Medical expenditure joint prediction method and system based on structural consistency constraint, storage medium and electronic device

Abstract

This invention belongs to the field of medical data processing technology and discloses a joint prediction method, system, storage medium, and electronic device for medical expenditures based on structural consistency constraints. Addressing the inherent defect of inconsistent prediction head output logic in existing multi-task learning methods, this invention forcibly eliminates semantic conflicts between different prediction heads during the training phase by simultaneously applying non-cross-quantile constraints, nested risk probability monotonic constraints, and cross-head consistency regularization. The measured quantile cross-rate and risk monotonic violation rate are both 0%, achieving deployment-level structural consistency guarantees and solving the technical problem of insufficient reliability in automated decision-making systems caused by contradictory output logic in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and specifically to a method, system, storage medium, and electronic device for joint prediction of medical expenditures based on structural consistency constraints. Background Technology

[0002] In the fields of managed care and actuarial science, accurately predicting an individual's future healthcare expenditure risk is a fundamental technical step in making healthcare resource allocation decisions. Healthcare expenditure data has physical characteristics that significantly distinguish it from conventional data: its distribution exhibits extreme heavy-tailing, meaning that a small group (such as the top 5%) bears a disproportionate share of total expenditure (such as approximately 50%) compared to their population share; at the same time, this type of data usually comes from complex sampling survey designs (such as MEPS-HC), with attached survey weights reflecting the overall structure. Ignoring these weights will lead to a bias in the estimation of overall representativeness.

[0003] The existing technologies for predicting medical expenditure risk mainly adopt the following methods, but all of them have objective technical defects: (1) Limitations of single-task prediction methods: Mean regression methods (such as generalized linear model GLM, Tweedie regression, XGBoost) aim to minimize the mean square error. Under heavy-tailed distribution, they have a systematic underestimation bias on high-cost tail data and cannot provide uncertainty quantification. Single-threshold classification methods (such as logistic regression) can only output risk probability for a single threshold and cannot provide decision ranking for multiple capacity points at the same time. They also lack characterization of the tail shape of the risk distribution. Although independent quantile regression methods can estimate conditional quantiles, independent training of each quantile often leads to quantile crossover, that is, the predicted value of the high quantile is lower than the predicted value of the low quantile, which violates the basic monotonicity of the probability distribution and makes the prediction results logically uninterpretable and cannot be directly used in clinical decision support systems. (2) Inherent defects of multi-task learning architecture: Multi-task deep learning methods developed in recent years adopt a shared representation layer and multi-predictor head architecture, attempting to output mean, quantile and classification probability at the same time. However, such methods only treat each prediction task as an independent optimization objective and process them in parallel without establishing a structural consistency constraint mechanism between prediction heads. This leads to inherent logical conflicts in the outputs of different prediction heads at the deployment level: for example, the overthreshold probability predicted by the risk classification head is inconsistent with the probability distribution implied by the quantile prediction head, or there is non-monotonic nesting between different threshold risk probabilities. Such logical conflicts make it impossible for the model output to form a unified and interpretable decision basis, which may lead to the decision system performing incorrect allocation based on contradictory data during automated deployment. (3) Lack of adaptation to complex sampling design: Existing deep learning methods are usually optimized based on the assumption of simple random sampling, without considering the complex sampling weights inherent in medical survey data. This leads to the model training process overfitting to subgroups with large sample sizes, failing to obtain statistically representative risk estimates at the overall level, and thus affecting the accuracy of population-level resource allocation decisions based on prediction results. (4) Technical gap in dynamic identification of cost status: Existing technologies mostly focus on static cost level prediction and lack technical means to identify the dynamic transfer of individual cost status (such as new entry into high risk, continuous high risk, exit from high risk), resulting in the inability to differentiate intervention strategies according to the dynamic trajectory of patients.

[0004] In view of this, the present invention is proposed. Summary of the Invention

[0005] The present invention aims to solve at least one of the above technical problems, and provides a method, system, storage medium and electronic device for joint prediction of medical expenditure based on structural consistency constraints.

[0006] This invention achieves the following technical effects by imposing specific structural consistency constraints on the heavy-tailed distribution characteristics and complex sampling designs of medical data, and by optimizing the multi-task learning architecture: To address the inherent flaw of inconsistent prediction head output logic in existing multi-task learning methods, this invention forcibly eliminates semantic conflicts between different prediction heads during the training phase by simultaneously applying non-cross quantile constraints, nested risk probability monotonic constraints, and cross-head consistency regularization. The measured quantile crossover rate and risk monotonic violation rate are both 0%, achieving deployment-level structural consistency guarantees and resolving the technical problem of insufficient reliability in automated decision-making systems caused by contradictory output logic in existing technologies.

