Methods, media, and equipment for assessing hospital admission necessity based on a diagnosis-treatment mismatch model.

Abstract

This invention discloses a method, medium, and device for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model, belonging to the field of medical insurance fund supervision technology. The method includes: acquiring diagnostic coding data of the target case, extracting joint diagnostic feature vectors, secondary diagnostic severity load values, and objective severity dimensionality reduction features, and generating a continuous diagnostic severity score through weighted fusion; acquiring detailed inpatient expense data, extracting multidimensional treatment intensity indicators and comparing them with baselines of the same primary diagnostic code and comorbidity stratification level to generate a continuous treatment intensity score; calculating the percentile ranking of diagnostic severity and treatment intensity within a reference group of the same diagnostic severity level, and calculating the mismatch coefficient based on the mismatch relationship between the two; inputting the diagnostic severity score, treatment intensity score, and mismatch coefficient into an integrated learning model, and outputting a comprehensive score for the necessity of hospital admission. This invention enables penetrating and accurate identification of low-standard hospital admission behavior, improving the accuracy of detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent supervision of medical insurance funds, specifically to a method, medium, and equipment for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model. Background Technology

[0002] Currently, with the full implementation of DRG / DIP payment methods, under-standard admission has become one of the most frequent types of medical behavioral distortions. Under-standard admission refers to medical institutions admitting patients without clear admission criteria, whose treatment could have been completed in outpatient clinics or day wards, in order to obtain DRG / DIP disease group surplus. Unlike practices such as splitting hospitalizations or over-coding, which require manipulation of coding or treatment processes, under-standard admission has the lowest operational cost and risk—hospitals only need to lower the admission standards without altering medical records or coding.

[0003] Existing low-standard hospital admission detection technologies mainly fall into three categories: The first category is based on single-dimensional statistical threshold methods, which screen for admission based on single indicators such as length of hospital stay or cost ratio. However, this method suffers from serious dimensionality limitations; a single rule can only capture one manifestation of low-standard admission, and the normal cost structure varies greatly across different diseases, leading to a large number of false alarms due to a uniform statistical threshold. The second category is based on disease-specific benchmark comparison methods, which determine anomalies by calculating the static deviation of each cost dimension from the baseline of the same disease. However, this method only focuses on deviations in "account numbers" and fails to establish a non-linear matching relationship between the disease severity implied by the diagnostic code and the actual treatment intensity reflected by the cost structure, making it impossible to make a penetrating judgment from the perspective of "whether the patient actually needs hospitalization." The third category is based on semantic analysis of medical record text, which uses NLP technology to analyze electronic medical record text to determine admission criteria. However, this method suffers from serious data availability problems—medical insurance departments often struggle to obtain complete medical record texts in actual supervision, and the writing quality of medical records varies greatly among different hospitals, resulting in poor model generalization ability and a significant decrease in cross-institutional transfer accuracy.

[0004] In summary, existing technologies lack a method to accurately identify low-standard hospital admissions at the case level by quantitatively modeling the nonlinear matching relationship between diagnostic severity and treatment intensity, based solely on detailed medical expense data and diagnostic coding data. Summary of the Invention

[0005] In view of the above problems, the present invention provides a technical solution for hospital admission necessity assessment based on a diagnosis-treatment mismatch model, which solves the following technical problem: how to achieve penetrating and accurate identification of low-standard hospital admission behavior at the case level by establishing a nonlinear mismatch model between diagnosis severity and treatment intensity, based solely on medical expense details and diagnostic coding data in the absence of medical record text data.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model, the method comprising:

[0007] S1: Obtain the diagnostic coding data of the target case. The diagnostic coding data includes a set of primary diagnostic codes and a set of secondary diagnostic codes. The primary diagnostic codes are decomposed into multiple levels and concatenated into a joint diagnostic feature vector. The set of secondary diagnostic codes is matched with a comorbidity index dictionary, and the secondary diagnostic severity load value is calculated based on the matching result. Objective severity indicators are aggregated and statistically analyzed by diagnostic category based on historical hospitalization data. The objective severity indicators are mapped to a high-dimensional space through kernel principal component analysis. Dimensionality reduction features that characterize the complexity of the disease are extracted. The joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused, and then normalized to the [0,1] interval through an activation function to generate a continuous diagnostic severity score.

[0008] S2: Obtain detailed inpatient cost data for the target case, extract multiple dimensions of treatment intensity indicators, use the primary diagnosis code and secondary diagnosis severity load values ​​as indexes, retrieve historical baseline datasets with the same primary diagnosis code and the same comorbidity stratification level, transform the treatment intensity indicators of each dimension into deviation features relative to the historical baseline with the same primary diagnosis code and the same comorbidity stratification level, and generate normalized continuous treatment intensity scores through nonlinear fusion model mapping.

[0009] S3: Select a group of historical cases from the historical case database whose diagnostic severity continuous scores are within the same preset threshold range as the target case as a reference group. Within the reference group, obtain the first percentile ranking of the target case's diagnostic severity continuous score and the second percentile ranking of the target case's treatment intensity continuous score. Calculate the nonlinear mismatch coefficient based on the first percentile ranking and the second percentile ranking.

[0010] S4: Input the multidimensional features, including at least the continuous score of diagnostic severity, the continuous score of treatment intensity, and the nonlinear mismatch coefficient, into the integrated learning model, output a comprehensive score representing the necessity of hospitalization, and output the assessment result of the necessity of hospitalization based on the comprehensive score.

[0011] Furthermore, in step S1, the joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality-reduced features are weighted and fused, then normalized to the [0,1] interval using an activation function to generate a continuous diagnostic severity score, specifically including:

[0012] The joint diagnostic feature vector is input into the first fully connected network and mapped to the first hidden vector. The first hidden vector is concatenated with the secondary diagnostic severity load value and then input into the second fully connected network to generate the fused feature vector.

[0013] The fused feature vector and the dimensionality-reduced feature are attention-weighted to output a weighted fusion vector. The weighted fusion vector is then mapped to the [0,1] interval using a Sigmoid activation function to generate the continuous score for the diagnostic severity.

[0014] Furthermore, prior to step S1, the following is also included:

[0015] Obtain the outpatient visit record sequence of the target case within a preset time window before admission, input the outpatient visit record sequence into a hidden Markov model, the state space of the hidden Markov model includes several preset outpatient visit modes, map the outpatient visit record sequence into an outpatient trajectory state sequence, and extract the transition probability matrix and dwell time distribution of the outpatient trajectory state sequence as outpatient trajectory features.

[0016] The outpatient trajectory features, the joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused together to generate the continuous diagnostic severity score.

[0017] Furthermore, the procedure after step S2 and before step S3 also includes:

[0018] The outpatient visit cost data and outpatient treatment item data of the target case within a preset time window before admission are obtained. The outpatient visit cost data and outpatient treatment item data are input into the counterfactual treatment intensity prediction model. The counterfactual treatment intensity prediction model is trained with historical data of outpatient cases that were not admitted to the hospital with the same primary diagnosis code and the same comorbidity stratification level. The model outputs a predicted treatment intensity score that represents the target case if it had not been admitted to the hospital but had received outpatient treatment.

[0019] The continuous score of the treatment intensity of the target case is compared with the predicted treatment intensity score to calculate the treatment intensity increment, and the treatment intensity increment is input as an additional feature into the ensemble learning model.

[0020] Furthermore, in step S2, the treatment intensity indicators of each dimension are transformed into deviation characteristics relative to the historical baseline of the same primary diagnosis code and the same comorbidity stratification level, specifically including:

[0021] The treatment intensity indicators of each dimension are transformed by an equal logarithmic ratio and mapped to an unconstrained real number space to obtain the transformed treatment intensity vector. The transformed treatment intensity vector is then input into an autoencoder model, which is trained using cost data from the historical baseline dataset that encodes the same primary diagnosis and is stratified at the same level of comorbidities. The reconstruction error between the treatment intensity vector and the autoencoder reconstruction output is calculated as the deviation feature.

[0022] Furthermore, the actual length of hospital stay and actual cost consumption index of the target case after admission are obtained. The actual length of hospital stay and actual cost consumption index are used as feedback signals. The evaluation error is calculated based on the feedback signals. The online gradient descent algorithm is used to update the weight parameters of the weighted fusion in S1 and the boundary values ​​of the preset threshold range in S3 according to the evaluation error, so that the continuous score of diagnostic severity and the comprehensive score of admission necessity of subsequent cases are dynamically adjusted according to the feedback of actual diagnosis and treatment results.

[0023] Furthermore, prior to S3, it also includes:

[0024] S101: Obtain M historical admission records of the target case within a preset historical time window, where M is an integer greater than or equal to 1. Extract the corresponding historical diagnostic severity continuous score and historical treatment intensity continuous score from each historical admission record, and construct a longitudinal score sequence for diagnostic severity. longitudinal scoring sequence of diagnosis and treatment intensity ,in, This represents the continuous score of diagnostic severity corresponding to the i-th historical hospital admission. This represents the continuous score of treatment intensity corresponding to the i-th historical hospital admission;

[0025] S102: Concatenate the longitudinal score sequence D of diagnostic severity and the longitudinal score sequence T of treatment intensity to generate a joint longitudinal score vector. The longitudinal score joint vector V is input into a time encoder, which is either a Transformer encoder based on a self-attention mechanism or a recurrent neural network based on a gated recurrent unit, and outputs the hidden vector H of the patient's diagnosis and treatment trajectory.

[0026] S103: Construct a contrastive learning pre-training task, construct positive sample pairs and negative sample pairs from the historical case database, wherein the positive sample pairs are longitudinal score joint vector pairs of two adjacent historical admission records of the same patient, and the negative sample pairs are longitudinal score joint vector pairs of different patients with the same primary diagnosis code. The temporal encoder is pre-trained using the InfoNCE or NT-Xent contrastive loss function to reduce the latent vector distance of the positive sample pairs and increase the latent vector distance of the negative sample pairs.

