optionally, a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space are constructed based on a rule vector library and standardized data, and then dynamically weighted and fused to obtain a fused feature tensor, including: Adaptive weights were calculated for the medical compliance feature space, DRG benefit feature space, and enrollment risk feature space, respectively. The feature vectors of the three dimensions are weighted and fused according to adaptive weights to obtain the fused feature tensor.

[0010] Optionally, the adaptive weights are calculated using the following formula: ;in, Let be the weight of the i-th dimension. Let be the historical average error rate of the i-th dimension. is the Laplace smoothing coefficient.

[0011] Optionally, inference is performed on the fused feature tensor based on the hybrid incremental prediction model to obtain the inference results, including: The fusion feature tensor is separated into patient static feature sub-tensor and diagnosis and treatment time-series dynamic feature sub-tensor; The patient's static feature subtensor is input into a multilayer perceptron to obtain the DRG group's basic probability distribution and the baseline cost value for each disease group. By inputting the dynamic feature subtensor of the diagnosis and treatment time series into the Transformer model, the predicted values ​​and trend curves of drug costs, examination costs, treatment costs, nursing costs and medical consumable costs for the next 3 days are obtained respectively. Based on the rule vector library, the basic probability distribution of DRG grouping is corrected by hard constraints, and the predicted values ​​are pruned for compliance, resulting in the corrected probability distribution and the predicted values ​​of individual expenses. The total cost forecast is obtained by weighting and summing the predicted values ​​of each item of cost according to the preset weights. The disease group with the highest probability in the modified probability distribution was selected as the DRG prediction group, and the group stability score and the probability of inclusion risk were calculated. The uncertainty of the total cost forecast is quantified using the Monte Carlo method, and the surplus probability is calculated in conjunction with the DRG payment standard.

[0012] Optionally, a multi-objective optimization search is performed on the inference results based on the dynamic constraint relaxation to obtain a set of candidate treatment item combinations, including: Determine the type of constraint; constraint types include: hard constraints and soft constraints; hard constraints include: minimum medical safety requirements, medical insurance compliance restrictions, and DRG enrollment requirements; soft constraints include: DRG payment standard caps and average length of stay limits. Construct a multi-objective optimization function based on the strictness of the constraint types; The Pareto optimal solution search for the multi-objective optimization function is performed using the NSGA-III algorithm to obtain a set of candidate treatment item combinations.

[0013] Optionally, the candidate treatment item combination set is subjected to time-series optimization and cost smoothing to obtain a treatment item time-series execution plan, including: Extract project characteristics from the candidate treatment project portfolio set; project characteristics include: earliest executable time, latest required executable time, execution duration, cost per session, and prerequisite dependent projects; Construct a dependency graph of diagnosis and treatment items based on project characteristics; Based on the dependency graph of treatment items, and with the objective function of minimizing the daily cost variance during the hospitalization period, high-cost items are evenly distributed over the remaining hospitalization days to generate a time-series execution plan. The time-series execution plan includes: the suggested execution date, time period, and estimated cost for each item. Calculate the cost fluctuation coefficient for each time-series execution plan, and select the plan with the smallest cost fluctuation coefficient as the time-series execution plan for the diagnosis and treatment project.

[0014] A DRG-DIP cost optimization system based on multi-dimensional fusion includes: The data processing module is used to perform spatiotemporal synchronous access and standardization processing of multi-source heterogeneous medical data to obtain standardized data. The standardized data includes: patient basic feature matrix, time series of diagnosis and treatment items, and cost consumption trajectory curve. The rule building module is used to build a rule vector library based on DRG grouping rules and medical insurance compliance rules; The feature fusion module is used to construct a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space based on the rule vector library and standardized data, and then dynamically weight and fuse them to obtain a fused feature tensor. The cost inference module is used to infer the fused feature tensor based on the hybrid incremental prediction model and obtain the inference results. The inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability. The project portfolio module is used to perform multi-objective optimization search on the inference results based on dynamic constraint relaxation to obtain a set of candidate treatment project portfolios. The optimization module is used to perform time-series optimization and cost smoothing on the candidate treatment item combination set to obtain the treatment item time-series execution plan. The DIP fusion module is used to perform compliance verification and optimization of the timing execution plan of diagnosis and treatment items through DIP billing rules, so as to obtain an optimized diagnosis and treatment execution plan.

