DRG Risk Identification Method, Device and Medium Based on Collaborative Learning

Abstract

This application provides a DRG risk identification method, device, and medium based on collaborative learning, belonging to the field of medical big data analysis technology. The method includes: constructing a collaborative learning framework containing a high-precision first prediction model and an interpretable second prediction model; training the second model through knowledge distillation to solidify its parameters into a "DRG-medical item" importance matrix; if the DRG grouping predicted by the model is inconsistent with the actual declared grouping and has a lower standard weight, it is determined to be a high-risk case; generating dual evidence for high-risk cases: item-level micro-attribution evidence based on the importance matrix, and macro-level corroboration based on group consensus from similar case retrieval in a historical database. This application unifies high-precision prediction with strong interpretability, achieving accurate and transparent risk identification by providing a dual chain of evidence, thus enhancing the reliability of the conclusions.

Description

Technical Field

[0001] This invention relates to the field of medical big data analysis technology, specifically to a method, device, and medium for DRG risk identification based on collaborative learning. Background Technology

[0002] Diagnosis Related Groups (DRG) payment is the core of current healthcare payment reform, aiming to guide medical institutions to standardize medical practices and control medical costs through standardized pricing of medical services. However, this payment method has also created new risks to the medical insurance fund, particularly the "over-coding" behavior. This involves some medical institutions selecting non-optimal but higher-paying diagnostic or procedural codes to categorize cases into higher-paying DRG groups, thereby obtaining unreasonable overpayments.

[0003] Existing technologies for identifying such risks have several shortcomings. For example, rule-based expert systems rely on manually pre-set rule bases, making it difficult to detect new and hidden high-level coding patterns. Furthermore, rule updates are often delayed, leading to high maintenance costs. Threshold-based early warning methods using statistical indicators detect anomalies by monitoring macro-level indicators such as average cost per group and length of hospital stay. However, due to their coarse analytical granularity, they cannot effectively distinguish between reasonably high costs due to complex patient conditions and unstructured high-level coding behavior, resulting in high false positive and false negative rates. Traditional machine learning-based classification methods typically require a large number of labeled "fraudulent" or "violation" samples for supervised learning. However, in real-world scenarios, such confirmed samples are extremely scarce, making model training difficult and resulting in insufficient generalization ability. While unsupervised learning-based anomaly detection methods do not rely on labels, their outputs often lack clear business semantics, failing to explain the causes of risks. Their "black box" nature makes the conclusions lack credibility and operability in regulatory audit scenarios requiring clear evidence.

[0004] In recent years, artificial intelligence technologies, represented by deep learning, have demonstrated a powerful ability to automatically learn features from high-dimensional and complex data. However, directly applying them to the field of DRG risk supervision introduces new technical contradictions: generally, the higher the complexity and prediction accuracy of a model, the less transparent its internal decision-making logic becomes, and the more prominent the "black box" problem becomes. Although some technical solutions attempt to improve the predictive performance of specific tasks through knowledge transfer between models, their main purpose is to improve prediction accuracy, and they have not effectively solved the problems of transparency and interpretability of the model's decision-making process. In the context of medical insurance supervision, where clear and traceable evidence is required to justify the reasonableness of every fund payment, this "black box" characteristic becomes the core obstacle to its application and promotion. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide a DRG risk identification method, device and medium based on collaborative learning.

[0006] According to the present invention, a DRG risk identification method based on collaborative learning is provided. This method is executed by a computer device and includes: Construct a collaborative learning framework that includes a first prediction model and a second prediction model, wherein the first prediction model is a prediction model used to predict the first DRG group based on case data, and the second prediction model is an interpretable model whose model parameters can form an interpretable knowledge representation. Train the first prediction model; The second prediction model is trained using knowledge distillation techniques, wherein the prediction results output by the first prediction model are used as soft labels to guide the parameter learning of the second prediction model, thereby obtaining the interpretable knowledge representation solidified into a global "DRG-medical item" importance matrix; Provide a historical case database; Obtain the standard weights corresponding to each DRG group; Obtain a test case containing the actual declared DRG group; The case to be tested is predicted. If the DRG grouping predicted based on the first prediction model is inconsistent with the actual declared DRG grouping, and the standard weight corresponding to the predicted first DRG grouping is lower than the standard weight corresponding to the actual declared DRG grouping, then a risk label is added to the case to be tested.