[0027] S104: Record the continuous score of diagnostic severity and the continuous score of treatment intensity for the target cases in the current admission cycle as follows: and Construct the joint longitudinal score vector for the current admission cycle. ,Will Input the pre-trained temporal encoder to generate the latent vector of the current diagnosis and treatment trajectory. ;

[0028] S105: Calculate the hidden vector of the current treatment trajectory. The Euclidean or cosine distance between the patient's treatment trajectory latent vector H and the longitudinal mismatch distance is used as the longitudinal mismatch distance, which represents the degree of deviation of the current hospitalization cycle's treatment mode from the patient's own historical trajectory.

[0029] In step S4, the longitudinal mismatch distance is input as an additional feature into the ensemble learning model.

[0030] Furthermore, prior to step S4, the following steps are also included:

[0031] S301: Extract a training sample set from the historical case database. Each training sample includes a continuous score for diagnostic severity, a continuous score for treatment intensity, a nonlinear mismatch coefficient, the patient's age, the patient's gender, the route of admission, the admitting department, the actual length of hospital stay, the actual total hospitalization cost, and the actual admission outcome. The actual admission outcome is a binary label, with a value of 1 indicating a reasonable admission and a value of 0 indicating a low-standard admission.

[0032] S302: The continuous score of diagnostic severity, continuous score of treatment intensity, nonlinear mismatch coefficient, patient age, patient gender, admission route, and admission department in the training sample set are used as a set of candidate causal variables. The actual hospital admission outcome is used as the target variable Y. A causal discovery algorithm based on conditional independence test is used to calculate the causal strength between each variable in the candidate causal variable set X and the target variable Y, and a directed acyclic causal graph structure G is learned.

[0033] S303: Based on the directed acyclic causal graph structure G, identify variables that have a direct causal edge with the target variable Y as a set of causal parent nodes P(Y), and each variable in the set of causal parent nodes P(Y) is a direct causal driving factor of the admission outcome.

[0034] S304: Using the variables in the set of causal parent nodes P(Y) as input and the actual hospital admission outcome Y as output, construct a structural causal model SCM. The structural causal model SCM includes structural equations and noise distribution. Parameters of the structural causal model SCM are estimated by maximum likelihood estimation or variational inference.

[0035] S305: Obtain the current set of observations of the causal parent node of the target case, take the continuous score of the treatment intensity of the target case as the intervention variable, set the counterfactual intervention value as the preset quantile of the continuous score of treatment intensity or the median of the corresponding dimension in the historical baseline dataset, calculate the counterfactual admission necessity probability of the target case under the counterfactual intervention value through the structural causal model SCM, and take the difference between the counterfactual admission necessity probability and the admission necessity probability under the current observation value as the causal effect estimate.

[0036] When performing step S4, the causal effect estimate is used as the input feature of the ensemble learning model, or the causal effect estimate and the nonlinear mismatch coefficient are weighted and fused together and then used as the input feature of the ensemble learning model.

[0037] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the hospital admission necessity assessment method based on a diagnosis-treatment mismatch model as described in the first aspect of the present invention.

[0038] In a third aspect, the present invention provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when executed by the processor, the computer program implements the hospital admission necessity assessment method based on the diagnosis-treatment mismatch model as described in the first aspect of the present invention.

[0039] Unlike existing technologies, the above-mentioned technical solution involves a method, medium, and equipment for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model, belonging to the field of medical insurance fund supervision technology. The method includes: acquiring diagnostic coding data of the target case, extracting joint diagnostic feature vectors, secondary diagnostic severity load values, and objective severity dimensionality reduction features, and generating a continuous diagnostic severity score through weighted fusion; acquiring detailed inpatient expense data, extracting multidimensional treatment intensity indicators and comparing them with baselines of the same primary diagnostic code and comorbidity stratification level to generate a continuous treatment intensity score; calculating the percentile ranking of diagnostic severity and treatment intensity within a reference group of the same diagnostic severity level, and calculating the mismatch coefficient based on the mismatch relationship between the two; inputting the diagnostic severity score, treatment intensity score, and mismatch coefficient into an integrated learning model, and outputting a comprehensive score for the necessity of hospital admission. This invention can achieve penetrating and accurate identification of low-standard hospital admission behavior, improving the accuracy of detection.

[0040] The above description of the invention is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical solution of the present invention and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of the present invention easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of the present invention. Attached Figure Description

[0041] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on the present invention.

[0042] In the accompanying drawings of the instruction manual:

[0043] Figure 1 This is a flowchart of the hospital admission necessity assessment method based on a diagnosis-treatment mismatch model as described in the first exemplary embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of an electronic device according to an exemplary embodiment of the present invention;

[0045] The reference numerals used in the above figures are explained as follows:

[0046] 10. Electronic equipment; 101. Processor; 102. Storage medium. Detailed Implementation

[0047] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that can be achieved by this invention, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this invention, and are therefore only examples, and should not be used to limit the scope of protection of this invention.

[0048] In the first aspect, such as Figure 1 As shown, this application provides a method for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model, the method comprising:

[0049] S1: Obtain the diagnostic coding data of the target case. The diagnostic coding data includes a set of primary diagnostic codes and a set of secondary diagnostic codes. The primary diagnostic codes are decomposed into multiple levels and concatenated into a joint diagnostic feature vector. The set of secondary diagnostic codes is matched with a comorbidity index dictionary, and the secondary diagnostic severity load value is calculated based on the matching result. Objective severity indicators are aggregated and statistically analyzed by diagnostic category based on historical hospitalization data. The objective severity indicators are mapped to a high-dimensional space through kernel principal component analysis. Dimensionality reduction features that characterize the complexity of the disease are extracted. The joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused, and then normalized to the [0,1] interval through an activation function to generate a continuous diagnostic severity score.

[0050] S2: Obtain detailed inpatient cost data for the target case, extract multiple dimensions of treatment intensity indicators, use the primary diagnosis code and secondary diagnosis severity load values ​​as indexes, retrieve historical baseline datasets with the same primary diagnosis code and the same comorbidity stratification level, transform the treatment intensity indicators of each dimension into deviation features relative to the historical baseline with the same primary diagnosis code and the same comorbidity stratification level, and generate normalized continuous treatment intensity scores through nonlinear fusion model mapping.

[0051] S3: Select a group of historical cases from the historical case database whose diagnostic severity continuous scores are within the same preset threshold range as the target case as a reference group. Within the reference group, obtain the first percentile ranking of the target case's diagnostic severity continuous score and the second percentile ranking of the target case's treatment intensity continuous score. Calculate the nonlinear mismatch coefficient based on the first percentile ranking and the second percentile ranking.

[0052] S4: Input the multidimensional features, including at least the continuous score of diagnostic severity, the continuous score of treatment intensity, and the nonlinear mismatch coefficient, into the integrated learning model, output a comprehensive score representing the necessity of hospitalization, and output the assessment result of the necessity of hospitalization based on the comprehensive score.

[0053] In this embodiment, the diagnostic coding data refers to the standardized ICD-10 code on the medical insurance medical record homepage, which is divided into a single primary diagnosis (the core disease of this hospitalization) and multiple secondary diagnoses (comorbid underlying diseases and complications). It is the only standardized data source for quantifying the severity of the condition and does not require medical record text.

[0054] The diagnostic category refers to the first three digits of the ICD-10 code, which are used to aggregate statistical baselines for diseases of the same major category and avoid the problem of sparse samples in the detailed coding.

[0055] The Comorbidity Index Dictionary (CCI Dictionary) is a clinically standardized Charlson comorbidity index reference table. Each type of chronic severe illness and complication has a fixed clinical weight, which is used to quantify the patient's basic physiological vulnerability.

[0056] Objective severity index refers to four-dimensional proxy indicators obtained from the statistical analysis of all historical hospitalization data in the region: average length of stay (ALOS), average total cost (AEXP), case mortality rate (MORT), and ICU admission rate (ICUR), which objectively reflect the overall severity of the corresponding disease.

[0057] Kernel principal component analysis (KPCA) is a nonlinear dimensionality reduction algorithm that uses radial basis function (RBF) kernels. Unlike linear PCA, it can capture nonlinear clinical associations among four types of objective indicators and output two-dimensional continuous disease urgency features.

[0058] Diagnostic severity continuous score (DPS) The value range of ) is fixed at [0,1]. The higher the value, the more severe the diagnosis code indicates and the stronger the theoretical indication for hospitalization. The lower the value, the milder the condition, and theoretically, hospitalization is not required.

[0059] Treatment intensity indicators refer to four-dimensional quantitative indicators derived from the breakdown of inpatient tertiary cost details, including: core treatment intensity, intravenous medication rate, etc. The intensity of monitoring and observation, and the density of high-end examinations and tests, among which the intravenous medication rate is the core hard indicator for distinguishing between outpatient mild cases and inpatient severe cases in this invention.

[0060] Comorbidity stratification refers to the classification of secondary diagnostic loads ( By segmenting and stratifying patients, we can achieve accurate baseline comparison for patients with the same disease type and the same underlying disease burden, thus eliminating confounding bias in the baselines of elderly patients with multiple diseases and those with only mild symptoms.

[0061] Historical baseline datasets refer to a collection of normal inpatient samples from hospitals of the same level, ICD category, and comorbidity stratification within a region, used to calculate the mean values ​​of various diagnostic and treatment indicators. Standard deviation .

[0062] Deviation characteristics refer to the standardized deviation of a single-dimensional diagnostic and treatment indicator relative to the baseline, eliminating interference from the natural differences in cost structures for different diseases.