[0007] By introducing complex sampling survey weights for weighted calculation during model training and evaluation, this invention overcomes the estimation bias problem caused by the simple random sampling assumption in existing technologies, making the model optimization objective consistent with the overall population distribution, ensuring that the risk identification results have statistical validity at the overall level, and improving the technical accuracy of medical risk prediction at the population level.

[0008] To address the heavy-tailed nature of healthcare expenditures, this invention significantly improves upon the systematic underestimation of high-cost populations by employing tail-weighted methods and a high-cost underestimation penalty mechanism. Under ultra-tight capacity constraints, the high-risk population capture rate is increased by 1.311 times compared to the traditional logistic regression baseline, and the accuracy is improved by 6.40 percentage points, significantly enhancing the allocation efficiency of scarce medical intervention resources.

[0009] This invention achieves technical identification of the dynamic trajectory of high-cost populations by generating state transition prediction results (probabilities of new entry, continued entry, and exit from high-cost populations), providing data support for formulating differentiated continuous management or impact intervention strategies, and filling the functional gaps of existing static prediction technologies. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the joint prediction method for medical expenditures based on structural consistency constraints provided for embodiments of the present invention; Figure 2 A schematic diagram of the structure of the medical expenditure joint prediction system based on structural consistency constraints provided for an embodiment of the present invention; Figure 3 A schematic block diagram of an example electronic device provided for an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be arbitrarily combined with each other. Those skilled in the art should understand that the following descriptions are merely exemplary means of implementing this invention and not limiting conditions; any technical means employed to achieve the same or similar technical effects as this invention should fall within the protection scope of this invention.

[0012] Figure 1 This is a flowchart illustrating a joint forecasting method for healthcare expenditures based on structural consistency constraints. (Refer to...) Figure 1 The first embodiment of the present invention provides a joint prediction method for medical expenditures based on structural consistency constraints, comprising the following steps: S101, acquire individual medical data, the medical data including demographic characteristics, medical utilization characteristics and insurance information characteristics reflecting individual medical risks, the medical data having heavy-tailed distribution characteristics.

[0013] This step aims to obtain statistically representative medical expenditure data to provide input for subsequent multi-task prediction models. The medical data includes demographic characteristics, medical utilization characteristics, and insurance information characteristics reflecting individual medical risks; these three types of characteristics constitute the basic and necessary feature set of the technical solution of this invention.

[0014] Demographic characteristics are used to characterize an individual's basic sociodemographic attributes and are fundamental determinants of healthcare expenditure. These characteristics typically include age, sex, race, region, and level of poverty. Age and sex are closely related to healthcare needs (e.g., the elderly bear a heavy burden of chronic diseases, and women in certain age groups have high healthcare utilization rates); race and region reflect the accessibility of healthcare resources and health disparities; and the level of poverty directly affects the ability to pay for healthcare and healthcare-seeking behavior.

[0015] Healthcare utilization characteristics reflect an individual's historical healthcare service usage patterns and health status, serving as a core basis for predicting future expenditures. These characteristics typically include total historical healthcare expenditures, out-of-pocket expenses, number of chronic diseases, number of outpatient visits, and top-level cost indicators (such as top 10%, top 5%, and top 1% cost indicators). Historical expenditures are the strongest predictor of future expenditures; the number of chronic diseases reflects the cumulative disease burden; and the number of outpatient visits reflects the intensity of healthcare utilization.

[0016] Insurance information features characterize an individual's healthcare coverage and affordability, directly impacting accessibility to healthcare services and actual expenditure levels. These features typically include insurance coverage type (e.g., commercial insurance, public health insurance, no insurance), coverage scope, and co-payment ratio. Insurance type determines the economic accessibility of healthcare services, thus influencing healthcare-seeking behavior and expenditure levels.

[0017] The above three types of characteristics constitute the minimum necessary feature set for medical risk prediction, and none can be omitted. Demographic characteristics alone cannot reflect an individual's health status; medical utilization characteristics alone cannot distinguish between affordability and social determinants; and insurance information characteristics alone cannot assess actual medical needs.

[0018] Beyond the minimum necessary feature set, the medical data may also include other feature categories reflecting an individual's medical risks. In some preferred embodiments, the medical data further includes prescription drug use characteristics, which reflect an individual's medication patterns and drug dependence levels, typically including the number of prescription drug events, total prescription drug costs, number of unique drug codes, and drug diversity ratio. In other preferred embodiments, the medical data may further include genetic characteristics, lifestyle characteristics, socioeconomic status characteristics, etc., which can supplement the reflection of an individual's medical risks or health determinants.