[0007] Preferably, it further includes: generating dual evidence for cases with risk labels, the dual evidence comprising: Based on the aforementioned "DRG-Health Item" importance matrix, the health items of the cases with risk labels are analyzed to generate first evidence for item-level attribution; and A set of cases similar to the risky cases in terms of medical resource consumption patterns is retrieved from the historical case database to generate a second piece of evidence supporting the consensus of the group.

[0008] Preferably, the first prediction model is a deep neural network model based on the Transformer architecture; The training of the first prediction model includes: Pre-training of a mask-based language model on medical project sequences.

[0009] Preferably, the second prediction model is a linear model, and the parameters of the linear model are the importance matrix of the "DRG-medical items".

[0010] Preferably, the step of generating the second evidence includes: The case data is encoded into a case feature vector using the first prediction model; A similarity search is performed on the case feature vectors to retrieve the set of similar cases.

[0011] Preferably, the method further includes: Calculate a comprehensive risk score for cases with risk labels.

[0012] Preferably, the comprehensive risk score is calculated based on one or more risk factors; The risk factors include at least one of the following: model prediction confidence output by the first prediction model, economic difference driving force calculated based on the standard weight difference between the predicted first DRG group and the actual declared DRG group, and similar case corroboration calculated based on the number of similar case sets retrieved from the historical case database or similarity scores.

[0013] A DRG risk identification device according to the present invention includes a processor and a memory. The memory stores a computer program, a historical case database, and standard weights corresponding to each DRG group. When the computer program is executed by the processor, it implements the following: The model building module is used to build a collaborative learning framework that includes a first prediction model and a second prediction model, wherein the first prediction model is a high-precision prediction model and the second prediction model is an interpretable model. The model training module is used to train the first prediction model; The knowledge distillation module is used to train the second prediction model through knowledge distillation technology to obtain an interpretable knowledge representation that is solidified into a global "DRG-medical item" importance matrix; The case determination module is used to add a risk label to the case to be tested if the DRG group predicted based on the first prediction model is inconsistent with the actual declared DRG group, and when comparing the standard weights stored in the memory, the standard weight corresponding to the predicted first DRG group is lower than the standard weight corresponding to the actual declared DRG group. The evidence generation module is used to generate dual evidence for cases identified as at risk. The dual evidence includes a first piece of evidence generated based on the importance matrix of the "DRG-medical items" and a second piece of evidence generated based on similarity retrieval of the historical case database.

[0014] Preferably, when the computer program is executed by the processor, it further implements: The risk quantification module is used to calculate a comprehensive risk score for cases with risk labels.

[0015] According to the present invention, a computer-readable storage medium is provided thereon storing a computer program, which, when executed by a processor, implements the method described herein.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This application constructs a collaborative learning framework that includes a high-precision prediction model and an interpretable model. By utilizing knowledge distillation technology, it transfers and solidifies the powerful pattern learning capabilities of complex models into a simple model with completely transparent parameters. This unifies high-precision prediction and strong model interpretability at the system level, overcomes the technical obstacle of "black box" decision-making when artificial intelligence models are applied to serious regulatory scenarios, and achieves a technical leap from "black box" prediction to "white box" decision-making.

[0017] 2. This invention automatically learns grouping logic from massive historical data, without relying on scarce, manually labeled risk samples, and can proactively discover unknown or evolving abnormal patterns, breaking through the data bottleneck of traditional supervised learning methods.

[0018] 3. This invention provides project-level micro-attribution evidence and macro-level corroboration based on similar case retrieval for risk conclusions, constructing a dual technical evidence chain. This greatly enhances the robustness and persuasiveness of the risk conclusions and provides complete technical evidence support for subsequent audits and verifications. Simultaneously, by conducting multi-dimensional quantitative assessments of risks, it achieves refined risk classification and ranking. Attached Figure Description

[0019] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic diagram of the overall process of a DRG risk identification method based on collaborative learning provided in this application embodiment; Figure 2 A schematic diagram of the architecture of a DRG risk identification system provided in this application embodiment; Figure 3 This is a schematic diagram illustrating the interaction timing of an online risk identification method provided in an embodiment of this application.

[0020] Explanation of reference numerals in the attached figures: Detailed Implementation

[0021] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0022] Example 1 In one embodiment of this application, a collaborative learning-based DRG risk identification method and its complete implementation process are described in detail. This embodiment aims to achieve high-precision identification of the risk of over-coding in medical insurance and provide transparent and traceable dual evidence for risk conclusions.