[0063] Continuous score of treatment intensity ( The value range of ) is fixed at [0,1]. A higher value indicates more substantial invasive and critical medical interventions during hospitalization, while a lower value indicates only oral medication and short-term observation, which is consistent with the characteristics of outpatient treatment.

[0064] The preset threshold range refers to the pre-defined continuous scoring interval, which selects historical cases with the same severity of illness as the target case, to achieve "horizontal comparison of patients with the same degree of severity" and avoid mixing severe and mild cases.

[0065] The percentile ranking refers to the percentile of the target case's diagnostic severity within the reference group, sorted from smallest to largest. The higher the value, the more severe the condition is within the same disease range.

[0066] Second percentile ranking ( The conditional quantile refers to the percentile of treatment intensity within a population after fixing the severity range of a diagnosis. The lower the value, the more significantly the actual treatment intensity is insufficient for the same condition.

[0067] Nonlinear mismatch coefficient ( The calculation formula for ) is as follows:

[0068] ;

[0069] When diagnostic criteria are emphasized but actual treatment is weak, the coefficient exhibits non-linear amplification, accurately capturing low-standard admission and high-standard coding behavior.

[0070] Ensemble learning models refer to pre-trained gradient boosting trees (XGBoost / LightGBM) that non-linearly fit complex relationships among multiple features and output a final comprehensive score indicating the necessity of hospital admission. ).

[0071] The results of the admission necessity assessment refer to the three-level early warning system based on a comprehensive score: Red Alert ( <0.3), highly suspected low-standard admission), yellow alert (0.3≤ <0.6), manual verification), green mark ( (≥0.6), compliant and reasonable hospital admission).

[0072] In step S1, the system retrieves all ICD-10 diagnostic codes for the target case from the medical insurance medical record homepage database. First, it performs string slicing on the main diagnostic codes, breaking them down into four layers of coding features: chapter, category, subcategory, and detail. These layers are then concatenated dimensionally to generate a dense joint diagnostic feature vector. This approach addresses the technical shortcomings of sparse and feature-ineffective rare detailed coding samples by using multi-granularity feature concatenation. For all secondary diagnostic codes, each code is matched against the built-in CCI comorbidity weight dictionary, and the weights of all comorbidities are summed and then smoothed using a logarithmic smoothing formula (…). The secondary diagnostic load is calculated, and the logarithmic transformation is used to suppress the excessive amplification of the weights of multiple comorbidities, thus stably characterizing the patient's long-term underlying disease vulnerability level.

[0073] The system retrieves the full historical inpatient baseline dataset for the region, groups it by ICD category, and calculates the four-dimensional objective severity surrogate vectors for ALOS, AEXP, MORT, and ICUR for each group. Because the four indicators have complex nonlinear coupling relationships in clinical scenarios, linear dimensionality reduction would lose key critical nonlinear correlation information. Therefore, RBF kernel KPCA is introduced to first map the four-dimensional vector to a high-dimensional feature manifold, and then perform dimensionality reduction to extract two-dimensional features representing the complexity of the disease, separately representing the urgency of disease treatment. The multi-level joint diagnostic vector, log-smoothed secondary diagnostic payload, and KPCA-reduced disease features are uniformly weighted and fused. The weights are obtained from pre-training with region-labeled samples. The fusion result is input into a Sigmoid activation function, forcibly mapped to a continuous interval of 0-1, and outputs a standardized diagnostic severity score. This quantifies the intensity of patients' theoretical hospital admission needs from a diagnostic semantic perspective.

[0074] In step S2, the system reads the detailed inpatient tertiary expense list and drug administration route code, and automatically aggregates and calculates four core treatment agency indicators, including the proportion of treatment cost / surgery cost / anesthesia cost constituting the core treatment intensity; the ratio of intravenous drug cost to total drug cost yields the intravenous drug use rate. Nursing fees, monitoring fees, and daily bed fees are combined to calculate the intensity of monitoring and observation; the proportion of high-end examination fees such as MRI and CT scans forms the examination density indicator. Intravenous medication rate serves as a hard differentiating indicator; outpatient mild cases almost always receive only oral medication, while inpatient essential cases generally receive continuous intravenous infusion, naturally eliminating the interference of differences in the absolute value of total hospital costs.

[0075] The system uses "ICD diagnostic category + secondary diagnostic load stratification" as a composite index key to accurately retrieve the stratified historical baseline database. It standardizes the standard deviation of each of the four diagnostic and treatment indicators, generating Z-features of deviation for each dimension. This completely resolves the baseline distortion problem caused by the inherent cost structure differences among internal medicine, surgery, and chronic diseases. All standardized deviation features are concatenated into a feature vector, input into a pre-trained logistic regression nonlinear fusion operator, and normalized to the [0,1] interval using the Sigmoid activation function, outputting a unified diagnostic and treatment intensity score. This quantifies the actual medical intervention efforts implemented during this hospitalization.

[0076] In step S3, the system traverses the historical case database and filters all diagnostic severity levels. Samples falling within the pre-defined scoring range of the target cases are used to construct a horizontal reference population with perfectly matched disease severity, preventing the invalidation of comparisons caused by mixed samples of mild and severe cases. Within this homogeneous reference population, the empirical cumulative distribution function (ECDF) is used to calculate two quantiles: one is the global percentile of the target case's diagnostic severity within the population (…). Secondly, the conditional quantiles of treatment intensity after fixing the diagnostic score interval. Then, the core nonlinear mismatch calculation formula is entered, and when the case diagnosis code is marked as critical ( While the mismatch coefficient approaches 1, treatment intervention is extremely weak under similar conditions (conditional quantile approaches 0). The mismatch coefficient will amplify non-linearly and exponentially. This method accurately captures hidden under-standard admission behaviors such as "deliberately inflating diagnoses and admitting mild cases to obtain DRG surpluses," which is different from the shortcomings of traditional linear single-point outliers that only compare cost values ​​and cannot match the condition and treatment logic. If the mismatch coefficient exceeds the 90th percentile threshold for this disease, a kernel Shapley value attribution decomposition is triggered simultaneously to distinguish whether the mismatch originates from over-reporting of diagnoses or under-reporting of treatment, and to divert to different audit and handling paths.

[0077] The nonlinear mismatch coefficient was calculated. Then, the system pre-stores the 90th percentile of the mismatch coefficient under the corresponding diagnostic category as the anomaly judgment threshold; when When the threshold is exceeded, the KernelShapley value attribution decomposition operator is automatically activated. The operator uses a virtual alliance game model to quantify the marginal contributions of the diagnostic feature side and the treatment intensity feature side to the mismatch coefficient shift, respectively, and outputs the diagnostic side contribution value. Contribution value of treatment side .like The system automatically marks the case as a violation of the code type and pushes a code compliance review task; if The system marks patients as being admitted to hospitals below the standard and pushes on-site audit tasks for admission criteria to automatically divert two types of medical insurance misconduct and output regulatory and disposal clues in a differentiated manner.

[0078] In step S4, the system collects three core basic features: diagnostic severity score, treatment intensity score, and nonlinear mismatch coefficient. The complete feature vector is then fed into a gradient boosting ensemble model trained on offline labeled samples. The model learns the nonlinear coupling relationships between multiple features and outputs a single comprehensive hospital admission necessity score. The system performs graded interception and judgment according to three intervals, and directly outputs standardized assessment results. A red warning indicates a highly suspicious low-standard admission, a yellow warning indicates an abnormality that requires manual review, and a green mark indicates that the hospitalization fully meets the reasonable admission criteria.

[0079] The aforementioned solution overcomes the limitations of existing single-indicator threshold and static cost baseline comparison technologies by employing a dual-dimensional quantitative matching paradigm of disease diagnosis and actual treatment. It can be implemented solely based on standardized medical insurance settlement and medical record data, eliminating data barriers associated with electronic medical record acquisition. By using KPCA to non-linearly extract the critical characteristics of disease types and the logarithmic loading of the CCI to quantify comorbidities, it achieves refined stratification of disease conditions. Based on conditional quantiles of the same disease group, it constructs a non-linear mismatch coefficient, which can identify hidden under-standard cases completely missed by traditional methods, such as "normal book costs but pre-outpatient examinations." The integrated model combines multi-dimensional quantitative indicators, balancing sensitivity and specificity in single-case identification. The tiered early warning output directly connects to the medical insurance audit process, significantly reducing the workload of manual screening. The mismatch indicators, coupled with Shapley attribution, can automatically distinguish between under-standard admission and over-standard coding, supporting differentiated enforcement by regulatory authorities.

[0080] In some embodiments, in step S1, the joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction feature are weighted and fused, then normalized to the [0,1] interval using an activation function to generate a continuous diagnostic severity score, specifically including:

[0081] The joint diagnostic feature vector is input into the first fully connected network and mapped to the first hidden vector. The first hidden vector is concatenated with the secondary diagnostic severity load value and then input into the second fully connected network to generate the fused feature vector.

[0082] The fused feature vector and the dimensionality-reduced feature are attention-weighted to output a weighted fusion vector. The weighted fusion vector is then mapped to the [0,1] interval using a Sigmoid activation function to generate the continuous score for the diagnostic severity.

[0083] In this embodiment, the first fully connected network refers to a shallow dimensionality reduction neural network that takes a high-dimensional, multi-level diagnostic coding vector as input and outputs a low-dimensional dense hidden vector, filtering out redundant and noisy features within the coding and reducing computational load.

[0084] The first latent vector refers to the low-dimensional representation after high-dimensional diagnostic coding compression, which retains the core semantics of the disease level and eliminates meaningless and redundant coding dimensions.

[0085] The second fully connected network refers to a cross-modal fusion network that enables feature interaction between discrete coding features and continuous comorbidity load values.