[0019] In one specific embodiment of the present invention, publicly available panel survey data on medical expenditures (exemplarily Panel 27, 2022-2023) is used to construct a feature set that includes the three essential characteristics mentioned above and prescription drug utilization characteristics, exemplarily totaling thirty-two dimensions of features (including 9 dimensions of demographic characteristics, 5 dimensions of historical medical utilization signals, 12 dimensions of engineering characteristics, 3 dimensions of prescription drug aggregation characteristics, and 3 dimensions of cross-derived characteristics). This data exhibits a heavy-tailed distribution, meaning that the top 5% of the population bears approximately 50% of the total expenditure, and the top 1% bears approximately 25% of the total expenditure; it also incorporates complex sampling survey weights (such as PERWT22F) to correct for sample bias caused by complex sampling design.

[0020] In other specific embodiments, the feature dimensions can be adjusted within the range of 20 to 50 dimensions. For example, the technical effects of the present invention can be achieved by using only a 26-dimensional basic feature set (excluding prescription drug aggregation features) or by using a 45-dimensional enhanced feature set (adding chronic disease status markers, medical utilization counts, and health function assessment indicators to the basic features), and the enhanced feature set can further improve the model's performance in terms of capture rate and discrimination.

[0021] Those skilled in the art should understand that the 32-dimensional features described are merely exemplary configurations for this specific data source and are not necessary limitations for implementing the present invention. Depending on the specific application scenario (such as Chinese medical insurance data, European health survey data, or commercial insurance claims data), the feature dimensions can be adjusted accordingly (e.g., from 20 to 50 dimensions), as long as the above three essential features are included, the technical effects of the present invention can be achieved.

[0022] Regarding data acquisition methods, those skilled in the art can employ any conventional medical data collection methods, including but not limited to extraction from electronic health records (EHRs), queries of insurance claims databases, and downloads of government survey data. Importantly, the acquired data should include weight information reflecting its sampling structure to ensure the overall representativeness of the model training.

[0023] S102, the demographic features, medical utilization features and insurance information features are respectively subjected to feature encoding processing, including standardization of numerical features and embedding encoding of categorical features, to obtain a shared representation vector.

[0024] This step aims to transform the raw, heterogeneous medical data into a numerical representation suitable for deep learning models, resulting in a shared representation vector. By specifically encoding the three essential features mentioned above, effective representation of different types of data is ensured.

[0025] For numerical features (such as age, historical medical expenditures, number of chronic diseases, etc.), conventional normalization or standardization methods in the field can be used, such as min-max normalization, Z-score standardization, etc., to eliminate dimensional differences. In a specific embodiment of the present invention, cost-related features in medical utilization features (such as total expenditures in the current year, out-of-pocket expenditures, prescription drug expenditures) are first logarithmically transformed to compress the extreme values ​​of the heavy-tailed distribution, and then Z-score standardization is performed to make their mean 0 and standard deviation 1.

[0026] For categorical features (such as gender, race, region, insurance type, etc.), conventional embedding encoding or one-hot encoding can be used. In a specific embodiment of the present invention, token-based embedding encoding is used for the categorical variables in demographic features and insurance information features, mapping each category to a low-dimensional dense vector.

[0027] Regarding the neural network architecture used for feature encoding, those skilled in the art can employ any conventional deep learning backbone network, including but not limited to multilayer perceptrons, convolutional neural networks, recurrent neural networks, or Transformer architectures. In a preferred embodiment of the present invention, a transformer architecture, specifically FT-Transformer (FeatureTransformer), is used, with an exemplary configuration including: a model dimension of 192, 3 Transformer layers, 8 multi-head attention heads, and a dropout ratio of 0.1. Each Transformer block includes a multi-head self-attention mechanism, layer normalization, a feedforward network, and residual connections.

[0028] After feature encoding, each individual's medical data is mapped into a fixed-dimensional shared representation vector, which will serve multiple subsequent prediction tasks simultaneously, thus achieving representation sharing.

[0029] S103, based on the shared representation vector, simultaneously generate mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results.

[0030] Based on the shared representation vector, this step utilizes a multi-head output architecture to simultaneously generate mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results. Here, "simultaneously" means that during the model's forward propagation, all types of predictions are computed in parallel based on the same shared representation, rather than being trained sequentially and independently.