[0023] Reference Figure 1 This document illustrates a schematic diagram of the overall process of a collaborative learning-based DRG risk identification method provided in an embodiment of this application. The method is mainly divided into two stages: an offline model building and knowledge base generation stage, and an online risk identification and evidence generation application stage.

[0024] First, in the offline phase, model building and knowledge base generation are performed, which includes the following steps: Step 1: Data Preparation and Preprocessing. The system obtains massive amounts of historical inpatient medical record front page data and cost details from the medical insurance information platform or hospital information system as raw training data. This data typically includes the patient's demographic information (such as age and gender), diagnostic information (such as primary diagnosis and other diagnoses, usually using the International Classification of Diseases coding), surgical and procedural information, and a detailed list of all medical services, medications, and consumables incurred during hospitalization.

[0025] Data preprocessing aims to provide a high-quality data foundation for subsequent model training. Specifically, this step includes: 1. Data cleaning: handling missing values, outliers, and inconsistent records in the raw data, such as removing or correcting outlier records with negative costs or zero days of hospitalization. 2. Medical item standardization: Given that different hospitals may have different names and codes for the same medical service item, a globally unified "medical item dictionary" needs to be established. Through fuzzy matching, rule mapping, etc., all medical item names in the raw data are uniformly mapped to this dictionary, and a unique item ID is assigned to each unique medical item. Assume that after standardization, the global medical item dictionary contains N items. 3. Case serialization: For each inpatient case, all medical items in its cost details list are organized into a sequence of item IDs according to their chronological order during hospitalization. To enable the model to handle variable-length inputs, a fixed maximum sequence length L is set (e.g., L=512). Sequences longer than L are truncated, for example, retaining only the latest 512 items; sequences shorter than L are padded at the end with a special padding marker (e.g., a [PAD] marker with ID 0) until the length reaches L. After this step, each case is converted into an integer sequence of length L, which serves as input to the subsequent model.

[0026] Step Two: Construct and train the first prediction model in the collaborative learning framework. In this embodiment, the first prediction model is designed as a high-precision prediction model with a complex structure and high prediction accuracy. This model, in Figure 2 The system architecture shown corresponds to teacher model 21, which is specifically implemented as a deep neural network model based on the Transformer architecture. The training of teacher model 21 employs a two-stage strategy: The first stage is self-supervised pre-training based on a masked language model. The goal of this stage is to enable the model to learn the inherent relationships between medical items and the semantic representation of treatment patterns from a massive amount of case sequences without DRG grouping labels. Specifically, a network structure containing multiple (e.g., 6) Transformer encoder layers is constructed, and a large number of case item ID sequences obtained from the aforementioned preprocessing are used as input. During training, for each input sequence, 15% of the item IDs are randomly selected and masked, i.e., the original item ID is replaced with a special [MASK] label. The model's task is to predict the original item ID at the masked position based on the other unmasked item IDs in the sequence (i.e., the context). By optimizing the loss function for this prediction task (usually cross-entropy loss), the model is forced to learn co-occurrence relationships, temporal relationships, and complex treatment logic between items. For example, the model learns that the item "appendectomy" often co-occurs with the diagnosis of "appendicitis" and certain specific antibiotics and anesthetics. This pre-training process enables the teacher model 21 to possess a deep understanding of the case data.

[0027] The second stage is supervised fine-tuning based on real DRG labels. After pre-training, a classification head (e.g., a fully connected layer followed by a Softmax activation function) is added to the top layer of the Transformer encoder. Then, the entire model is fine-tuned using a batch of historical case data with verified and accurate DRG grouping labels. At this point, the model's task becomes predicting the correct DRG grouping based on the input sequence of case item IDs. Through this fine-tuning process, the model adapts the general clinical knowledge learned in the pre-training stage to the specific DRG grouping prediction task, thus becoming a high-precision DRG grouping predictor.

[0028] The trained teacher model 21 possesses two key functions: first, it can predict the most likely DRG group to which any case belongs; second, it can encode the sequence information of any case into a low-dimensional, dense real-valued vector, namely the "case feature vector". This vector can be extracted from the output of the last layer of the Transformer encoder or the output of a specific label (such as the [CLS] label), and it highly summarizes the overall medical resource consumption pattern and clinical characteristics of the case.

[0029] Step 3: Train the second prediction model using knowledge distillation techniques. The second prediction model, in... Figure 2The system architecture diagram is represented by Student Model 22, which is designed as a simple, fully transparent, and interpretable model. In this embodiment, Student Model 22 is specifically a linear model, the core of which is a trainable weight matrix W with dimensions [M,N], where M is the total number of DRG groups and N is the size of the global medical item dictionary. Each element W(i,j) of matrix W has a clear physical meaning, that is, it represents the contribution or importance weight of medical item j in determining whether a case belongs to DRG group i.