[0086] Attention-weighted mechanisms refer to adaptive weight allocation operators that dynamically adjust the weight ratio of KPCA dimensionality-reduced critical features and coding fusion features based on the current importance of disease characteristics. Fixed weights cannot adapt to the clinical patterns of different diseases.

[0087] The weighted fusion vector refers to the unified diagnostic feature representation after adaptive attention weight allocation, which integrates all information from three categories: coding level, comorbidity burden, and objective severity of disease.

[0088] The original joint diagnostic feature vector is an ultra-high-dimensional sparse vector composed of multiple codes. Directly participating in the fusion would lead to a large amount of invalid computation and introduce coding noise. Therefore, it is first fed into the first fully connected network for dimensional compression mapping to generate the first hidden vector with a significantly reduced dimension. Only the core semantic information that distinguishes the severity of the disease within the ICD coding is retained, and redundant coding features without discrimination are filtered out.

[0089] The first hidden vector is a dense vector derived from discrete coding, and the secondary diagnostic load L cc Since the two types of heterogeneous data are single continuous values ​​and have different dimensions, they cannot be directly weighted. Therefore, after performing vector concatenation, the data is fed into a second fully connected network. The network automatically learns the clinical association between the encoded features and the burden of comorbidities, and outputs a unified dimension fused feature vector to complete the deep interactive fusion of the two types of heterogeneous diagnostic information.

[0090] The fusion feature vector only reflects the patient's own diagnostic coding information and lacks the objective criticality pattern of the disease obtained from regional historical big data statistics. Therefore, an attention weighting mechanism is introduced to dynamically assign adaptive weights to the fusion vector and the 2-dimensional reduced disease features output by KPCA: automatically increase the weight of objective criticality features for acute severe diseases and increase the weight of comorbidity load for chronic underlying diseases. This avoids the defect that fixed weighting cannot adapt to all diseases and outputs a weighted fusion vector that takes into account both individual diagnosis and group disease patterns.

[0091] Finally, the weighted fusion vector is input into the Sigmoid activation function to map the unbounded real numbers to the standard continuous interval of 0-1, generating a stable and horizontally comparable continuous score for the diagnostic severity, ensuring that the quantitative scale of the condition is consistent for all cases.

[0092] The above scheme addresses the low computational efficiency and noise interference issues caused by the high-dimensional sparsity of multi-level ICD encoding through hierarchical compression and fusion of heterogeneous diagnostic features using a two-level fully connected network. It achieves deep interaction between encoded semantics and comorbidity burden. Adaptive attention weighting replaces fixed weighting coefficients, dynamically adjusting the contribution weights of different diagnostic dimensions according to the clinical characteristics of the disease. This significantly improves the accuracy of disease quantification for different diseases and avoids systematic biases in the determination of acute, severe, and chronic diseases caused by uniform fixed weights. The entire fusion process is based entirely on numerical computation, eliminating the need for manually setting feature weight thresholds. It adapts to scenarios with dynamically updated regional medical insurance data, outputting standardized interval scores. This provides a stable and unified baseline for subsequent mismatch coefficient calculations, effectively reducing the probability of misjudgment caused by coarse feature fusion methods.

[0093] In some embodiments, the method further includes the following steps prior to step S1:

[0094] Obtain the outpatient visit record sequence of the target case within a preset time window before admission, input the outpatient visit record sequence into a hidden Markov model, the state space of the hidden Markov model includes several preset outpatient visit modes, map the outpatient visit record sequence into an outpatient trajectory state sequence, and extract the transition probability matrix and dwell time distribution of the outpatient trajectory state sequence as outpatient trajectory features.

[0095] The outpatient trajectory features, the joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused together to generate the continuous diagnostic severity score.

[0096] In this embodiment, the preset time window refers to the system's configurable pre-admission observation period, which is 30 days by default. It is used to capture all outpatient and emergency visits of patients before hospitalization, covering the typical violation time sequence scenario of "completing a full set of outpatient examinations before hospitalization".

[0097] Outpatient visit record sequence refers to a time-series event chain arranged in ascending order of visit time. Each event carries the visit time, treatment items, and cost details, completely reconstructing all medical behaviors before hospitalization.

[0098] Hidden Markov Model (HMM) is a time-series probability model where the observed values ​​are outpatient items / costs and the hidden states represent clinical treatment intentions. It can decode the underlying motivations for seeking medical treatment from surface outpatient data.

[0099] The preset outpatient visit mode (hidden state) refers to five standardized clinical intention states, including: routine follow-up, diagnostic examination, waiting for diagnosis, treatment decision, and emergency hospitalization.

[0100] Outpatient trajectory state sequence refers to the optimal hidden state time sequence obtained by decoding with the Viterbi algorithm, which restores the complete path of changes in the patient's medical intentions before hospitalization.

[0101] The transition probability matrix records the probability of switching between different diagnostic and treatment intentions. For example, "diagnosis and examination → direct hospitalization" corresponds to a typical low-standard hospitalization transfer path.

[0102] The duration of stay distribution refers to the statistical distribution of the duration of continuous medical visits for each intention state, which is used to distinguish between short-term single examinations and long-term repeated follow-ups.

[0103] Before performing S1 diagnostic feature extraction, the system retrieves all outpatient events within a preset window before the target case's admission from the outpatient settlement database. These events are then sorted by visit time from oldest to newest to construct an ordered temporal event chain, comprehensively capturing all pre-hospitalization diagnostic and treatment behaviors. This addresses the core deficiency of existing technologies that only analyze inpatient data and separate outpatient and inpatient scenarios. The system uses the treatment items and cost percentages of each outpatient record as HMM observation variables, and incorporates five types of clinical diagnostic and treatment intentions as unobservable hidden states. Through a Viterbi globally optimal decoding algorithm, discrete outpatient event sequences are mapped to continuous outpatient trajectory state sequences, intuitively reconstructing the underlying motivations of patients seeking medical care. Based on the decoded state sequences, two types of temporal features are statistically analyzed: a transition probability matrix between different intention states and a distribution of the duration of a single continuous visit for each state. These two indicators together constitute the multidimensional features of the outpatient trajectory. These outpatient temporal trajectory features are added as a new diagnostic dimension, merged with the fused encoded features, secondary diagnostic payloads, and KPCA disease-specific dimensionality reduction features, and uniformly fed into an attention-weighted fusion process to participate in the diagnostic severity score calculation. If the Hidden Markov Model (HMM) decodes the typical transfer path of "directly jumping to hospitalization while waiting for diagnostic examinations / confirmation, skipping the emergency hospitalization status," the outpatient trajectory features will significantly increase the tendency to under-standardize the disease assessment, causing the diagnostic severity score to be adjusted accordingly. This corrects the misjudgment of the disease caused by relying solely on static ICD coding from a temporal behavioral perspective. At the same time, this temporal feature can be linked to the logic of identifying pre-examination examinations: if outpatient events involve a large number of examination items and the cost of inpatient examinations is significantly lower than the baseline, the outpatient trajectory features will amplify the abnormality of the subsequent nonlinear mismatch coefficient, accurately identifying hidden under-standard admissions due to pre-examination examinations.

[0104] When constructing the outpatient visit record sequence, the system uses the admission time t0 as the benchmark and analyzes each outpatient event e. i Configure exponential time decay weights, the weight calculation formula is as follows: , The system allows for a configurable time decay factor, with a default value of 0.05. Outpatient treatment behaviors closer to the admission time receive higher weights, while the weight of remote follow-up events automatically decays. Simultaneously, a Dynamic Time Warping (DTW) algorithm is used to cluster all outpatient time-series events, classifying them into three standard visitation patterns: concentrated examination type, long-term follow-up type, and emergency outbreak type. If the clustering result is concentrated examination type, combined with HMM decoding of the "diagnostic examination directly jumps to hospitalization" state path, and the proportion of high-end examinations and tests during hospitalization is lower than 25% of the baseline for the same disease, the system generates a strong risk label for low-standard admission due to pre-examination transfer, reinforcing the abnormal weight of the case.

[0105] This embodiment adds pre-hospitalization outpatient time-series behavior as an auxiliary dimension for disease assessment, breaking through the limitations of traditional methods that rely solely on static diagnostic coding to assess admission criteria. It determines the necessity of hospitalization from the patient's complete treatment behavior chain. By decoding the hidden medical intention through HMM, it can accurately capture irregular time-series patterns such as "concentrated outpatient examinations and planned hospitalization" that traditional static algorithms cannot identify. Two quantitative features, transfer probability and residence duration, can quantify the degree of time-series anomalies and directly participate in the fusion calculation of diagnostic severity, allowing disease assessment to take into account both static coding and dynamic disease progression changes. The entire solution only retrieves routine outpatient settlement data, without requiring outpatient medical record text, has a low deployment threshold, and effectively improves the identification and recall rate of low-standard hospitalization due to transfer from pre-examination, reducing the risk of missed detection.

[0106] In some embodiments, the method further includes the following steps after step S2 and before step S3:

[0107] The outpatient visit cost data and outpatient treatment item data of the target case within a preset time window before admission are obtained. The outpatient visit cost data and outpatient treatment item data are input into the counterfactual treatment intensity prediction model. The counterfactual treatment intensity prediction model is trained with historical data of outpatient cases that were not admitted to the hospital with the same primary diagnosis code and the same comorbidity stratification level. The model outputs a predicted treatment intensity score that represents the target case if it had not been admitted to the hospital but had received outpatient treatment.

[0108] The continuous score of the treatment intensity of the target case is compared with the predicted treatment intensity score to calculate the treatment intensity increment, and the treatment intensity increment is input as an additional feature into the ensemble learning model.