[0015] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a multi-dimensional fusion-based DRG-DIP cost optimization method and system. The method includes: spatiotemporally and synchronously accessing and standardizing multi-source heterogeneous medical data to obtain standardized data; constructing a rule vector library based on DRG grouping rules and medical insurance compliance rules; constructing a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space based on the rule vector library and standardized data, and dynamically weighting and fusing them to obtain a fusion feature tensor; inferring the fusion feature tensor based on a hybrid incremental prediction model to obtain inference results; the inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score, and enrollment risk probability; performing multi-objective optimization search on the inference results based on dynamic constraint relaxation to obtain a candidate treatment item combination set; performing time-series optimization and cost smoothing processing on the candidate treatment item combination set to obtain a treatment item time-series execution plan; and performing compliance verification and optimization on the treatment item time-series execution plan through DIP billing rules to obtain an optimized treatment execution plan. This method balances the interests of all three parties, reduces the risk of overspending and violations, and improves the level of refined management of DRG payment and the overall operational efficiency of medical institutions by processing multi-source heterogeneous medical data in a time and space, dynamically weighting and fusing three-dimensional features, and combining hybrid incremental prediction with dynamic constraint multi-objective optimization. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0017] Figure 1 This is a flowchart of the DRG-DIP cost optimization method based on multi-dimensional fusion of the present invention; Figure 2 This is a schematic diagram of the DRG-DIP cost optimization system structure based on multi-dimensional fusion of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this invention provides a DRG-DIP cost optimization method based on multi-dimensional fusion, comprising the following steps: Step 100: Perform spatiotemporal synchronous access and standardization processing on multi-source heterogeneous medical data to obtain standardized data; the standardized data includes: patient basic feature matrix, time series sequence of diagnosis and treatment items, and cost consumption trajectory curve; Step 200: Construct a rule vector library based on DRG grouping rules and medical insurance compliance rules; Step 300: Construct a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space based on the rule vector library and standardized data, and perform dynamic weighted fusion to obtain a fused feature tensor; Step 400: Perform inference on the fused feature tensor based on the hybrid incremental prediction model to obtain the inference results; the inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability; Step 500: Perform multi-objective optimization search on the inference results based on dynamic constraint relaxation to obtain a set of candidate treatment item combinations; Step 600: Perform time-series optimization and cost smoothing on the candidate treatment item combination set to obtain the treatment item time-series execution plan; Step 700: Perform compliance verification and optimization of the treatment item timing execution plan through DIP billing rules to obtain an optimized treatment execution plan.

[0021] In the specific implementation process, step 100 first deploys a message middleware and RESTful API adapter cluster based on the HL7v2.5 / v3.0 standard, connecting to the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Medical Image Archiving and Communication System (PACS), and Medical Insurance Settlement System, respectively. NTP network time protocol is used to perform global time synchronization for all connected systems. A 100ms incremental extraction polling interval is set for real-time diagnosis and treatment data, and a batch incremental synchronization mode is used for historical data from the 7 days prior to admission. A global data association index is constructed using the patient's unique hospital number as the primary key. Subsequently, multi-stage data cleaning is performed for the patient... For missing values ​​in required fields such as age and gender in the patient's basic information, the mode of patients of the same age group in the same DRG disease group was used for filling. For continuous missing values ​​in test results, K-nearest neighbor interpolation with K=5 was used for filling. For outliers, the 3σ principle was used to remove test data that exceeded the normal physiological range and to mark abnormal records. At the same time, unstructured redundant text in the image report summary was filtered by regular expressions to extract key positive indicators. Then, the coding standardization conversion was performed. The coding mapping dictionary was constructed based on the ICD-10 diagnostic code, ICD-9-CM-3 surgical operation code and the "National Medical Service Item Classification and Code" uniformly issued by the National Healthcare Security Administration. String fuzzy matching combined with BER was used. The heterogeneous coding was automatically converted using T-semantic similarity calculation, with a semantic similarity threshold set at 0.85. Next, a patient baseline feature matrix was constructed. Standardized demographic features were converted into one-hot encoded vectors, and diagnostic, surgical, and comorbidity / complication codes were converted into embedded vectors. The results of 24 core laboratory tests were min-max normalized. Finally, a patient baseline feature matrix with dimensions [1, 128] was formed, where the first 16 dimensions are demographic features, the middle 64 dimensions are diagnostic and surgical features, and the last 48 dimensions are core laboratory test indicators. A time-series sequence of treatment items was then generated, with the patient's admission time as the zero point, and the actual execution timestamps of each treatment item added. The data is sorted in ascending order, and redundant items that are executed repeatedly and have no clinical indication are removed. Each treatment item is labeled with execution duration, executing department, operator, and cost attribute tags to form a structured time-series sequence containing timestamp, item code, item name, execution attribute, and single-session cost. The sequence length is dynamically adapted to the patient's hospital stay. Finally, a cost consumption trajectory curve is generated, using the natural day as the statistical granularity. The daily drug cost, examination cost, treatment cost, nursing cost, medical consumable cost, and other costs are calculated from the patient's hospital stay day 1 to the current hospital stay day. The daily cumulative total cost and the proportion of each item cost are calculated, and the main trajectory curve and the stacked curve of the proportion of each item cost are plotted with the hospital stay day as the horizontal axis and the cumulative cost as the vertical axis.