[0031] The mean prediction results are intended to provide a point estimate of healthcare expenditure for the next period. Conventional regression head designs in the art can be used, such as a single-layer linear layer followed by an activation function. In one specific embodiment of the invention, the mean head employs a single-layer linear mapping followed by an exponential transformation, and preferably, regression training is performed in logarithmic space. During prediction, the original scale is restored through an inverse transformation to accommodate the right-skewed characteristics of the heavy-tailed distribution.

[0032] Conditional quantile predictions aim to characterize the uncertainty of conditional expenditure distributions, providing quantile estimates at different probability levels. Conventional quantile regression head designs in the art can be employed. In one specific embodiment of the invention, five key quantiles are output simultaneously: the median (50th quantile), 75th quantile, 90th quantile, 95th quantile, and 99th quantile.

[0033] The nested high-cost threshold probability prediction aims to output the probability that an individual's future medical costs will exceed a specific threshold. Conventional classification head designs in the art can be used, such as a single-layer linear layer followed by a sigmoid activation function. In a specific embodiment of this invention, three nested threshold probabilities are output: the top 10%, the top 5%, and the top 1% high-cost threshold probabilities, and a monotonic nesting relationship is required (the probability of the top 1% is no greater than the probability of the top 5%, and the probability of the top 5% is no greater than the probability of the top 10%).

[0034] In some preferred embodiments, state transition prediction results are also generated based on the shared representation vector, including three-class probabilities of new high cost, continued high cost, and exiting high cost, which are used to distinguish the dynamic transition trajectory of individual cost states and provide data support for formulating differentiated continuous management or shock intervention strategies.

[0035] S104, During the model training phase, structural consistency constraints are applied to the heavy-tailed distribution characteristics of the medical data. These constraints include: non-cross-quantile constraints, which ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile through non-negative incremental cumulative parameterization; and nested risk probability monotonic constraints, which ensure that the probability of the higher threshold is not greater than the probability of the lower threshold through incremental parameterization and activation functions.

[0036] This step aims to address the technical problem of multi-head output logic conflicts in existing multi-task learning methods. During the model training phase, taking into account the heavy-tailed distribution characteristics of the medical data, the following three structural consistency constraints are applied simultaneously: Non-crossing quantile constraints: These ensure monotonicity between different quantile levels, meaning the predicted value at lower quantiles is no greater than the predicted value at higher quantiles (e.g., 50th quantile ≤ 75th quantile ≤ 90th quantile ≤ 95th quantile ≤ 99th quantile). Those skilled in the art can employ any parameterization method or penalty function capable of forcing monotonicity. In a preferred embodiment of the invention, non-negative increment accumulation parameterization is used: the predicted value of the j-th quantile is represented as the sum of the previous j non-negative increments, i.e. ,in, Let τ be the quantile prediction value, X be the medical data feature vector of the input individual, and τ be the quantile prediction value. j Let be the j-th quantile level, l be the summation index variable, representing the summation from the 1st quantile level to the j-th quantile level, exp(•) be the exponential function, and g be the quantile level. l (X) is the mapping output of the l-th incremental network to the shared representation. By forcing the increment to be non-negative through exponential transformation, the possibility of quantile crossover is fundamentally eliminated.

[0037] Nested risk probability monotonic constraint: This ensures a monotonic nested relationship between different threshold risk probabilities, i.e., a higher threshold probability is no greater than a lower threshold probability (e.g., the top 1% probability ≤ the top 5% probability ≤ the top 10% probability). Those skilled in the art can employ any technique capable of forcing probability monotonicity, such as ordered Logit models, isotonic regression constraints, or monotonic neural networks. In a preferred embodiment of the invention, incremental parameterization and Softplus transformation (Ordered Logits structure) are used: the threshold probability is represented as the sigmoid output of the cumulative increment after a Softplus transformation, and the non-negativity of the Softplus function forces different threshold probabilities to monotonically increase.

[0038] By applying the two constraints in synergy, this invention achieves deployment-level structural consistency, with both the measured quantile crossover rate and the risk monotonic violation rate reaching zero. In some preferred embodiments, the structural consistency constraint further includes cross-head consistency regularization to further constrain the deviation between the risk probability prediction and the implicit probability of the quantile prediction. Specifically, the quantile prediction result implies a probability that the cost exceeds a certain threshold (e.g., the 95th quantile implies the probability of the top 5% risk), and this implicit probability should be close to the risk probability prediction result. Those skilled in the art can use any metric that measures the difference in probability distributions as a regularization term, such as the L1 norm, L2 norm, KL divergence, or JS divergence. In a specific embodiment of this invention, an L2 regularization term is used: , where L coherence The loss is the cross-head consistency regularization loss, where λ is the regularization coefficient (exemplarily taken as 0.01). Let be the output probability of the risk head, representing the high-cost threshold probability directly output by the risk prediction head for the i-th individual. Let ||•|| be the implied probability of the quantile head, representing the implied overthreshold probability derived from the quantile prediction results. 2 It is the square of the L2 norm.