[0030] The purpose of knowledge distillation is to transfer and "distill" the complex, implicit knowledge learned by the teacher model 21 into the structurally simple student model 22. The specific training process is as follows: The same training dataset as used during the fine-tuning of the teacher model 21 is used. For each case in the dataset, the following operations are performed: 1. Obtain the "hard label," which is the true, verified DRG group label for the case. 2. Input the case into the trained teacher model 21 and obtain its output "soft label." The soft label is not a single prediction result, but rather a probability distribution vector predicted by the teacher model for all M DRG groups. This vector contains richer information; for example, the teacher model might consider a case to have a 70% probability of belonging to DRG_A and a 20% probability of belonging to DRG_B, reflecting the similarity between DRG_A and DRG_B.

[0031] The training objective of student model 22 is to make its output prediction distribution as close as possible to the soft labels output by teacher model 21, while also taking into account the true hard labels. Its total loss function is... It can be represented as: in, This is the output probability distribution of student model 22. These are soft labels output by Teacher Model 21. These are real hard labels (usually represented as one-hot vectors). It is the Kolbec-Leibler divergence loss, which measures the difference between the output distribution of the student model and the teacher model, prompting the student model to imitate the teacher model; It is the cross-entropy loss, used to ensure that the student model's predictions are consistent with the true labels. and This is a hyperparameter used to balance the weights of the two loss terms.

[0032] Once the student model 22 has converged during training, its parameter matrix W is fixed. This fixed matrix W constitutes the global, interpretable "DRG-healthcare item" importance matrix 24. This matrix is ​​a core knowledge output of this application, which quantitatively reveals the basis for the judgment of each healthcare service item for each DRG group in a completely "white-box" manner.

[0033] Step 4: Constructing an Offline Knowledge Base. This step primarily generates two core knowledge bases for use in the online application phase. 1. Case Vector Database 23: All historical case data collected by the system are processed one by one using the trained teacher model 21 to extract the case feature vector for each case. These millions or even tens of millions of vectors are stored in a dedicated vector database, and an efficient index is established (e.g., using an index structure based on an approximate nearest neighbor search algorithm such as HNSW) to facilitate rapid similarity retrieval later. 2. Global DRG - Item Importance Matrix 24: This is the weight matrix W that is fixed after the aforementioned student model training is completed.

[0034] At this point, the offline modeling phase is complete, and the system has generated the core model and knowledge base for online identification. Accordingly, the methodology then moves to the online risk identification and evidence generation application phase.

[0035] Combined with reference Figure 3 This illustrates the interaction sequence of the online risk identification method. When a pending inpatient record containing the actual declared DRG group (denoted as Real_DRG) requires risk assessment, the system performs the following steps: Step 1 (S301): Receive cases to be reviewed. The client (e.g., medical insurance settlement system or audit front-end) sends the data of the cases to be reviewed to the risk calculation engine 31.

[0036] Step 2 (S302, S303): Model Prediction. Risk calculation engine 31 calls the teacher model service and inputs the medical item sequence of the case to be reviewed into teacher model 21. Teacher model 21 returns three key pieces of information: the predicted DRG group (denoted as Pred_DRG), the confidence level of the prediction result (i.e., the Softmax probability value corresponding to Pred_DRG), and the case feature vector.

[0037] Step 3 (S304): Core Risk Assessment. The risk calculation engine 31 executes a crucial screening logic. First, it retrieves the standard weights (or payment weights, CMI values) corresponding to Pred_DRG and Real_DRG from the pre-configured DRG standard weight table. These standard weights reflect the technical difficulty and resource consumption levels of different DRG groups and are core parameters for medical insurance payment. Then, the engine determines whether the following two conditions are met simultaneously: 1. Grouping inconsistency, i.e., Pred_DRG and Real_DRG are not the same; 2. Weight decreasing trend, i.e., the standard weight corresponding to Pred_DRG is lower than the standard weight corresponding to Real_DRG. This combination of conditions aims to accurately identify the typical characteristics of "high-risk coding," i.e., a medical institution submits a case that should belong to a lower payment standard group to a higher payment standard group through coding operations. Only when both conditions are met simultaneously is the case initially identified as a high-risk case and enters the subsequent evidence generation and quantification stage. Otherwise, the system determines it to be low-risk, and the process terminates.