[0022] In the specific implementation process, step 200 deploys a distributed crawler cluster based on the Scrapy framework. Targeted crawling tasks are set up for the official websites of the National Healthcare Security Administration and provincial healthcare security bureaus, official government new media platforms, and the healthcare security inspection notification column. A 3-minute real-time monitoring trigger mechanism is set up for policy release sections. The system automatically crawls DRG payment policy documents, grouping rule adjustment notices, healthcare insurance catalog update announcements, policy interpretation documents, and a database of violation penalty cases. PyPDF2, python-docx, and BeautifulSoup are used to parse the content of PDF, Word, and HTML documents respectively, extracting plain text and performing MD5 hashing. The system performs global deduplication, classifying and storing the deduplicated text by publication time, publication level, and document type. Then, it loads a BERT semantic segmentation model pre-trained on a Chinese medical corpus and fine-tuned with annotated DRG policy documents. First, a sentence-level binary classifier filters out irrelevant administrative notices, meeting minutes, etc. Then, a nested named entity recognition module extracts core entities such as DRG grouping rules, payment standards, medical insurance restricted items list, definition of violations, and penalty standards. Simultaneously, it extracts attribute information such as the applicable disease group scope, effective time, expiration time, constraints, and exceptions of the rules. Finally, a conflict resolution algorithm based on rule priority is used to process different layers. For rule conflicts at different levels and times, the priority order is: new local regulations > new national regulations > old local regulations > old national regulations, with rules at the same level ordered in reverse chronological order. Next, the extracted key information is standardized and converted into a 7-dimensional rule vector. The rule ID uses the snowflake algorithm to generate a globally unique 19-bit numerical identifier. The rule type is converted into a 4-dimensional one-hot encoding. The applicable disease group range is mapped to a 128-dimensional DRG disease group encoding embedding vector consistent with step 100. The effective and expiration times are converted into Unix timestamps. Constraints are converted into a Boolean logic abstract syntax tree using the ANTLR4 parser. Execution actions are converted into 4-dimensional one-hot encodings. Confidence levels are determined based on the publishing entity and... Document types are dynamically assigned values. In this embodiment, official documents from the National Healthcare Security Administration are assigned a value of 1.0, official documents from provincial governments are assigned a value of 0.95, official documents from municipal governments are assigned a value of 0.9, official interpretations are assigned a value of 0.8, and derivations of violation cases are assigned a value of 0.7. Finally, the Neo4j graph database is used to store rule vectors and the relationships between rules, and each rule is assigned an incremental version number and its complete history of modification, review, effectiveness, and repeal is recorded. After a new rule is imported, the historical medical insurance settlement data verification module is automatically triggered. The accuracy and coverage of the rules are calculated using 100,000 settlement data from the past 12 months. Rules with an accuracy rate of less than 88% are marked as pending review and pushed to manual verification.

[0023] In the specific implementation process, step 300 first constructs a medical compliance feature space, matches the medical compliance rule set corresponding to the initial predicted disease group of the current patient from the rule vector library, and then quantifies and calculates 6 core features in sequence: 1. Matching degree of necessary examinations / treatments: The ratio of the total number of examinations and treatments that must be completed before the Nth day of admission as clearly required by the statistical rules to the sum of the number of examinations and treatments that the patient has performed and has effective results and the number of scheduled examinations to be performed. The formula is: Matching degree = (Number of completed necessary examinations + Number of scheduled necessary examinations to be performed × 0.5) / Total number of necessary examinations; 2. Clinical guideline compliance: Load the structured diagnosis and treatment pathways for the corresponding diseases issued by the National Health Commission, convert the time sequence of the diagnosis and treatment items performed by the patient into a one-hot encoded vector, calculate the cosine similarity with the diagnosis and treatment pathway vector recommended by the guideline for the same hospitalization day, and deduct 0.2 if the items prohibited by the guideline have been performed, and normalize the results to the 0-1 interval; 3. Medical insurance restriction level: Iterate through all the medical treatment items that the patient has performed and plans to perform, query the restriction level of each item from the rule vector library, namely no restriction, limited indication, limited second-line drugs, limited medical institution level, and completely prohibited, and assign values ​​of 0, 0.25, 0.5, 0.75, and 1.0 respectively. The overall level is obtained by weighting the average of the expected cost of the item. 4. Medical safety risk level is calculated by weighting the number of comorbidities and complications of the patient (0.1 points for each additional comorbidity), the degree of abnormality of core test indicators (0.05 points for each deviation from the normal range), and the surgical grade (0.2 points for each grade improvement). The total score is normalized to the 0-1 range. 5. Informed consent completion rate: Calculate the ratio of the total number of informed consent forms that need to be signed to the number that have been signed. Items that have been informed but not signed are weighted at 0.5. Finally, the above 6 feature values ​​are concatenated in order to form a medical compliance feature vector with dimensions [1,6].

[0024] Furthermore, a DRG benefit feature space is constructed, and based on the cost consumption trajectory curve and DRG payment standard rules, six benefit features are quantitatively calculated: 1. Cost contribution of treatment items: The proportion of the cost of executed items to the current cumulative total cost is calculated. The proportion of the average cost of the same disease group in history is used for planned execution items. The weight of recent execution is 1.0 and the weight of long-term plans is 0.7. The entropy value of the cost contribution distribution is calculated and normalized to the 0-1 range. 2. Marginal cost: Calculate the sum of the direct costs (in this example, the costs of medicines, consumables, and labor) and indirect costs (in this example, the costs of equipment depreciation and site costs) borne by the hospital when adding a new treatment item; 3. Surplus contribution value, calculated using the formula: Surplus contribution value = Increase in medical insurance payments - Marginal cost; 4. Potential for improving disease group weight: Query the impact of each planned surgery and diagnosis on the disease group weight from the rule vector library; 5. Medical insurance reimbursement ratio: The medical insurance reimbursement ratio for all items is calculated by weighting the total estimated cost to obtain the overall ratio. 6. Cost recovery period, calculated as follows: Cost recovery period = (Marginal cost of the project - Patient's out-of-pocket expenses) / Average daily medical insurance payment; Finally, the above 6 feature values ​​are concatenated in order to form a DRG benefit feature vector with dimensions [1,6].