[0039] S105, optimize the model parameters using a joint loss function that includes the structural consistency constraints.

[0040] This step aims to adjust the model parameters through algorithm optimization so that the model output both fits the observed data and satisfies structural consistency constraints. End-to-end training is performed using a joint loss function, which includes a task-specific loss term and a structural consistency constraint term.

[0041] The joint loss function includes, for example, mean regression loss (such as weighted mean square error), quantile pinball loss, risk head binary cross-entropy loss, and regularization loss corresponding to the above structural consistency constraints (non-cross constraint, monotonic constraint, cross-head consistency).

[0042] Mean regression loss is used to optimize mean prediction results, and exemplarily employs mean squared error or weighted mean squared error. In a specific embodiment of the present invention, the mean head loss is calculated in log(1+Y) space, that is, regressing the logarithmic transformation value of medical expenditures. During evaluation, the original scale is restored through inverse transformation to adapt to the right-skewed characteristics of the heavy-tailed distribution.

[0043] Quantile Pinball loss, used to optimize conditional quantile predictions, is the standard loss function for quantile regression. For each quantile level τ, Pinball loss applies different penalties based on the direction of the prediction error: a proportional penalty of τ is applied when the actual value is higher than the predicted quantile, and a proportional penalty of (1-τ) is applied when it is lower, thereby driving the model to estimate the conditional quantile.

[0044] The risk head binary cross-entropy loss is used to optimize the probability prediction results of nested high-cost thresholds. The binary cross-entropy is calculated for each threshold (top 10%, top 5%, top 1%) to measure the difference between the predicted probability and the actual binary classification label (whether it exceeds the threshold).

[0045] In some preferred embodiments, the medical data also includes complex sampling survey weights, and the joint loss function is calculated based on the complex sampling survey weights, that is, the loss of each sample is weighted according to its survey weight, to ensure that the optimization objective reflects the overall distribution rather than the simple sample distribution, to overcome the estimation bias problem caused by the simple random sampling assumption in the prior art, and to ensure that the risk identification results have statistical validity at the overall level.

[0046] In some preferred embodiments, taking into account the heavy-tailed distribution characteristics of the medical data, the present invention may optionally introduce a tail-weighted term and / or a high-cost underestimation penalty term into the joint loss.

[0047] The tail-weighting term applies higher weights to the loss of high-cost samples to account for the heavy-tailed distribution characteristics. In a specific embodiment of the invention, the tail weighting is determined based on the 90th digit of medical expenditures in the training set, and the weighting function is exemplarily w(Y)=1+γσ((YT) tail ) / s), where w(Y) is the tail-weighted function; Y is the actual medical expenditure; T tail γ is the 90th quantile of medical expenditure in the training set; γ is the tail weighting strength coefficient; σ is the Sigmoid function; s is the scaling factor; this loss function achieves a smooth transition of the loss weights for tail samples, and the weighting factor is usually in the range of two to five times.

[0048] The high-cost underestimation penalty is used to mitigate the systematic underestimation of high-cost groups by the mean forecast, and additional penalties are imposed only on samples where actual expenditures exceed the 90th quantile and the predicted value is lower than the actual value: Among them, L under The penalty loss is for underestimating high costs; E[•] is the expectation operator, which represents the weighted average of the training sample set; 1{•} is the indicator function, which takes the value of 1 when the condition in the curly braces is met and takes the value of 0 when it is not met; To predict healthcare expenditures; η is the penalty intensity coefficient. This loss function applies additional penalties only to samples where actual expenditures exceed the 90th percentile and the predicted value is lower than the actual value, in order to mitigate the systematic underestimation of high-cost populations by mean predictions.

[0049] Structural consistency constraint losses include: non-crossing regularization: intrinsically implemented through non-negative incremental cumulative parameterization, requiring no additional penalty term; nested monotonic regularization: intrinsically implemented through incremental parameterization and Softplus transformation, requiring no additional penalty term; and cross-head consistency regularization: i.e., the aforementioned L... coherence Constrain the consistency between the risk head and quantile head outputs.

[0050] Importantly, the joint loss function is calculated based on the weights of the complex sampling survey, that is, the loss of each sample is weighted according to its survey weight, ensuring that the optimization objective reflects the overall distribution rather than the simple sample distribution, and overcoming the estimation bias problem caused by the assumption of simple random sampling in the prior art.