[0109] In this embodiment, the counterfactual treatment intensity prediction model refers to a supervised regression model. The training set consists of compliant mild cases of the same disease type, the same stratification, and those treated only on the outpatient basis without hospitalization. The model takes outpatient expenses and treatment items as input and outputs standardized outpatient treatment intensity.

[0110] Non-hospitalized outpatient cases refer to compliant samples of patients diagnosed with the same disease in the region's history who completed treatment solely through outpatient follow-up / prescription / examination and had no hospitalization records. These cases represent the benchmark for the intensity of standard outpatient treatment for the disease.

[0111] The predicted treatment intensity score ranges from [0,1], simulating the maximum medical intervention that the target case can achieve by relying solely on outpatient services without hospitalization.

[0112] The incremental treatment intensity refers to the actual inpatient treatment intensity minus the counterfactual outpatient treatment intensity. The smaller the increment, the less additional substantial treatment is provided in inpatient care compared to outpatient care, which is highly consistent with the characteristics of low-standard admission.

[0113] In step S2, the actual inpatient treatment intensity is output. Before entering the S3 mismatch coefficient calculation stage, the system reuses the full data of outpatient expenses and treatment items from the pre-admission window as input variables for the counterfactual model. The counterfactual model is trained offline using compliant outpatient cases of the same ICD category and comorbidity stratification. It learns the intensity patterns of examinations, medications, and basic treatments that can be completed solely with outpatient resources, and can simulate "the maximum intervention intensity that can be achieved by outpatient treatment alone if the patient is not admitted to the hospital," solving the counterfactual inference problem of "unobservable outpatient alternatives" in admission assessment. The model output value range is also [0,1], which is a predicted outpatient treatment intensity score. It uses a unified quantitative scale with the inpatient score and can be directly calculated by difference. The difference between the inpatient treatment intensity and the counterfactual outpatient predicted intensity is calculated to obtain the treatment intensity increment: if the increment is close to 0, it means that no severe interventions that cannot be completed in the outpatient department are provided during hospitalization, and only outpatient oral medications and basic observations are repeated, which is a typical low-standard admission; if the increment is significantly high, it means that the hospital provides essential treatments that cannot be carried out in the outpatient department, such as intravenous infusion, surgery, and monitoring, and the admission criteria are sufficient. The incremental treatment intensity is treated as an independent additional feature and is combined with the diagnostic severity, treatment intensity, and nonlinear mismatch coefficient. This combination is then fed into the integrated learning model in step S4 to participate in the comprehensive score calculation. The model can autonomously learn the correlation between the incremental size and the risk of low-standard hospitalization, further enhancing its ability to distinguish between outpatient and inpatient settings across different scenarios.

[0114] After the integrated model outputs a comprehensive score indicating the necessity of hospital admission and triggers red and yellow level warnings, the system reverse-engineers the entire feature calculation chain, capturing the underlying database physical storage information corresponding to each anomaly judgment to form rigid physical row anchors. A single anchor completely includes the data source table name, database physical row RowID, trigger judgment field name, field original value, and corresponding algorithm rule name. The system encapsulates all anchors and model quantification indicators into a four-layer standardized JSON audit evidence message: the first layer, the decision summary layer, stores the comprehensive score, warning level, violation type, and total number of anchors; the second layer, the core signal layer, stores the diagnostic severity, mismatch coefficient, Shapley attribution, and differential reconciliation score; the third layer, the cross-scenario outpatient trajectory layer, stores the time decay weight sequence and HMM intent transfer path; and the fourth layer, the physical row number anchor layer, stores the index of the original records in the entire database. Before message output, a self-check for integrity is performed: if any judgment conclusion does not have a corresponding physical anchor point, the system will intercept the message output, ensuring that all risk conclusions can be drilled down to the original settlement, medical records, and outpatient data with one click, and directly used as evidence for medical insurance audit, thus avoiding the defect that pure mathematical models cannot provide evidence.

[0115] This embodiment introduces counterfactual inference logic to quantify the additional medical intervention value of hospitalization compared to outpatient services, addressing the technical shortcoming of traditional methods in determining whether hospitalization constitutes a non-essential value-added service. Model training relies solely on routine outpatient settlement data, incurring no additional data collection costs. A unified intensity score across a consistent value range ensures horizontal comparability of differences. The incremental treatment intensity, as a new input feature of the integrated model, provides direct quantitative evidence for determining low-standard hospitalization. Cases with incremental intensity approaching zero automatically receive increased risk weights, accurately distinguishing between "mild cases that can be cured on outpatient leave but are forcibly hospitalized" and "severe cases requiring hospitalization." This effectively reduces false alarms caused by relying solely on the internal cost structure of inpatient care, further improving the accuracy of identifying hidden low-standard hospitalization cases.

[0116] In some embodiments, in step S2, the treatment intensity indicators of each dimension are converted into deviation features relative to the historical baseline of the same primary diagnosis code and the same comorbidity stratification level, specifically including:

[0117] The treatment intensity indicators of each dimension are transformed by an equal logarithmic ratio and mapped to an unconstrained real number space to obtain the transformed treatment intensity vector. The transformed treatment intensity vector is then input into an autoencoder model, which is trained using cost data from the historical baseline dataset that encodes the same primary diagnosis and is stratified at the same level of comorbidities. The reconstruction error between the treatment intensity vector and the autoencoder reconstruction output is calculated as the deviation feature.

[0118] In this embodiment, the ILR transformation is a nonlinear orthogonal transformation operator designed for closed-component data such as cost percentages. It can remove the constraint of "dimension sum to 1" and map the percentages to an unconstrained Euclidean real space, while preserving the geometric distance of the original data.

[0119] The transformed diagnostic intensity vector refers to the multidimensional weight vector after eliminating closure interference through ILR transformation. Each dimension is independent of the others and can be directly used for unsupervised anomaly detection.

[0120] Disease-specific sparse autoencoders refer to shallow neural networks trained offline using only compliant normal inpatient cost samples of the same ICD category and comorbidity stratification. The encoding layer performs non-linear dimensionality reduction to learn the manifold surface of normal costs for that disease, and the decoding layer restores the standard cost structure.

[0121] Reconstruction error refers to the Euclidean distance between the input transformed vector and the decoded and restored vector of the autoencoder. The larger the value, the more serious the deviation of the current case cost structure from the historical normal pattern of the disease, and it is directly used as the deviation feature of the cost dimension.

[0122] After extracting four original treatment intensity indicators—core treatment intensity, intravenous medication rate, monitoring density, and proportion of high-end examinations—in step S2, these original indicators are proportion-based component data. Since there is a natural closed coupling relationship between their dimensions, directly calculating the standard deviation would lead to statistical distortion. The system first calls the ILR orthogonal transformation operator to construct an orthogonal basis matrix and perform a logarithmic mapping on the four-dimensional proportion data. This completely eliminates the spurious correlations caused by the closed constraints, outputting an unconstrained, dimensionally independent transformed treatment intensity vector.

[0123] The system then uses "major diagnostic category + comorbidity stratification" as the search key to retrieve the ILR transform vectors of all historical compliant inpatient samples under that stratification, completing the offline training of the sparse autoencoder. Specifically, the encoding layer learns the nonlinear manifold features of the normal cost distribution for that disease, while the decoding layer is responsible for reconstructing the standard cost vector. It only fits the normal treatment structure and cannot reconstruct the cost patterns of low-standard admission. The target case's ILR transform vector is input into the trained autoencoder. The model first reduces dimensionality to extract manifold features and then reconstructs the standard vector in reverse. The mean square error between the input vector and the reconstructed output vector is calculated as the reconstruction error. If a case is a low-standard admission, it will exhibit an asymmetric structure of "high examination ratio and low core treatment / intravenous medication ratio." The autoencoder cannot reconstruct this abnormal pattern, and the reconstruction error will be significantly amplified. This is used as a quantified deviation feature in the treatment intensity fusion model. This method replaces the traditional single-dimensional Z-score linear comparison, capturing the complex low-standard fingerprint of multi-dimensional joint shifts and distinguishing between two opposite types of cost anomalies: low-standard admission and over-treatment.

[0124] After completing the autoencoder reconstruction error calculation, the system constructs disease-specific normal manifold centers and historical low-standard admission case cluster centers based on the unconstrained feature vectors after ILR transformation, and calculates two sets of cosine distances in parallel: the distance between case features and normal manifold centers. ), case characteristics and distance from low-standard cluster centers ( Substitute into the differential reconciliation formula ( ) to obtain the differential reconciliation score; The larger the value, the closer the cost structure of the case is to the low-standard admission characteristic pattern. At the same time, a four-dimensional asymmetric abnormality fingerprint code is constructed: the examination and test dimension deviates from the baseline code by +1, and the core treatment, intravenous medication, and hospitalization duration are lower than the baseline code by -1. A feature fingerprint vector is generated, and the cosine of the vector angle is used to distinguish between two completely different types of cost abnormalities: low-standard admission (multi-dimensional reverse offset) and over-treatment (all-dimensional positive offset), so as to avoid confusion between the two types of violations and false alarms.

[0125] This embodiment addresses the statistical distortion problem of cost proportion component data through ILR transformation. It relies on a disease-specific autoencoder to learn a nonlinear normal cost manifold, using reconstruction error as a deviation feature. This overcomes the limitation of traditional linear comparisons, which can only identify single-dimensional numerical deviations. It can accurately capture the multi-dimensional reverse joint offset patterns unique to low-standard hospitalizations, effectively distinguishing between low-standard hospitalizations for mild cases and over-treatment for severe cases, significantly reducing screening false alarm rates. The entire process relies solely on routine medical insurance cost details, requiring no additional clinical text. The autoencoder is stratified and trained according to disease type to adapt to the natural cost differences of different diseases, providing a more accurate quantitative basis for structural anomalies in treatment intensity scoring and improving the sensitivity of subsequent nonlinear mismatch coefficient discrimination.