[0025] Furthermore, an inclusion risk feature space was constructed, and based on the DRG grouping rule vector library and 12 months of historical grouping data, six inclusion risk features were quantitatively calculated: 1. The impact of diagnosis on grouping is calculated using the information gain method to determine the change in entropy of each diagnosis on the grouping result, i.e., the change in entropy of the grouping probability distribution after removing the diagnosis. 2. The impact of surgical procedures on grouping was determined by calculating the change in entropy of each surgical procedure on the grouping results using the information gain method and summing the results. 3. The impact of comorbidities on CC / MCC determination: Query the rule vector library to see if each comorbidity belongs to the CC / MCC catalog and the corresponding weight adjustment coefficient. Calculate the product of the probability of CC / MCC determination that all comorbidities may bring and the weight increase value, and normalize it. 4. Grouping error probability: A logistic regression model is trained based on historical grouping data. The patient's diagnosis, surgery, and comorbidity characteristics are input to obtain the grouping error probability of the current patient. 5. Low-risk probability: Statistically calculate the percentage of relevant minor diagnoses or procedures that were missed during the current patient's recorded diagnosis or surgery, in accordance with relevant rules. 6. High-risk probability: Statistically calculate the percentage of secondary diagnoses or procedures for patients that have been recorded but have no clinical indication during surgery. Finally, the above 6 feature values ​​are concatenated in order to form an inclusion risk feature vector with dimensions [1,6].

[0026] Furthermore, the historical average error rates of the three feature spaces used individually for DRG grouping prediction and cost forecasting over the past 30 days were statistically analyzed. Specifically, the error rate for the medical compliance dimension was the error rate for predicting compliance violations; the error rate for the DRG benefit dimension was the ratio of the average absolute error of cost forecasting to the DRG payment standard; and the error rate for the enrollment risk dimension was the error rate for predicting enrollment risk. Then, the adaptive weights for the three dimensions were calculated using the following formula: ;in, Let be the weight of the i-th dimension. Let be the historical average error rate of the i-th dimension. The Laplace smoothing coefficient is 0.01 in this embodiment. The three 6-dimensional feature vectors are then multiplied by their corresponding weights and concatenated column-wise to form a fused feature tensor of dimensions [1, 18].

[0027] In the specific implementation process, step 400 separates the fused feature tensor into a patient static feature sub-tensor and a treatment time-series dynamic feature sub-tensor according to feature attributes. The static feature sub-tensor is extracted from the fused features of 8 dimensions that do not change with the treatment process, including the diagnostic impact degree, surgical impact degree, basic value of medical insurance payment ratio, and impact degree of comorbidities CC / MCC corresponding to the patient's basic features. The dynamic feature sub-tensor contains the remaining 10 time-series dimensions that are updated in real time with the medical orders, including the matching degree of necessary items, cost contribution degree, current compliance level, and real-time enrollment risk probability. Sine and cosine position codes based on the hospitalization date are added to the dynamic feature sub-tensor, and the sequence length is uniformly padded to 30 days.

[0028] The patient's static feature subtensor was then input into a pre-trained 3-layer multilayer perceptron (MLP). The number of neurons in the hidden layers of this MLP was 128, 64, and 32 respectively, and the ReLU activation function was used. Random deactivation with a dropout rate of 0.2 and a coefficient of 1e was added to each layer. -5 The L2 regularization is applied, and the output layers are a 376-dimensional DRG disease group baseline probability distribution with softmax activation and a disease group baseline cost benchmark value with linear activation, respectively. The MLP is trained based on past de-identified discharge settlement data, using the AdamW optimizer with an initial learning rate of 1e. -4 The batch size is 64, and an early stopping mechanism is set if the validation set loss does not decrease for 5 consecutive rounds. Simultaneously, the dynamic feature sub-tensor of the treatment time series is input into a 2-layer Transformer encoder. The encoder is configured with 4 self-attention heads, an embedding dimension of 64, and a feedforward network dimension of 256. Residual connections and layer normalization are used. The output layer predicts the current values ​​and daily trend curves of drug costs, examination fees, treatment fees, nursing fees, and medical consumable costs for the next 3 days through 5 independent linear heads. The Transformer uses an incremental learning mechanism; for each new medical order, only the attention weights and output head parameters of the last layer are updated. The incremental batch size is 32, and the learning rate is 1e. -5 .