[0051] Through the collaborative optimization of the above multiple loss terms, the model satisfies the structural consistency constraint while fitting the observed data, and performs special processing on the tail characteristics of the heavy-tailed distribution.

[0052] Regarding optimization algorithms, those skilled in the art can use any conventional gradient descent optimizer, such as stochastic gradient descent (SGD), Adam, AdamW, RMSprop, etc., in conjunction with appropriate learning rate scheduling strategies (such as cosine annealing, step decay, exponential decay).

[0053] In one specific embodiment of the present invention, the AdamW optimizer is employed, with an exemplary learning rate of 1×10⁻⁶. -3 The weight decay coefficient is 1×10 for example. -5 The learning rate is scheduled using cosine annealing. Regarding the training epoch configuration, the maximum number of training epochs is no more than 100, preferably 50, and an early stopping mechanism is used to prevent overfitting. In a preferred embodiment of the invention, verified through Optuna hyperparameter search, the early stopping patience value is exemplarily 11 epochs, meaning training terminates when the validation set performance does not improve for 11 consecutive epochs, thus balancing training sufficiency and generalization performance.

[0054] Regarding the training strategy, a phased training strategy or a single-stage joint training strategy can be adopted. The phased training strategy, for example, includes: first optimizing the parameters corresponding to the mean prediction result, and then gradually introducing the parameters corresponding to the quantile prediction result and the nested high-cost threshold probability prediction result for joint optimization. The single-stage joint training strategy simultaneously activates all loss terms for end-to-end optimization. Those skilled in the art can select an appropriate training strategy based on the data scale and convergence stability. The single-stage joint training strategy simultaneously activates all loss terms (including mean regression loss, quantile pinball loss, risk head binary cross-entropy loss, and structural consistency constraint loss) for end-to-end optimization. In a preferred embodiment of the present invention, a batch size of 512 and a learning rate multiplier of 1× are used, and all performance data (including: a 1.311-fold improvement in high-risk population capture rate compared to the traditional logistic regression baseline under 0.5% capacity constraint, a 6.40 percentage point improvement in accuracy under 1% capacity, a reduction in expected calibration error to 0.010, and zero structural consistency violation rate, etc.) are all derived from this single-stage joint training configuration.

[0055] S106 generates mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results based on optimized model parameters, which are used for population risk identification under capacity constraints.

[0056] After optimizing the model parameters through the above training process, this step uses the optimized model to predict the target population and generates the final mean prediction result, conditional quantile prediction result, and nested high-cost threshold probability prediction result.

[0057] For nested high-cost threshold probability prediction results, probability calibration processing may be optionally included to further improve the reliability of the probability output. In a specific embodiment of the present invention, a temperature scaling method is used: the temperature parameter t is learned on the validation set, and the logarithmic probability of the risk head output is scaled to reduce the expected calibration error to below 0.015.

[0058] The generated predictions are ultimately used for population risk identification under capacity constraints. Specifically, based on the budget capacity of the managed care program (e.g., covering only 0.5%, 1%, or 5% of the population), the population is ranked according to risk probability or quantile predictions, and the top k% of individuals with the highest risk are selected as high-risk candidates. In practice, under a very tight capacity constraint (0.5%), the high-risk population capture rate is 1.311 times higher than the traditional logistic regression baseline, and the accuracy is improved by 6.40 percentage points.

[0059] Those skilled in the art should understand that the above-mentioned joint forecasting method for medical expenditures based on structural consistency constraints can also be applied to various medical management scenarios.

[0060] As an example, this can be applied to budget range and capacity decision-making scenarios. Individual-level budget ranges are constructed using the mean prediction results and conditional quantile prediction results. For instance, the mean prediction of the previous year's medical expenditure is used as the point estimation center, and the prediction range is constructed by combining the 50th quantile (median) and the 95th quantile, or a higher-confidence budget upper limit range is constructed using the 75th quantile and the 99th quantile. By aggregating the population's prediction ranges with the nested high-cost threshold probability prediction results, a capture curve is plotted, i.e., a decision support curve where the horizontal axis represents the coverage capacity ratio and the vertical axis represents the captured high-cost expenditure ratio. Decision-makers can identify the optimal capacity point on the capture curve based on a given budget constraint (e.g., only 1% of the population can be covered), balancing coverage rate and intervention costs to achieve the optimal capacity-value match.