[0038] Step 4: Generate dual evidence for high-risk cases. For cases identified as high-risk, the system automatically constructs two types of evidence to support the determination.

[0039] The first type of evidence is micro-level evidence, which is project-level attribution evidence generated based on the "DRG-Healthcare Items" importance matrix 24. Specifically, the risk calculation engine 31 (corresponding to...) Figure 3 S307 and S308 of the model utilize this matrix to analyze the medical items included in high-risk cases. The analysis may include: 1. Identifying key items supporting the Pred_DRG, i.e., identifying items in the case's item list that have high positive weights for the Pred_DRG in the importance matrix 24; the presence of these items is the primary basis for the model's Pred_DRG judgment. 2. Identifying key items that contradict or are missing from the Real_DRG, i.e., identifying core diagnostic and treatment items that have high positive weights for the Real_DRG but are missing in the case, or identifying items that are present in the case but have significant negative weights for the Real_DRG. These analytical results are presented in list form, intuitively explaining to reviewers the reasons for discrepancies between the model's judgment and the hospital's report, thereby attributing the risk to the specific medical service item level.

[0040] The second type of evidence is macro-level evidence, which is corroborated by group consensus generated by retrieving similar case sets from the case vector database 23. Specifically, the risk calculation engine 31 (corresponding to...) Figure 3In steps S305 and S306, the case feature vectors obtained in step two are used to perform an efficient approximate nearest neighbor search in the case vector database 23 to retrieve a batch of historical cases (e.g., Top 10 or Top 20) that are most similar to the current high-risk case in terms of medical resource consumption patterns. Then, the DRG grouping distribution of this batch of similar cases is statistically analyzed. If the vast majority (e.g., 90%) of these similar cases belong to the Pred_DRG group, while very few or no cases belong to the Real_DRG group, this forms strong evidence of a collective consensus: other cases most similar to this high-risk case have historically been generally assigned to the lower-weighted group predicted by the model. This provides macro-level corroboration for the rationality of the model's judgment.

[0041] Step 5 (S309): Calculate the comprehensive risk score. To refine the risk classification, the risk calculation engine 31 calculates a comprehensive risk score. In this embodiment, the score integrates the following four risk factors: *R1 (Model Macro Reliability): Set based on the predictive performance metrics (such as F1 score) of the teacher model 21 on the historical test set for the specific group Pred_DRG. If the model's predictions for this group are consistently accurate, this factor will have a high score. *R2 (Case Prediction Confidence): The confidence probability value of the Pred_DRG output by the teacher model 21. *R3 (Economic Difference Driving Force): Calculated based on differences in standard weights, for example, it can be defined as... The closer this value is to 1, the greater the economic benefit and the stronger the motivation for using high-level coding. *R4 (Similar Case Corroboration): Calculated based on the retrieved set of similar cases, for example, it can be defined as the proportion of cases in the similar case set whose DRG group is Pred_DRG. Finally, these factors are combined into a final comprehensive risk score using a weighted formula, for example: Risk_Score=60+(w1*R1+w2*R2+w3*R3+w4*R4)*10, where w1 to w4 are weights that can be dynamically adjusted according to regulatory priorities. This score (e.g., 60-100 points) helps regulators rank a large number of risky cases to prioritize the review of the highest-risk cases.

[0042] Step Six (S310): Output and Storage. The risk calculation engine 31 integrates the detailed information of risk cases, the comprehensive risk score, and the generated micro and macro evidence into a structured risk report, stores it in the risk results database, and presents it to auditors through the visualization workbench 41.

[0043] Reference Figure 2This document illustrates a schematic diagram of the architecture of a DRG risk identification system provided in an embodiment of this application. Logically, the system can be divided into four layers: a data layer 10, responsible for storing raw and standardized data; a model and knowledge base layer 20, including the teacher model 21, student model 22, case vector library 23, and global DRG-item importance matrix 24 trained and generated in this embodiment; a service layer 30, whose core is a risk calculation engine 31, responsible for scheduling the model and knowledge base and executing all the logic of online risk identification; and an application layer 40, including a visualization workbench 41 for human-computer interaction and a database for storing results. The complete functionality described in this embodiment corresponds to the full-function deployment mode of this system architecture.

[0044] Example 2 As an optional implementation of Embodiment 1, this embodiment aims to illustrate that the collaborative learning framework proposed in this application has good model universality, and its core idea does not depend on a specific type of complex model. In this embodiment, the structure of the first prediction model (i.e., teacher model 21) has changed, while the rest of the framework remains unchanged.