[0029] Next, hard constraint correction is performed based on the rule vector library. First, all DRG grouping hard constraint rules are traversed. For patients who meet the condition of "primary diagnosis + surgical operation" and have a unique corresponding disease group, the probability of the corresponding disease group is set to 1.0 and the probabilities of other disease groups are cleared to zero. The probability of the corresponding disease group of the soft constraint rules is weighted by a factor of 1.2 and then normalized. Then, the medical insurance restriction rules are traversed. The part of the predicted cost that exceeds the daily or single-time limit is cut to the rule limit. The predicted cost of completely prohibited items is set to 0 and marked with a violation label.

[0030] Then, based on the average proportion of the costs of each item in the current predicted disease group over the past 12 months as a preset weight, the weighted sum of the predicted costs of the five items is used to obtain the total predicted cost. The disease group with the highest probability in the corrected probability distribution is selected as the DRG predicted group, and the group stability score P is calculated using the entropy values ​​of the top three high-probability disease groups. The calculation formula is as follows: ,in The probability of being included in the first three disease groups is calculated as 0.6 × (1 - probability of the highest disease group) + 0.2 × low risk probability + 0.2 × high risk probability.

[0031] Finally, the Monte Carlo method was used to quantify uncertainty. Based on historical data, the total cost prediction error was fitted to a normal distribution with a mean of 0 and a standard deviation of 5% of the total cost. 10,000 random samples were performed, and the percentage of times the sampled value was less than or equal to the corresponding DRG payment standard was counted to obtain the surplus probability. At the same time, a 95% confidence interval was calculated. The top-level parameters of the model were fine-tuned once every 32 new diagnosis and treatment data were accumulated. Every Sunday morning, a full training was completed with the full amount of data to update all parameters in order to balance the real-time performance and long-term accuracy of the model.

[0032] In the specific implementation process, step 500, based on the rule vector library and inference results, divides all constraints into inviolable hard constraints and dynamically adjustable soft constraints. Hard constraints include: minimum medical safety requirements (including all guideline-recommended essential treatment items; the absence of any one item directly renders the solution infeasible), medical insurance compliance restrictions (no treatment items marked as completely prohibited in the rule vector library are allowed; restricted items must meet constraints such as indications and usage levels), and essential DRG enrollment characteristics (diagnostic, surgical, and comorbidity-related treatment items that play a decisive role in the current predicted disease group enrollment must be retained). Soft constraints include: the upper limit of DRG payment standards and the average length of hospital stay limit. Subsequently, the patient's disease severity score is calculated to determine the dynamic constraint relaxation. The disease severity score formula is: ; in, For the number of comorbidities and complications, each item in the CC category is worth 1 point, and each item in the MCC category is worth 2 points. The core test indicator is the degree of abnormality. Each deviation from the normal range by 1 time is scored as 0.5 points, with a maximum of 3 points. Surgical procedures were graded, with grades 1-4 scored from 1 to 4 respectively. Non-surgical patients received 0 points. The soft constraint relaxation coefficient was then calculated based on this score. This means that the upper limit of soft constraints for patients with the most severe conditions can be relaxed by 50%.

[0033] Furthermore, a multi-objective optimization function with three optimization objectives is constructed, the first objective being to maximize medical compliance: ,in For required item matching degree, For clinical guideline compliance, In this embodiment, the number of medical insurance violations is [number]. , , The values ​​are 0.4, 0.4, and 0.2 respectively. The second objective is to maximize the DRG balance return. ,in The amount of medical insurance payment for the current disease group. This represents the predicted total cost of the treatment package. The third objective is to minimize enrollment risk. ,in For the probability of grouping error, To reduce the probability of programming risks, To mitigate the high probability of programming errors, a soft-constraint penalty function is constructed: The results are then superimposed on the second objective function. Individuals that do not meet the hard constraints are directly given an infinite penalty value and eliminated. Finally, the NSGA-III algorithm is used to search for Pareto optimal solutions. The combination of treatment items is encoded as a binary chromosome, with each gene bit corresponding to a candidate treatment item. The chromosome length is equal to the total number of available treatment items in the current disease group. The initial population size is 200, the maximum number of iterations is 100, the crossover probability is 0.8, and a two-point crossover method is used. The mutation probability is 0.1, and a bit-flip mutation method is used. Twelve uniformly distributed reference points are set. After each generation of evolution, non-dominated solutions are retained and the reference points are updated. After the iteration is completed, the top 50 non-dominated solutions that satisfy the hard constraints and whose soft constraint violation is within an acceptable range are selected from the final Pareto front to form a candidate treatment item combination set.