[0061] As an example, this can be used in scenarios involving uncertainty triage and review triggering. The interval width of the conditional quantile prediction results is used as a quantitative indicator of individual uncertainty. For example, the difference between the 99th and 95th quantiles is calculated as a tail uncertainty measure, or the difference between the 95th and 50th quantiles is calculated as an overall uncertainty measure. An uncertainty threshold is set (for example, twice the median interval width of the entire population). Individuals whose quantile differences exceed the threshold are marked as high-uncertainty individuals. This triggers priority for manual clinical review, initiating additional information collection processes (such as supplementary laboratory tests, imaging data, or past medical history), or postponing automated intervention decisions until uncertainty decreases. This reduces erroneous automated interventions for high-uncertainty individuals and improves the reliability of intervention decisions.

[0062] As an example, this can also be used to differentiate between persistent and shock-induced high-cost management scenarios. The three-category probabilities output by the state transition prediction results—the probability of new high-cost entry, the probability of persistent high-cost entry, and the probability of exiting high-cost entry—are used to distinguish the individual's cost dynamic patterns. Individuals with a persistent high-cost probability higher than a threshold (exemplarily 0.5) are classified as "persistent high-cost individuals," characterized by long-term medical burden due to chronic disease, making them suitable for long-term continuous care management resources (such as chronic disease management programs, regular follow-ups, and medication adherence management). Individuals with a new high-cost probability higher than the threshold are classified as "one-off cost shock individuals," characterized by short-term high expenditures due to acute events, making them suitable for short-term shock response strategies (such as acute-phase interventions and short-term rehabilitation support), avoiding misallocation of long-term management resources. Through this differentiation mechanism, precise matching of intervention resources and individual cost dynamics is achieved, improving management efficiency.

[0063] In summary, the embodiments of the present invention, through the coordinated implementation of the above steps, solve the technical problem of inconsistent output logic of multi-task prediction models for medical data with heavy-tailed distribution characteristics and originating from complex sampling designs, achieve deployment-level structural consistency assurance, and significantly improve the technical reliability of medical risk prediction in automated decision-making systems.

[0064] Those skilled in the art should understand that, without departing from the concept of the present invention, the specific implementation methods of the above steps (such as network architecture selection, parameterization methods, optimization algorithms, etc.) can be conventionally adjusted or replaced, and these equivalent transformations should all be covered within the protection scope of the present invention.

[0065] Figure 2 This is a schematic diagram of the structure of the joint medical expenditure prediction system 200 based on structural consistency constraints, referenced. Figure 2 The second embodiment of the present invention provides a joint prediction system 200 for medical expenditures based on structural consistency constraints, including: an acquisition module 201, an encoding module 202, a generation module 203, a constraint module 204, an optimization module 205, and an identification module 206. The functions of each module are described below: The acquisition module 201 is used to acquire the medical data of an individual, wherein the medical data has a heavy-tailed distribution characteristic; Encoding module 202 is used to perform feature encoding processing on the medical data, including standardizing numerical features and embedding and encoding categorical features to obtain a shared representation vector; The generation module 203 is used to simultaneously generate mean prediction results, conditional quantile prediction results, and nested risk probability prediction results based on the shared representation vector; Constraint module 204 is used to apply structural consistency constraints to the heavy-tailed distribution characteristics of the medical data during the model training phase. The constraints include: non-cross quantile constraints, which ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile through non-negative incremental cumulative parameterization; and nested risk probability monotonic constraints, which ensure that the high threshold probability is not greater than the low threshold probability through incremental parameterization and activation function. Optimization module 205 is used to optimize model parameters using a joint loss function that includes the structural consistency constraints; The identification module 206 is used to generate mean prediction results, conditional quantile prediction results and nested high-cost threshold probability prediction results based on the optimized model parameters, and to identify population risks under capacity constraints.

[0066] Based on the above embodiments, the present invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the joint prediction method for medical expenditure based on structural consistency constraints in the first embodiment.

[0067] Figure 3 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0068] like Figure 3 As shown, the electronic device 300 may include a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0069] Multiple components in device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0070] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as a method for joint prediction of medical expenditures based on structural consistency constraints or a model training method for joint prediction of medical expenditures based on structural consistency constraints. For example, in some embodiments, the method for joint prediction of medical expenditures based on structural consistency constraints or the model training method for joint prediction of medical expenditures based on structural consistency constraints can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by computing unit 301, one or more steps of the medical expenditure joint prediction method based on structural consistency constraints or the model training method for medical expenditure joint prediction based on structural consistency constraints described above can be performed. Alternatively, in other embodiments, computing unit 301 can be configured by any other suitable means (e.g., by means of firmware) to perform the medical expenditure joint prediction method based on structural consistency constraints or the model training method for medical expenditure joint prediction based on structural consistency constraints.