[0045] Specifically, the main difference from Embodiment 1 lies in the specific structure and training method of the teacher model 21. In Embodiment 1, the teacher model adopts a Transformer-based architecture; while in this embodiment, the teacher model 21 adopts a recurrent neural network model based on gated recurrent units.

[0046] The structure of Teacher Model 21 is adjusted as follows: The model input layer also receives a preprocessed sequence of item IDs of length L. Each ID in the sequence is first converted into a low-dimensional dense word vector through an embedding layer. Then, the word vector sequence is fed into a bidirectional gated recurrent unit network. It should be noted that the bidirectional network can capture the temporal dependencies in the sequence from both the forward and backward directions simultaneously.

[0047] Accordingly, the training method of teacher model 21 is also adjusted. Since the gated recurrent unit network itself is good at processing sequence data, in this embodiment, the complex masked language model pre-training stage in embodiment 1 can be omitted. Instead, an end-to-end approach is adopted, using data with real DRG group labels to supervise the training of the bidirectional gated recurrent unit network. In order to obtain a fixed-length vector representing the entire case, the hidden states of the bidirectional network at all time steps can be aggregated, for example, through average pooling or max pooling operations, to obtain a comprehensive case representation vector. This vector is then fed into the classification head (fully connected layer + Softmax) for DRG group prediction.

[0048] Although the internal structure of Teacher Model 21 has changed from Transformer to bidirectional gated recurrent unit network, its external functions remain the same: 1. For an input case sequence, it can still predict a DRG group Pred_DRG and its confidence. 2. It can still extract a "case feature vector" representing the overall features of the case from the intermediate layers of the model (e.g., the aggregation vector before the pooling layer).

[0049] Therefore, all subsequent steps, including the training of student model 22 (knowledge distillation), the generation of the global DRG-item importance matrix 24, the construction of the case vector database 23, and the entire online risk identification and evidence generation process (such as... Figure 3 As shown, all are exactly the same as in Example 1. Student model 22 still learns the soft labels output by the teacher model based on the gated loop unit through the knowledge distillation process, thereby solidifying its implicit knowledge into an interpretable importance matrix.

[0050] The intended effect of this embodiment is to demonstrate the robustness and flexibility of the proposed "teacher-student" collaborative learning framework. Regardless of whether the teacher model is based on a Transformer with a self-attention mechanism or a recurrent neural network with a recurrent mechanism (such as a gated recurrent unit or a long short-term memory network), this framework can effectively transform its powerful sequence modeling capabilities into a unified and interpretable knowledge representation (i.e., an importance matrix) through knowledge distillation techniques, thus achieving an organic combination of high-precision prediction and strong interpretability.

[0051] Example 3 This embodiment is also a variation of Embodiment 1, primarily demonstrating the flexibility of risk quantification methods. Embodiment 1 employs a linear weighted fusion approach to calculate the comprehensive risk score, a method that is intuitive and easy to configure. As an optional implementation, this embodiment provides a machine learning-based fusion method, aiming to more intelligently combine multiple risk factors through a data-driven approach to achieve more accurate risk ranking capabilities.

[0052] The main difference between this embodiment and Embodiment 1 lies in the "calculating the comprehensive risk score" step in the online risk identification process (corresponding to...). Figure 3 (S309 in the text).

[0053] In this embodiment, the calculation of risk factors is exactly the same as in Embodiment 1. That is, for a case that is initially identified as risky, the system still calculates four basic risk factors: R1 (macroscopic reliability of the model), R2 (case prediction confidence), R3 (economic difference driving force) and R4 (corroboration by similar cases).

[0054] However, in fusing these factors, this embodiment introduces a pre-trained "risk score fusion model". This model can be a simple machine learning model, such as a logistic regression model or a gradient boosting decision tree model.

[0055] The training process of this risk-based fusion model is as follows: 1. Data Preparation: First, a small batch of sample data, reviewed and confirmed by domain experts (such as senior medical insurance audit experts), needs to be obtained. These samples are all cases screened through the core risk assessment steps in Example 1. Experts need to make a final judgment on them, labeling them as "confirmed high risk" (positive samples) or "reasonable difference" (negative samples). 2. Feature Engineering: For this batch of labeled samples, calculate the corresponding R1, R2, R3, and R4 risk factor values ​​for each sample. These four values ​​will serve as the input features of the fusion model. 3. Model Training: Using the above features and the labels marked by experts, a binary logistic regression model is trained. The training objective of the model is to learn a function that can predict the probability that a case belongs to the "confirmed high risk" category based on the four input risk factors.