[0034] In the specific implementation process, step 600 iterates through each candidate treatment item combination, extracting the structured features of all items one by one. The earliest executable time is determined based on the clinical attributes of the item, while the latest mandatory execution time is calculated by combining the average maximum length of hospital stay and the clinical timeliness of the item. In this embodiment, the latest time for preoperative examination is 24 hours before surgery, and the latest time for postoperative rehabilitation treatment is 24 hours before discharge. The execution duration is taken from the 90th percentile of the same item in the historical execution data of the hospital's HIS system. The single cost is the average actual cost of the same item in the same disease group over the past three months. Pre-dependent items are generated through association mining between the clinical pathway rule base and historical treatment data. Subsequently, a directed acyclic graph (DAG) is constructed based on the extracted features, with treatment items as nodes and pre-dependent relationships as directed edges. The weight of each edge is the minimum time interval between the completion of the pre-dependent item and the execution of the subsequent item. Then, an integer programming model is constructed with the minimization of the daily cost variance within the hospitalization period as the core objective function. The expression is: ,in For the remaining days of hospitalization, The total estimated cost for day t is... The objective function is defined as the average daily estimated cost for the remaining hospitalization period. Simultaneously, the execution dates of all preceding procedures must be earlier than the current procedure, and the same equipment / doctor can execute a maximum of one procedure within the same time period. The daily number of procedures performed by each department must not exceed 120% of the department's maximum daily capacity. A genetic algorithm is then used to solve this objective function, generating 3-5 feasible time-series execution plans for each candidate combination. The cost fluctuation coefficient for each time-series execution plan is further calculated using the following formula: ,in The standard deviation of daily costs. The average daily cost is calculated. Finally, all time-series plans for all candidate combinations are sorted by cost fluctuation coefficient from smallest to largest. The plan with the smallest coefficient is selected as the time-series execution plan for the treatment project. If multiple plans have a coefficient difference of less than 0.01, the plan with the highest inpatient day utilization and the most balanced equipment load is prioritized. The final optimal plan output includes a daily cost detail bar chart, cost fluctuation trend curve, total estimated cost, estimated surplus, compliance score, enrollment risk score, and key project execution reminders.

[0035] In the specific implementation process, step 700 first expands and generates a subset of DIP-specific rule vectors from the existing rule vector library. Based on the original distributed crawler cluster, a targeted crawling task is added to the DIP payment policy column of the medical insurance bureau of the unified planning area. The DIP disease score directory, annual settlement method, core disease and comprehensive disease classification table, disease cost structure control standard, violation audit case library and point adjustment notice are automatically crawled. After document parsing and global deduplication, the data is input into the BERT semantic segmentation model that has been fine-tuned with DIP policy corpus to extract core entities and attributes such as disease code, disease benchmark score, regional average point value, core disease identifier, drug cost ratio limit, consumable cost ratio limit, and violation deduction rules. The data is standardized and converted into an 8-dimensional DIP rule vector containing rule ID, applicable disease range, score benchmark, cost constraint, effective time, and confidence level, and stored in a rule vector library for unified management. Subsequently, based on the patient's standardized primary diagnosis code, surgical procedure code, and comorbidity / complication characteristics, DIP disease matching is completed using precise coding matching combined with BERT semantic similarity verification. The semantic similarity threshold is set to 0.9, prioritizing matching core diseases. If no matching core disease is found, the corresponding comprehensive disease is matched according to the diagnostic subcategory, and the corresponding disease conversion coefficient is obtained. The corresponding disease score and benchmark cost structure threshold for the case are then calculated. Finally, the daily itemized costs in the treatment project's time-series execution plan are calculated according to DIP rules. The IP cost statistics are summarized, and the proportions of overall drug costs, examination fees, treatment fees, consumable costs, and nursing fees are calculated. The deviation from the corresponding DIP disease cost structure control limit is calculated. Simultaneously, the expected total settlement amount under the DIP model is calculated based on regional annual point values ​​and case scores, and a surplus is calculated compared with the total expected cost of the plan. Then, secondary time-series optimization and structural correction under DIP constraints are performed. A greedy adjustment algorithm is used, retaining all clinically necessary treatment items and not changing the clinical necessity of the items. For high-cost items exceeding the cost structure proportion limit, priority is given to adjusting the execution date to evenly distribute them across the remaining hospitalization days to reduce the peak daily proportion. If it still exceeds the control limit... The threshold is then replaced with a low percentage of compliant items in the DIP catalog within the same clinical efficacy range. At the same time, the matching between diagnosis and operation is verified to exclude item combinations that may lead to high scores or inconsistencies between diagnosis and operation. The DIP violation case library is used to identify and mark violation risk points such as item decomposition and double charging through rule matching. Then, the DIP settlement suitability score of the optimized plan is calculated. The scoring formula is 0.4×(1-maximum deviation of cost structure)+0.3×DIP surplus rate+0.3×(1-violation risk value). The daily cost details and total cost prediction of the plan are updated simultaneously. Finally, an optimized treatment execution plan that meets both the DRG payment optimization goal and the DIP payment settlement rule is generated.

[0036] like Figure 2As shown, the present invention also provides a DRG-DIP cost optimization system based on multi-dimensional fusion, comprising: The data processing module is used to perform spatiotemporal synchronous access and standardization processing of multi-source heterogeneous medical data to obtain standardized data. The standardized data includes: patient basic feature matrix, time series of diagnosis and treatment items, and cost consumption trajectory curve. The rule building module is used to build a rule vector library based on DRG grouping rules and medical insurance compliance rules; The feature fusion module is used to construct a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space based on the rule vector library and standardized data, and then dynamically weight and fuse them to obtain a fused feature tensor. The cost inference module is used to infer the fused feature tensor based on the hybrid incremental prediction model and obtain the inference results. The inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability. The project portfolio module is used to perform multi-objective optimization search on the inference results based on dynamic constraint relaxation to obtain a set of candidate treatment project portfolios. The optimization module is used to perform time-series optimization and cost smoothing on the candidate treatment item combination set to obtain the treatment item time-series execution plan. The DIP fusion module is used to perform compliance verification and optimization of the timing execution plan of diagnosis and treatment items through DIP billing rules, so as to obtain an optimized diagnosis and treatment execution plan.