[0126] In some embodiments, the actual length of hospital stay and the actual cost index of the target case after admission are obtained. The actual length of hospital stay and the actual cost index are used as feedback signals. The evaluation error is calculated based on the feedback signals. An online gradient descent algorithm is used to update the weight parameters of the weighted fusion in S1 and the boundary values ​​of the preset threshold range in S3 according to the evaluation error, so that the continuous score of diagnostic severity and the comprehensive score of admission necessity of subsequent cases are dynamically adjusted according to the feedback of actual diagnosis and treatment results.

[0127] In this embodiment, the actual length of hospital stay and the cost consumption index refer to the actual labels of the treatment outcome. If the length of hospital stay is too short or the cost consumption index is much lower than the baseline of the disease group, it means that the actual intervention in this hospitalization is insufficient, and it can be used as a feedback basis for the model to evaluate the correctness of the model.

[0128] Assessment error refers to the difference between the model's output comprehensive hospital admission necessity score and the actual treatment outcome label, quantifying the judgment bias under the current model parameters.

[0129] Online gradient descent algorithm refers to an incremental parameter update optimizer that does not require retraining with all historical samples. It can fine-tune the model parameters based on feedback error after each case evaluation.

[0130] The weighted fusion weight parameter refers to the adaptive weights that control the importance of the joint encoding vector, comorbidity load, and KPCA dimensionality reduction features during the diagnostic feature fusion stage.

[0131] The preset threshold boundary refers to the upper and lower limits of the diagnostic severity score used to screen the reference group with the same condition. The threshold directly controls the range of comparison samples for calculating the mismatch coefficient.

[0132] After the system outputs the case classification and early warning results, it automatically captures the actual total number of hospital stays and the DRG group cost consumption index after the case is settled as objective feedback signals. The difference between the comprehensive score of admission necessity output by the model and the actual treatment outcome of the case is calculated to generate a quantitative assessment error: if the model determines that the admission is compliant, but the actual length of hospital stay and cost index are significantly lower, the judgment error is positive; if the model marks it as low-risk, but there is actually severe invasive treatment, the error is negative.

[0133] The system initiates an online gradient descent incremental optimization process, using the evaluation error as the backpropagation of the loss gradient, and simultaneously performs two types of dynamic parameter updates: The first type updates all weights of the attention-weighted fusion of diagnostic features. If the model frequently underestimates the risk of mild cases, the weights of severe case features are automatically reduced, while the weights of comorbidities and outpatient trajectory features are increased. The second type fine-tunes the threshold boundary for selecting a reference group with similar disease conditions. If frequent mixing of severe and mild cases leads to mismatch coefficient distortion, the threshold range is automatically narrowed to reduce the range of disease fluctuations in the comparison group. Parameter updates are performed incrementally online, eliminating the need for periodic full model retraining. The system continuously adapts to changes in regional medical policies, DRG grouping rules, and local treatment habits. Each new discharged case's treatment data continuously corrects the model's judgment scale, preventing performance degradation that can occur after long-term static model operation.

[0134] This embodiment supports multiple sets of configurable parameters to adapt to the medical insurance supervision needs of different regions: the outpatient pre-observation window can be adjusted within the range of 15-60 days, and the time decay factor λ can be set from 0.02 to 0.1; the nonlinear mismatch anomaly threshold supports customization of the 85th to 95th quantile; the admission necessity level three warning threshold scores of 0.3 and 0.6 can be fine-tuned according to the degree of supervision; the CUSUM control upper limit H and the offset constant k can be modified according to the base number of hospitals in the region. In the online gradient descent update process, in addition to adjusting the feature fusion weights and the reference group threshold boundaries, the above globally adjustable parameters are updated simultaneously, dynamically adapting with the local mild case admission baseline, and adapting to the differences in DRG / DIP policies and the differences in treatment behaviors of primary / tertiary hospitals in different provinces and cities.

[0135] This embodiment introduces an online self-updating mechanism to address the pain points of static models failing to adapt to regional differences in diagnosis and treatment and the iteration of medical insurance policies. Relying on actual post-discharge hospital stay duration and cost indices as low-cost feedback labels, it uses gradient descent to fine-tune feature fusion weights and mismatch calculation thresholds in real time, eliminating the need for manual labeling of large numbers of samples for retraining and significantly reducing model maintenance costs. As cases accumulate, the model gradually adapts to local baselines for mild and severe cases, continuously reducing judgment errors and maintaining stable recognition accuracy and recall rates over the long term. This avoids the problem of a surge in missed detections and false alarms over time in fixed-parameter models, making it suitable for diverse deployment scenarios across multiple regions and hospitals nationwide.

[0136] In some embodiments, the method further includes the following steps before S3:

[0137] S101: Obtain M historical admission records of the target case within a preset historical time window, where M is an integer greater than or equal to 1. Extract the corresponding historical diagnostic severity continuous score and historical treatment intensity continuous score from each historical admission record, and construct a longitudinal score sequence for diagnostic severity. longitudinal scoring sequence of diagnosis and treatment intensity ,in, This represents the continuous score of diagnostic severity corresponding to the i-th historical hospital admission. This represents the continuous score of treatment intensity corresponding to the i-th historical hospital admission;

[0138] S102: Concatenate the longitudinal score sequence D of diagnostic severity and the longitudinal score sequence T of treatment intensity to generate a joint longitudinal score vector. The longitudinal score joint vector V is input into a time encoder, which is either a Transformer encoder based on a self-attention mechanism or a recurrent neural network based on a gated recurrent unit, and outputs the hidden vector H of the patient's diagnosis and treatment trajectory.

[0139] S103: Construct a contrastive learning pre-training task, construct positive sample pairs and negative sample pairs from the historical case database, wherein the positive sample pairs are longitudinal score joint vector pairs of two adjacent historical admission records of the same patient, and the negative sample pairs are longitudinal score joint vector pairs of different patients with the same primary diagnosis code. The temporal encoder is pre-trained using the InfoNCE or NT-Xent contrastive loss function to reduce the latent vector distance of the positive sample pairs and increase the latent vector distance of the negative sample pairs.

[0140] S104: Record the continuous score of diagnostic severity and the continuous score of treatment intensity for the target cases in the current admission cycle as follows: and Construct the joint longitudinal score vector for the current admission cycle. ,Will Input the pre-trained temporal encoder to generate the latent vector of the current diagnosis and treatment trajectory. ;

[0141] S105: Calculate the hidden vector of the current treatment trajectory. The Euclidean or cosine distance between the patient's treatment trajectory latent vector H and the longitudinal mismatch distance is used as the longitudinal mismatch distance, which represents the degree of deviation of the current hospitalization cycle's treatment mode from the patient's own historical trajectory.

[0142] In step S4, the longitudinal mismatch distance is input as an additional feature into the ensemble learning model.

[0143] In this embodiment, the preset historical admission window refers to a configurable patient history hospitalization retrospective period, used to extract time-series data of the same patient being admitted multiple times, targeting scenarios of repeated hospitalizations for chronic diseases such as hypertension and lower back pain.

[0144] A longitudinal scoring sequence refers to a historical scoring sequence of diagnosis and treatment in two dimensions, ordered by the patient's historical admission time, taking the patient as the unit, to achieve longitudinal comparison "with the patient as the control".

[0145] A temporal encoder refers to a Transformer self-attention network / GRU recurrent network, which is specifically designed to extract long-term dependencies in temporal sequences and characterize a patient's stable treatment habits over many years.

[0146] Contrastive learning pre-training refers to an unsupervised representation learning method that uses the patient's own historical trajectory to construct positive and negative samples, and can complete encoder training without manual labeling.

[0147] InfoNCE / NT-Xent contrastive loss refers to a standardized contrastive learning loss function that constrains patients with the same trajectory features to be close together, while patients with different diseases with the same trajectory features to be far apart.

[0148] The longitudinal mismatch distance refers to the vector distance between the current hospitalization trajectory and the patient's historical stable trajectory. The larger the value, the more serious the deviation between the current hospitalization treatment pattern and the patient's long-term habits, and the higher the suspicion level.

[0149] In step S101, the system retrieves all historical hospitalization records in the target patient window, reuses the S1 and S2 algorithm steps to batch calculate the diagnostic severity and treatment intensity scores for each historical admission, sorts them according to the admission time, and generates two independent longitudinal sequences to completely preserve the patient's long-term treatment patterns.

[0150] In step S102, the two sequences are concatenated into a one-dimensional joint vector, which is then input into the temporal encoder to capture temporal dependencies. The self-attention mechanism automatically weights the long-term and recent admission records, while the GRU efficiently compresses long sequence information and outputs a fixed-dimensional latent vector H that represents the patient's stable treatment habits.

[0151] In step S103, the offline completion of encoder contrastive learning pre-training specifically includes: positive samples are taken from the joint vector of two adjacent admissions of the same patient, and negative samples are taken from the trajectory vector of the same disease of other patients; by optimizing the network parameters through contrastive loss, the encoder after training can accurately identify the unique and stable diagnosis and treatment pattern of the same patient and distinguish individual differences between different patients.

[0152] In step S104, the current hospitalization... , Appended to the end of the historical sequence, generating a complete current trajectory vector containing both historical and current data, which is then fed into a pre-trained encoder to obtain the current trajectory latent vector. .

[0153] In step S105, calculate The vector distance with the historical baseline latent vector H yields the longitudinal mismatch distance. If a patient has long-term mild symptoms and only receives low-intensity interventions, but this time experiences frequent short-term hospitalizations without indications, the distance will increase significantly. This distance is incorporated as an independent feature into the S4 ensemble model, forming a bidirectional complementary discrimination with the horizontal disease mismatch coefficient.