[0056] During the online application phase, after the risk calculation engine 31 calculates the four factors R1 to R4 for risky cases, it no longer uses a linear weighting formula. Instead, it feeds these four factors as input to the pre-trained logistic regression model. The model outputs a probability value between 0 and 1, which represents the likelihood that the case is a true high-risk case.

[0057] To make the results more intuitive, this probability value can be linearly mapped to a comprehensive risk score, for example, between 60 and 100. For instance, the formula could be: Risk_Score = 60 + 40 * Probability.

[0058] The expected effect of this embodiment is that, compared to the simple linear weighting method in Embodiment 1, using a machine learning model to fuse risk factors can automatically learn the nonlinear relationships and complex interactions that may exist between the various factors. For example, the model may learn that when the economic difference driver (R3) is extremely high, even if the corroboration of similar cases (R4) is not particularly high, the overall risk should be significantly increased. This data-driven fusion approach is expected to provide a more accurate risk ranking capability than manually setting weights, thereby helping regulatory authorities to more effectively allocate and schedule limited audit resources and prioritize the most suspicious cases.

[0059] Example 4 This embodiment aims to demonstrate the architectural flexibility and high adaptability to different deployment environments of the technical solution of this application, specifically describing a lightweight deployment scheme. This scheme is particularly suitable for deployment environments with relatively limited computing and storage resources, such as the local servers of some municipal-level medical security departments.

[0060] The core idea of ​​this embodiment is to retain the most essential risk identification and micro-evidence generation functions based on an interpretable model, while sacrificing some prediction accuracy (which may be lower than that of the teacher model) and macro-evidence capability.

[0061] Refer again Figure 2 The system architecture diagram shown illustrates that, in the full-featured deployment mode described in Example 1, the service layer 30 needs to simultaneously deploy and invoke the teacher model 21, student model 22, case vector database 23, and importance matrix 24. The teacher model 21 (especially large Transformer-based models) consumes significant computational resources during inference, while the case vector database 23 requires substantial storage space (potentially reaching TB levels for databases with tens of millions of cases) and memory.

[0062] In the lightweight deployment scheme of this embodiment, the process is as follows: 1. Offline Centralized Training: On a central server with sufficient computing resources (e.g., a provincial or national data center), the entire offline training process described in Example 1 is executed. This includes data preprocessing, pre-training and fine-tuning of the teacher model 21, and knowledge distillation of the student model 22. The process ultimately produces the trained teacher model 21, student model 22 (and their weight matrix W, i.e., importance matrix 24), and case vector database 23.

[0063] 2. Lightweight Local Deployment: Deploy only a subset of the entire system on a resource-constrained local server. Specifically, deploy only: Student model 22 (essentially its weight matrix W); Global DRG - Project Importance Matrix 24 (same entity as the student model); A simplified version of the Risk Calculation Engine 31; The medical project dictionary and DRG standard weight table are used. However, the teacher model21 and case vector library23, which have huge computational and storage costs, are not deployed locally.

[0064] 3. Localized Risk Identification Process: In the localized risk identification process, when the local system receives a case to be reviewed, it executes a simplified and computationally efficient process that does not rely on complex deep learning inference and large-scale vector retrieval. Input processing: The list of medical items for the cases to be reviewed is dictionary-looked and ID-based, then converted into an N-dimensional multi-hot encoded vector, where N is the total number of medical items globally. In the vector, items used in the case correspond to a dimension of 1, while unused items correspond to 0.

[0065] Model prediction: Prediction is performed using Student Model 22. Since the Student Model is a linear model, its prediction process is simplified to a single matrix multiplication operation: multiplying the case's multi-hot encoding vector with the importance matrix 24 (i.e., the weight matrix W of the Student Model) yields the case's score across all M DRG groups, with the highest score being the predicted DRG group Pred_DRG_S. Understandably, this calculation is extremely fast and can be completed instantaneously on a standard CPU.

[0066] Risk assessment: Apply the same core risk assessment logic as in Example 1, that is, determine whether Pred_DRG_S is inconsistent with the actual declared Real_DRG, and whether the standard weight corresponding to Pred_DRG_S is lower than the weight of Real_DRG.

[0067] Evidence Generation: Only micro-level evidence is generated. If a case is determined to be at risk, the system directly analyzes and provides a list of key items supporting Pred_DRG_S and / or questioning Real_DRG based on the importance matrix 24, as a basis for explanation and attribution. Since the case vector library 23 is not deployed, this embodiment does not generate macro-level evidence.