[0071] Based on the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions, the computer instructions being used to cause the computer to execute the medical expenditure joint prediction method based on structural consistency constraints disclosed in the embodiments of the present invention.

[0072] Based on the above embodiments, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the medical expenditure joint prediction method based on structural consistency constraints disclosed in the embodiments of the present invention.

[0073] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0074] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0079] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A joint forecasting method for healthcare expenditure based on structural consistency constraints, characterized in that, include: Acquire individual medical data, which includes demographic characteristics, medical utilization characteristics, and insurance information characteristics reflecting individual medical risks, and the medical data has a heavy-tailed distribution characteristic; The demographic features, medical utilization features, and insurance information features are respectively subjected to feature encoding processing, including standardization of numerical features and embedding encoding of categorical features, to obtain a shared representation vector; Based on the shared representation vector, mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results are generated simultaneously. During the model training phase, structural consistency constraints are applied to the heavy-tailed distribution characteristics of the medical data. These constraints include: non-cross quantile constraints, which ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile through non-negative incremental cumulative parameterization; and nested risk probability monotonic constraints, which ensure that the high threshold probability is not greater than the low threshold probability through incremental parameterization and activation functions. The model parameters are optimized using a joint loss function that includes the structural consistency constraints. Based on the optimized model parameters, mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results are generated for population risk identification under capacity constraints.

2. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, The structural consistency constraint also includes crosshead consistency regularization, which constrains the deviation between the risk probability prediction and the implicit probability of the quantile prediction.

3. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, The medical data also includes complex sampling survey weights. The joint loss function is calculated based on the complex sampling survey weights, that is, the loss of each sample is weighted according to its survey weight, to ensure that the optimization objective reflects the overall distribution.

4. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, The conditional quantile prediction results include the median, 75th quantile, 90th quantile, 95th quantile, and 99th quantile; the nested high-cost threshold probability prediction results include the top 10%, top 5%, and top 1% high-cost threshold probabilities, and satisfy a monotonic nesting relationship; the non-cross quantile constraint is implemented by representing the quantile prediction values ​​as the accumulation of non-negative increments, and the non-negative increments are forced to be monotonic by applying an exponential transformation to the learnable function; the nested risk probability monotonic constraint is implemented by representing the threshold probability as the output of the cumulative increment after a Softplus transformation.

5. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, The joint loss function also includes a tail-weighted term and / or a high-cost underestimation penalty term. The tail-weighted term applies higher weights to samples exceeding the 90th decimal place based on the heavy-tailed distribution characteristics. The underestimation penalty term only penalizes samples where the actual expenditure exceeds the 90th decimal place and the predicted value is lower than the actual value.

6. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, The application of the structural consistency constraint ensures that both the quantile crossover rate and the risk monotonic violation rate are zero, achieving deployment-level structural consistency.

7. The joint prediction method for medical expenditure based on structural consistency constraints as described in claim 1, characterized in that, Based on the shared representation vector, state transition prediction results are also generated. These results include three-class probabilities: new entry into high cost, continued high cost, and exit from high cost, which are used to distinguish the dynamic transition trajectory of an individual's cost state.

8. A joint prediction system for medical expenditures based on structural consistency constraints, characterized in that, include: The acquisition module is used to acquire an individual's medical data, which has a heavy-tailed distribution characteristic; The encoding module is used to perform feature encoding processing on the medical data, including standardizing numerical features and embedding categorical features to obtain a shared representation vector. The generation module is used to simultaneously generate mean prediction results, conditional quantile prediction results, and nested risk probability prediction results based on the shared representation vector; The constraint module is used to apply structural consistency constraints to the heavy-tailed distribution characteristics of the medical data during the model training phase. The constraints include: non-cross quantile constraints, which ensure that the predicted value of the lower quantile is not greater than the predicted value of the higher quantile through non-negative incremental cumulative parameterization; and nested risk probability monotonic constraints, which ensure that the high threshold probability is not greater than the low threshold probability through incremental parameterization and activation function. The optimization module is used to optimize the model parameters using a joint loss function that includes the structural consistency constraints; The identification module is used to generate mean prediction results, conditional quantile prediction results, and nested high-cost threshold probability prediction results based on optimized model parameters, and to identify population risks under capacity constraints.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the joint prediction method for medical expenditure based on structural consistency constraints as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: A processor and a memory, the memory storing a computer program, the processor executing the computer program to implement the joint prediction method for medical expenditures based on structural consistency constraints as described in any one of claims 1-7.

Images