[0154] This embodiment constructs a longitudinal comparison system for individual patients to overcome the shortcomings of only making horizontal comparisons within the same disease, and is specifically optimized for scenarios of repeated under-standard hospitalizations for chronic diseases. Relying on unsupervised pre-training through contrastive learning, the temporal encoder can be trained using only the patient's natural historical hospitalization data, without the need for a large number of manually labeled samples. The longitudinal mismatch distance quantifies the degree of deviation between the current hospitalization and the patient's long-term treatment habits, and can identify hidden violation patterns such as "patients only had outpatient follow-ups in previous years, and recently had a large number of short-term hospitalizations to collect surplus funds." As a new input feature of the integrated model, it forms a two-layer verification with the horizontal nonlinear mismatch coefficient, further improving the identification and recall rate of under-standard hospitalizations for chronic diseases and reducing misjudgments caused by occasional hospitalizations for minor illnesses.

[0155] In some embodiments, prior to step S4, the method further includes:

[0156] S301: Extract a training sample set from the historical case database. Each training sample includes a continuous score for diagnostic severity, a continuous score for treatment intensity, a nonlinear mismatch coefficient, the patient's age, the patient's gender, the route of admission, the admitting department, the actual length of hospital stay, the actual total hospitalization cost, and the actual admission outcome. The actual admission outcome is a binary label, with a value of 1 indicating a reasonable admission and a value of 0 indicating a low-standard admission.

[0157] S302: The continuous score of diagnostic severity, continuous score of treatment intensity, nonlinear mismatch coefficient, patient age, patient gender, admission route, and admission department in the training sample set are used as a set of candidate causal variables. The actual hospital admission outcome is used as the target variable Y. A causal discovery algorithm based on conditional independence test is used to calculate the causal strength between each variable in the candidate causal variable set X and the target variable Y, and a directed acyclic causal graph structure G is learned.

[0158] S303: Based on the directed acyclic causal graph structure G, identify variables that have a direct causal edge with the target variable Y as a set of causal parent nodes P(Y), and each variable in the set of causal parent nodes P(Y) is a direct causal driving factor of the admission outcome.

[0159] S304: Using the variables in the set of causal parent nodes P(Y) as input and the actual hospital admission outcome Y as output, construct a structural causal model SCM. The structural causal model SCM includes structural equations and noise distribution. Parameters of the structural causal model SCM are estimated by maximum likelihood estimation or variational inference.

[0160] S305: Obtain the current set of observations of the causal parent node of the target case, take the continuous score of the treatment intensity of the target case as the intervention variable, set the counterfactual intervention value as the preset quantile of the continuous score of treatment intensity or the median of the corresponding dimension in the historical baseline dataset, calculate the counterfactual admission necessity probability of the target case under the counterfactual intervention value through the structural causal model SCM, and take the difference between the counterfactual admission necessity probability and the admission necessity probability under the current observation value as the causal effect estimate.

[0161] When performing step S4, the causal effect estimate is used as the input feature of the ensemble learning model, or the causal effect estimate and the nonlinear mismatch coefficient are weighted and fused together and then used as the input feature of the ensemble learning model.

[0162] In this embodiment, the binary admission outcome label refers to the actual result after manual review by the medical insurance auditors, where 1 = compliant hospitalization and 0 = confirmed low-standard admission, used for causal model training.

[0163] The causal discovery algorithm (PC / FCI algorithm) is an undirected graph learning tool based on conditional independence test, which distinguishes between correlation and true causal relationship between variables and removes confounding variables such as age and gender.

[0164] Directed acyclic causal graphs (DAGs) visualize the causal transmission links of variables, with arrows representing the direction of causal action, thus eliminating the model confounding biases caused by simple correlations.

[0165] The causal parent node P(Y) refers to the core variable that directly drives whether hospital admission is low-standard, filtering out interference features that are only correlated but have no causal effect.

[0166] Structural causal model (SCM) is a quantitative mathematical model derived from causal graphs. It includes variable structural equations and random noise distributions, and supports counterfactual intervention reasoning.

[0167] Counterfactual intervention refers to modifying the intensity of medical care to a reasonable baseline level, simulating a hypothetical scenario of "whether hospitalization would still be necessary if standard outpatient treatment were given".

[0168] The causal effect estimate refers to the difference between the current observation and the probability of low-standard risk under the counterfactual scenario, which quantifies the true causal contribution of "insufficient treatment intensity" to the low-standard admission outcome.

[0169] In step S301, historical cases with manual audit labels are collected, and multi-dimensional variables such as diagnosis, treatment, mismatch, demographics, and admission channels are integrated with binary outcome labels to construct a dedicated sample set for causal learning.

[0170] In step S302, the conditional independence test causal discovery algorithm is used to traverse all candidate variables, distinguish the correlation and causal relationship between variables and the low-standard outcome, output the complete DAG causal graph G, calculate the direct causal strength of each variable on the low-standard outcome, and remove confounding variables that only have statistical correlation but no actual clinical causality (such as simple age correlation without direct hospitalization driving effect).

[0171] In step S303, the causal graph is traversed, and the set of causal parent nodes that directly point to the target variable Y is filtered out. Only the core features that truly affect the necessity of hospital admission are retained, thereby reducing the noise interference of irrelevant variables on the model.

[0172] In step S304, an SCM structural causal model is built with the causal parent node as input. The structural equations of each variable and the noise distribution are fitted by maximum likelihood estimation. The true causal weight of each core feature to the low-standard risk is quantified, which is different from the ensemble model that only fits the correlation.

[0173] In step S305, the treatment intensity score is used as an intervention variable and artificially adjusted to the baseline median (outpatient standard treatment level) for the same disease. This score is then substituted into the SCM to calculate the probability of low-standard admission under the counterfactual scenario. The difference between this probability and the current actual observed probability yields the causal effect estimate. This value quantifies the causal contribution of "the current excessively low treatment intensity directly leading to the low-standard admission decision." The causal effect estimate is then fed into the ensemble model as a core additional feature, or weighted and fused with nonlinear mismatch coefficients, allowing the comprehensive score to simultaneously consider correlation fitting and causal mechanisms, avoiding misjudgments caused by confounding variables.

[0174] This embodiment introduces a causal inference framework to address the shortcomings of traditional machine learning, which only learns variable correlations and is susceptible to confounding factors. A causal discovery algorithm is used to screen the core features that truly drive under-standard hospital admissions, eliminating irrelevant confounding variables such as demographics, thus improving the clinical interpretability of the model's judgment logic. Based on a structural causal model, counterfactual intervention inference is completed, quantifying the true causal contribution of inadequate treatment to under-standard behavior, providing quantitative causal evidence for audit assessments. The causal effect estimate, as input to the ensemble model, complements the nonlinear mismatch coefficient mechanism, significantly reducing false alarms caused by spurious correlations of the dependent variable. Simultaneously, the model output possesses a complete causal explanation chain, better meeting the stringent interpretability requirements of medical insurance audit evidence collection.

[0175] In some embodiments, in addition to the single-case admission necessity assessment, a new institutional systemic low-standard admission identification process is added, periodically summarizing all case warning data within the region: The first step involves three-dimensional aggregation by medical institution ID, department code, and ICD diagnostic category, calculating the percentage of red-alert cases for each single disease in each department. If the percentage exceeds 30% and is more than three times higher than the baseline of hospitals of the same level in the region, it is determined that the department is systematically admitting mild cases in batches. The second step involves constructing a monthly warning rate time series and using the CUSUM cumulative sum control chart algorithm to calculate the positive cumulative statistic. ,in, This represents the positive cumulative sum statistic for the t-th statistical month, where t is the time index. The low standard admission red alert rate of the target medical institution in month t is calculated as: number of red alert cases in the current month ÷ total number of discharged cases in the current month. The long-term baseline warning rate (steady-state mean) of medical institutions is the average percentage of monthly red warnings during normal treatment cycles in the history of the medical institution when there are no systemic mild cases admitted, which is taken as the normal risk baseline; k represents a constant that allows for deviation. This represents the net positive risk deviation of the monthly warning rate after deducting the benchmark and the allowable small fluctuations. If the result is >0, it means that the risk of admission to the hospital under the low standard is significantly higher than the normal range in the current month and needs to be accumulated into the CUSUM statistic. If the result is ≤0, it means that the risk fluctuation in the current month is within the allowable range and there is no accumulation. After processing by the function max(0,·), the current period's cumulative sum is maintained at 0.

[0176] When the accumulated amount exceeds the preset control limit H, it is determined that the institution has a sudden change in the trend of low-standard admission; the third step integrates the departments with abnormal clustering, the types of mild diseases with high incidence, and the time points of the time sequence change to generate a multi-dimensional risk profile of low-standard admission of medical institutions. The profile is directly connected to the medical insurance audit task scheduling system to achieve targeted on-site inspection and greatly reduce the scope of manual screening.

[0177] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the hospital admission necessity assessment method based on a diagnosis-treatment mismatch model as described in the first aspect of the present invention.

[0178] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0179] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.

[0180] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.

[0181] like Figure 2 As shown, in a third aspect, the present invention provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the hospital admission necessity assessment method based on the diagnosis-treatment mismatch model as described in the first aspect of the present invention.

[0182] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may use at least one of the following: circuit, single or multiple application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), central processing units (CPUs), controllers, microcontrollers, and microprocessors, thereby enabling the processor to execute some or all of the steps, or any combination thereof, of the hospital admission necessity assessment method based on the diagnosis-treatment mismatch model described in the various embodiments of the present invention.

[0183] Finally, it should be noted that although the above embodiments have been described in the description and drawings of this invention, this should not limit the scope of patent protection of this invention. Any technical solutions that are based on the essential concept of this invention, utilize the content described in the description and drawings of this invention to make equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this invention.