[0068] The intended effect of this embodiment is that it achieves the core risk identification and interpretation functions of this application with extremely low computational and storage costs. Although its prediction accuracy may be slightly lower than that of the teacher model in full-featured mode, and it lacks corroborating macroscopic evidence, it fully retains the ability of "white-box" decision-making, that is, each risk judgment can provide project-level, quantifiable micro-attribution evidence. Understandably, this deployment mode, which combines high cost-effectiveness and high adaptability, enables the advanced technology proposed in this application to be widely promoted and applied to medical insurance regulatory agencies at different levels and with different levels of informatization, providing great flexibility and scalability for realizing intelligent and transparent medical insurance supervision nationwide.

[0069] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A DRG risk identification method based on collaborative learning, characterized in that, This method is executed by a computer device and includes: Construct a collaborative learning framework that includes a first prediction model and a second prediction model, wherein the first prediction model is a prediction model for predicting the first DRG group based on case data, and the second prediction model is an interpretable model whose model parameters can form an interpretable knowledge representation. Train the first prediction model; The second prediction model is trained using knowledge distillation techniques, wherein the prediction results output by the first prediction model are used as soft labels to guide the parameter learning of the second prediction model, thereby obtaining the interpretable knowledge representation solidified into a global "DRG-medical item" importance matrix; Provide a historical case database; Obtain the standard weights corresponding to each DRG group; Obtain a test case containing the actual declared DRG group; The case to be tested is predicted. If the DRG grouping predicted based on the first prediction model is inconsistent with the actual declared DRG grouping, and the standard weight corresponding to the predicted first DRG grouping is lower than the standard weight corresponding to the actual declared DRG grouping, then a risk label is added to the case to be tested. Generate dual evidence for cases with risk labels, the dual evidence including: Based on the aforementioned "DRG-Health Item" importance matrix, the health items of the cases with risk labels are analyzed to generate first evidence for item-level attribution; and The historical case database is used to retrieve case sets that are similar to high-risk cases in terms of medical resource consumption patterns, in order to generate a second piece of evidence that supports the consensus of the group.

2. The method according to claim 1, characterized in that, The first prediction model is a deep neural network model based on the Transformer architecture; The training of the first prediction model includes: Pre-training of a mask-based language model on medical project sequences.

3. The method according to claim 1, characterized in that, The second prediction model is a linear model, and the parameters of the linear model are the importance matrix of the "DRG-medical items".

4. The method according to claim 1, characterized in that, The steps for generating the second piece of evidence include: The case data is encoded into a case feature vector using the first prediction model; A similarity search is performed on the case feature vectors to retrieve the set of similar cases.

5. The method according to claim 1, characterized in that, The method further includes: Calculate a comprehensive risk score for cases with risk labels.

6. The method according to claim 5, characterized in that, The comprehensive risk score is calculated based on one or more risk factors; The risk factors include at least one of the following: model prediction confidence output by the first prediction model, economic difference driving force calculated based on the standard weight difference between the predicted first DRG group and the actual declared DRG group, and similar case corroboration calculated based on the number of similar case sets retrieved from the historical case database or similarity scores.

7. A DRG risk identification device, characterized in that, The system includes a processor and a memory, the memory storing a computer program, a historical case database, and standard weights corresponding to each DRG group. When the computer program is executed by the processor, it implements the following: The model building module is used to build a collaborative learning framework that includes a first prediction model and a second prediction model, wherein the first prediction model is a high-precision prediction model and the second prediction model is an interpretable model. The model training module is used to train the first prediction model; The knowledge distillation module is used to train the second prediction model through knowledge distillation technology to obtain an interpretable knowledge representation that is solidified into a global "DRG-medical item" importance matrix; The case determination module is used to add a risk label to the case to be tested if the DRG group predicted based on the first prediction model is inconsistent with the actual declared DRG group, and when comparing the standard weights stored in the memory, the standard weight corresponding to the predicted first DRG group is lower than the standard weight corresponding to the actual declared DRG group. The evidence generation module is used to generate dual evidence for cases identified as at risk. The dual evidence includes a first piece of evidence generated based on the importance matrix of the "DRG-medical items" and a second piece of evidence generated based on similarity retrieval of the historical case database.

8. The apparatus according to claim 7, characterized in that, When the computer program is executed by the processor, it also implements: The risk quantification module is used to calculate a comprehensive risk score for cases with risk labels.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images