[0037] The beneficial effects of this invention are as follows: 1) The system employs distributed crawlers and the BERT model to perform deep semantic parsing and has a built-in priority-based rule conflict resolution mechanism. This solves the problems of low efficiency in manual data entry, delayed policy updates, and cross-level rule conflicts. New rules are automatically verified for accuracy using historical settlement data. This avoids medical insurance violations and grouping errors caused by policy changes from the source of the rules, and significantly improves the system's adaptability to different regions and different versions of DRG policies. 2) A three-dimensional quantitative feature space was constructed, which includes medical compliance, DRG benefits, and enrollment risks, accurately quantifying the impact of treatment items on group stability; and adaptive weights were calculated based on historical error rates, which can automatically adjust the weights of each dimension according to the differences in the condition of different disease groups and patients, solving the problem of poor generalization of fixed weights; at the same time, the generated 18-dimensional fusion feature tensor comprehensively covers the core influencing factors of DRG cost optimization, and comprehensively improves the feature expression capability. 3) A hybrid prediction architecture combining multilayer perceptron for static features and Transformer for temporal dynamic features is adopted, and hard constraint correction from the rule vector library is used to improve the accuracy of DRG grouping. At the same time, an incremental learning mechanism is introduced, which updates only the top-level parameters of the model for each new medical order, solving the problem of insufficient real-time performance caused by full recalculation of traditional models, and enhancing the accuracy and real-time response capability of the prediction model. 4) The soft constraint relaxation of DRG payment standards and average length of hospital stay is dynamically adjusted based on the severity score of the patient's condition, which avoids the optimization scheme from deviating from clinical reality due to fixed constraints; at the same time, the NSGA-III algorithm is used to search for Pareto optimal solution, which also considers three core objectives: maximizing medical compliance, maximizing DRG surplus benefits, and minimizing enrollment risk, thus solving the limitations of single-objective optimization. 5) An integer programming model was constructed based on the dependency graph of treatment items. With the core objective of minimizing the daily cost variance during the hospitalization cycle, the optimal time series scheme was also selected through the cost fluctuation coefficient to avoid the phased overspending caused by concentrated expenditures. At the same time, the hospital department capacity, equipment load and doctor scheduling were comprehensively considered to reduce the risk of cost fluctuation and overspending.

[0038] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0039] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A cost optimization method for DRG-DIP based on multi-dimensional fusion, characterized in that, Includes the following steps: Multi-source heterogeneous medical data is spatiotemporally synchronized and standardized to obtain standardized data; the standardized data includes: patient basic feature matrix, time sequence of diagnosis and treatment items, and cost consumption trajectory curve. A rule vector library is constructed based on DRG grouping rules and medical insurance compliance rules; Based on the rule vector library and the standardized data, a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space are constructed and dynamically weighted and fused to obtain a fused feature tensor. The fusion feature tensor is inferred based on the hybrid incremental prediction model to obtain the inference results; the inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability. Based on the dynamic constraint relaxation, a multi-objective optimization search is performed on the inference results to obtain a set of candidate diagnosis and treatment item combinations; The candidate treatment item combination set is subjected to time-series optimization and cost smoothing to obtain a treatment item time-series execution plan. The DIP billing rules are used to perform compliance verification and optimization on the timing execution plan of the treatment items, resulting in an optimized treatment execution plan.

2. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, Multi-source heterogeneous medical data is spatiotemporally synchronized and standardized to obtain standardized data, including: Patient data is extracted in real time from HIS, EMR, LIS, PACS and medical insurance settlement systems via a standard HL7 interface and API adapter; the patient data includes: ICD diagnosis code, surgical procedure code, age, gender, comorbidities and complications, performed medical services, incurred costs, laboratory test results and imaging report summaries; The patient's basic feature matrix is ​​obtained by standardizing and converting the ICD diagnostic code, surgical operation code, and treatment item code using a unified coding standard. The diagnostic and treatment items are arranged in chronological order of execution to obtain the time sequence of the diagnostic and treatment items; Using calendar days as the statistical unit, the daily cumulative cost and the percentage of each itemized cost are calculated to obtain the cost consumption trajectory curve.

3. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, A rule vector library is constructed based on DRG grouping rules and medical insurance compliance rules, including: We collect DRG payment policy documents, adjustment notices, interpretive documents, and violation case databases published on the official websites of the National Healthcare Security Administration and provincial healthcare security bureaus. The DRG payment policy document was semantically segmented using the BERT model to identify key information, including: DRG grouping rules, payment standards, list of restricted medical insurance items, definition of violations and penalty standards. The key information is converted into rule vectors, and the rule vector library is generated. The rule vector includes: rule ID, rule type, applicable disease group scope, effective time, constraints, execution action, and confidence level.

4. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, The medical compliance feature space includes: matching degree of necessary examination items, matching degree of necessary treatment items, compliance with clinical guidelines, medical insurance restriction level, medical safety risk level, and informed consent completion rate; the DRG benefit feature space includes: cost contribution of diagnosis and treatment items, marginal cost, surplus contribution value, potential for improvement of disease group weight, medical insurance payment ratio, and cost recovery period; the enrollment risk feature space includes: the impact of diagnosis on grouping, the impact of surgical procedure on grouping, the impact of comorbidities and complications on CC / MCC determination, grouping error probability, low grouping risk probability, and high grouping risk probability.

5. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, Based on the rule vector library and the standardized data, a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space are constructed and dynamically weighted and fused to obtain a fused feature tensor, including: Calculate the adaptive weights for the medical compliance feature space, the DRG benefit feature space, and the enrollment risk feature space, respectively; The feature vectors of the three dimensions are weighted and fused according to the adaptive weights to obtain the fused feature tensor.

6. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 5, characterized in that, The formula for calculating the adaptive weight is: ;in, Let i be the weight of the i-th dimension. Let be the historical average error rate of the i-th dimension. is the Laplace smoothing coefficient.

7. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, The fused feature tensor is inferred based on a hybrid incremental prediction model to obtain the inference results, including: The fused feature tensor is separated into a patient static feature sub-tensor and a diagnosis and treatment time-series dynamic feature sub-tensor; The patient static feature subtensor is input into a multilayer perceptron to obtain the DRG grouping baseline probability distribution and the baseline cost value of the disease group. The dynamic feature subtensor of the diagnosis and treatment time series is input into the Transformer model to obtain the predicted values ​​and trend curves of drug costs, examination costs, treatment costs, nursing costs and medical consumable costs for the next 3 days. Based on the rule vector library, the basic probability distribution of the DRG group is hard-constrained and corrected, and the predicted values ​​are pruned for compliance, to obtain the corrected probability distribution and the predicted values ​​of the itemized costs. The predicted values ​​of the individual expenses are weighted and summed according to preset weights to obtain the predicted total cost. The disease group with the highest probability in the modified probability distribution is selected as the DRG prediction group, and the group stability score and the enrollment risk probability are calculated. The uncertainty of the total cost prediction is quantified using the Monte Carlo method, and the surplus probability is calculated in conjunction with the DRG payment standard.

8. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, Based on the dynamic constraint relaxation, a multi-objective optimization search is performed on the inference results to obtain a set of candidate treatment item combinations, including: Determine the type of constraint; the constraint types include: hard constraints and soft constraints; the hard constraints include: minimum medical safety requirements, medical insurance compliance restrictions, and DRG enrollment requirements; the soft constraints include: the upper limit of DRG payment standards and the average length of stay limit; Construct a multi-objective optimization function based on the stringency of the constraint types; The Pareto optimal solution search for the multi-objective optimization function is performed using the NSGA-III algorithm to obtain the candidate treatment item combination set.

9. The DRG-DIP cost optimization method based on multi-dimensional fusion according to claim 1, characterized in that, The candidate treatment item combination set is subjected to time-series optimization and cost smoothing to obtain a treatment item time-series execution plan, including: Extract the project features from the candidate treatment item combination set; the project features include: earliest executable time, latest required execution time, execution duration, single cost, and prerequisite dependent items; Construct a dependency graph of diagnosis and treatment items based on the described project characteristics; Based on the dependency graph of the treatment items, and with the objective function of minimizing the daily cost variance during the hospitalization period, high-cost items are evenly distributed over the remaining hospitalization days to generate a time-series execution plan. The time-series execution plan includes: the suggested execution date, time period, and estimated cost for each item. Calculate the cost fluctuation coefficient for each of the time-series execution schemes, and select the scheme with the smallest cost fluctuation coefficient as the time-series execution scheme for the diagnosis and treatment project.

10. A DRG-DIP cost optimization system based on multi-dimensional fusion, characterized in that, include: The data processing module is used to perform spatiotemporal synchronous access and standardization processing of multi-source heterogeneous medical data to obtain standardized data; The standardized data includes: a patient basic characteristic matrix, a time series sequence of diagnosis and treatment items, and a cost consumption trajectory curve; The rule building module is used to build a rule vector library based on DRG grouping rules and medical insurance compliance rules; The feature fusion module is used to construct a medical compliance feature space, a DRG benefit feature space, and an enrollment risk feature space based on the rule vector library and the standardized data, and to dynamically weight and fuse them to obtain a fused feature tensor. The cost inference module is used to infer the fused feature tensor based on the hybrid incremental prediction model to obtain the inference results; the inference results include: DRG predicted grouping, total cost prediction value, surplus probability, grouping stability score and grouping risk probability. The project portfolio module is used to perform multi-objective optimization search on the inference results based on dynamic constraint relaxation to obtain a set of candidate treatment project portfolios. The optimization module is used to perform time-series optimization and cost smoothing on the candidate treatment item combination set to obtain a treatment item time-series execution plan. The DIP fusion module is used to perform compliance verification and optimization of the timing execution plan of the diagnosis and treatment items through DIP billing rules, so as to obtain an optimized diagnosis and treatment execution plan.