Claims

1. A method for assessing the necessity of hospital admission based on a diagnosis-treatment mismatch model, characterized in that, The method includes: S1: Obtain the diagnostic coding data of the target case. The diagnostic coding data includes a set of primary diagnostic codes and a set of secondary diagnostic codes. The primary diagnostic codes are decomposed into multiple levels and concatenated into a joint diagnostic feature vector. The set of secondary diagnostic codes is matched with a comorbidity index dictionary, and the secondary diagnostic severity load value is calculated based on the matching result. Objective severity indicators are aggregated and statistically analyzed by diagnostic category based on historical hospitalization data. The objective severity indicators are mapped to a high-dimensional space through kernel principal component analysis. Dimensionality reduction features that characterize the complexity of the disease are extracted. The joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused, and then normalized to the [0,1] interval through an activation function to generate a continuous diagnostic severity score. S2: Obtain detailed inpatient cost data for the target case, extract multiple dimensions of treatment intensity indicators, use the primary diagnosis code and secondary diagnosis severity load values ​​as indexes, retrieve historical baseline datasets with the same primary diagnosis code and the same comorbidity stratification level, transform the treatment intensity indicators of each dimension into deviation features relative to the historical baseline with the same primary diagnosis code and the same comorbidity stratification level, and generate normalized continuous treatment intensity scores through nonlinear fusion model mapping. S3: Select a group of historical cases from the historical case database whose diagnostic severity continuous scores are within the same preset threshold range as the target case as a reference group. Within the reference group, obtain the first percentile ranking of the target case's diagnostic severity continuous score and the second percentile ranking of the target case's treatment intensity continuous score. Calculate the nonlinear mismatch coefficient based on the first percentile ranking and the second percentile ranking. S4: Input the multidimensional features, including at least the continuous score of diagnostic severity, the continuous score of treatment intensity, and the nonlinear mismatch coefficient, into the integrated learning model, output a comprehensive score representing the necessity of hospitalization, and output the assessment result of the necessity of hospitalization based on the comprehensive score.

2. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, In step S1, the joint diagnostic feature vector, secondary diagnostic severity load value, and the dimensionality reduction feature are weighted and fused, then normalized to the [0,1] interval using an activation function to generate a continuous diagnostic severity score, specifically including: The joint diagnostic feature vector is input into the first fully connected network and mapped to the first hidden vector. The first hidden vector is concatenated with the secondary diagnostic severity load value and then input into the second fully connected network to generate the fused feature vector. The fused feature vector and the dimensionality-reduced feature are attention-weighted to output a weighted fusion vector. The weighted fusion vector is then mapped to the [0,1] interval using a Sigmoid activation function to generate the continuous score for the diagnostic severity.

3. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, The steps preceding step S1 also include: Obtain the outpatient visit record sequence of the target case within a preset time window before admission, input the outpatient visit record sequence into a hidden Markov model, the state space of the hidden Markov model includes several preset outpatient visit modes, map the outpatient visit record sequence into an outpatient trajectory state sequence, and extract the transition probability matrix and dwell time distribution of the outpatient trajectory state sequence as outpatient trajectory features. The outpatient trajectory features, the joint diagnostic feature vector, the secondary diagnostic severity load value, and the dimensionality reduction features are weighted and fused together to generate the continuous diagnostic severity score.

4. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 3, characterized in that, The steps following step S2 and before step S3 also include: The outpatient visit cost data and outpatient treatment item data of the target case within a preset time window before admission are obtained. The outpatient visit cost data and outpatient treatment item data are input into the counterfactual treatment intensity prediction model. The counterfactual treatment intensity prediction model is trained with historical data of outpatient cases that have not been admitted to the hospital and have the same primary diagnosis code and the same comorbidity stratification level. The model outputs a predicted treatment intensity score that represents the target case if it had not been admitted to the hospital but had received outpatient treatment. The continuous score of the treatment intensity of the target case is compared with the predicted treatment intensity score to calculate the treatment intensity increment, and the treatment intensity increment is input as an additional feature into the ensemble learning model.

5. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, In step S2, the treatment intensity indicators of each dimension are transformed into deviation features relative to the historical baseline of the same primary diagnosis code and the same comorbidity stratification level, specifically including: The treatment intensity indicators of each dimension are transformed by an equal logarithmic ratio and mapped to an unconstrained real number space to obtain the transformed treatment intensity vector. The transformed treatment intensity vector is then input into an autoencoder model, which is trained using cost data from the historical baseline dataset that encodes the same primary diagnosis and is stratified at the same level of comorbidities. The reconstruction error between the treatment intensity vector and the reconstruction output of the autoencoder is calculated as the deviation feature.

6. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, The actual length of hospital stay and actual cost consumption index of the target case after admission are obtained. The actual length of hospital stay and actual cost consumption index are used as feedback signals. The evaluation error is calculated based on the feedback signals. The online gradient descent algorithm is used to update the weight parameters of the weighted fusion in step S1 and the boundary values ​​of the preset threshold range in step S3 according to the evaluation error. This allows the continuous score of the diagnostic severity and the comprehensive score of the necessity of hospitalization for subsequent cases to be dynamically adjusted according to the feedback of actual diagnosis and treatment results.

7. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, Prior to S3, it also included: S101: Obtain M historical admission records of the target case within a preset historical time window, where M is an integer greater than or equal to 1. Extract the corresponding historical diagnostic severity continuous score and historical treatment intensity continuous score from each historical admission record, and construct a longitudinal score sequence for diagnostic severity. longitudinal scoring sequence of diagnosis and treatment intensity ,in, This represents the continuous score of diagnostic severity corresponding to the i-th historical hospital admission. This represents the continuous score of treatment intensity corresponding to the i-th historical hospital admission; S102: Concatenate the longitudinal score sequence D of diagnostic severity and the longitudinal score sequence T of treatment intensity to generate a joint longitudinal score vector. The longitudinal score joint vector V is input into a time encoder, which is either a Transformer encoder based on a self-attention mechanism or a recurrent neural network based on a gated recurrent unit, and outputs the hidden vector H of the patient's diagnosis and treatment trajectory. S103: Construct a contrastive learning pre-training task, construct positive sample pairs and negative sample pairs from the historical case database, wherein the positive sample pairs are longitudinal score joint vector pairs of two adjacent historical admission records of the same patient, and the negative sample pairs are longitudinal score joint vector pairs of different patients with the same primary diagnosis code. The temporal encoder is pre-trained using the InfoNCE or NT-Xent contrastive loss function to reduce the latent vector distance of the positive sample pairs and increase the latent vector distance of the negative sample pairs. S104: Record the continuous score of diagnostic severity and the continuous score of treatment intensity for the target cases in the current admission cycle as follows: and Construct the joint longitudinal score vector for the current admission cycle. ,Will Input the pre-trained temporal encoder to generate the latent vector of the current diagnosis and treatment trajectory. ; S105: Calculate the hidden vector of the current treatment trajectory. The Euclidean or cosine distance between the patient's treatment trajectory latent vector H and the longitudinal mismatch distance is used as the longitudinal mismatch distance, which represents the degree of deviation of the current hospitalization cycle's treatment mode from the patient's own historical trajectory. When performing step S4, the longitudinal mismatch distance is input as an additional feature into the ensemble learning model.

8. The method for assessing the necessity of hospitalization based on a diagnosis-treatment mismatch model as described in claim 1, characterized in that, Before step S4, the following is also included: S301: Extract a training sample set from the historical case database. Each training sample includes a continuous score for diagnostic severity, a continuous score for treatment intensity, a nonlinear mismatch coefficient, the patient's age, the patient's gender, the route of admission, the admitting department, the actual length of hospital stay, the actual total hospitalization cost, and the actual admission outcome. The actual admission outcome is a binary label, with a value of 1 indicating a reasonable admission and a value of 0 indicating a low-standard admission. S302: The continuous score of diagnostic severity, continuous score of treatment intensity, nonlinear mismatch coefficient, patient age, patient gender, admission route, and admission department in the training sample set are used as a set of candidate causal variables. The actual hospital admission outcome is used as the target variable Y. A causal discovery algorithm based on conditional independence test is used to calculate the causal strength between each variable in the candidate causal variable set X and the target variable Y, and a directed acyclic causal graph structure G is learned. S303: Based on the directed acyclic causal graph structure G, identify variables that have a direct causal edge with the target variable Y as a set of causal parent nodes P(Y), and each variable in the set of causal parent nodes P(Y) is a direct causal driving factor of the admission outcome. S304: Using the variables in the set of causal parent nodes P(Y) as input and the actual hospital admission outcome Y as output, construct a structural causal model SCM. The structural causal model SCM includes structural equations and noise distribution. Parameters of the structural causal model SCM are estimated by maximum likelihood estimation or variational inference. S305: Obtain the current set of observations of the causal parent node of the target case, take the continuous score of the treatment intensity of the target case as the intervention variable, set the counterfactual intervention value as the preset quantile of the continuous score of treatment intensity or the median of the corresponding dimension in the historical baseline dataset, calculate the counterfactual admission necessity probability of the target case under the counterfactual intervention value through the structural causal model SCM, and take the difference between the counterfactual admission necessity probability and the admission necessity probability under the current observation value as the causal effect estimate. When performing step S4, the causal effect estimate is used as the input feature of the ensemble learning model, or the causal effect estimate and the nonlinear mismatch coefficient are weighted and fused together and then used as the input feature of the ensemble learning model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the hospital admission necessity assessment method based on the diagnosis-treatment mismatch model as described in any one of claims 1 to 8.

10. An electronic device having a computer program stored thereon, characterized in that, The device includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and the computer program, when executed by the processor, implements the hospital admission necessity assessment method based on the diagnosis-treatment mismatch model as described in any one of claims 1 to 8.

Images