A multi-factor diagnosis prediction method, device and equipment based on gradient boosting tree

Abstract

This disclosure relates to the field of time series data prediction technology, and proposes a multi-factor consultation volume prediction method, apparatus, and device based on gradient boosting tree. The method includes: acquiring historical consultation volume data from medical service institutions, performing data preprocessing to form a historical dataset; automatically extracting features from the historical dataset; using a gradient boosting tree algorithm to filter the extracted static and time-series feature factors to generate key feature factors; performing stratified sampling on the historical dataset to generate mutually exclusive data subsets; independently training multiple consultation volume prediction models using the mutually exclusive data subsets and key feature factors; and using the trained consultation volume prediction models to predict consultation volume, obtaining multiple grouped prediction information and fused consultation volume prediction information. The technical solutions provided by one or more embodiments of this disclosure can accurately uncover dynamic time-series patterns in consultation volume, providing reliable data support for the refined operation of medical service institutions.

Description

Technical Field

[0001] This disclosure relates to the field of time series data prediction technology, specifically to a multi-factor diagnostic prediction method, apparatus, and equipment based on gradient boosting trees. Background Technology

[0002] For healthcare service providers, accurately predicting future patient volume is a core task in daily operations management. As a prerequisite for refined operations, the accuracy of patient volume forecasting directly determines the effectiveness of optimizing the allocation of medical resources—doctor scheduling, nursing staff allocation, clinic usage planning, and medical equipment scheduling all require this as a basis.

[0003] Because the time series of patient visits at medical service institutions exhibits significant regularities when changing dynamically, specifically manifested as trends in visit volume and fluctuations around these trends, theoretically, accurate prediction of patient visits can be achieved by training a patient visit prediction model using a regression algorithm and constructing high-quality feature factors that characterize these trends and fluctuations.

[0004] However, existing methods for predicting patient volume in the industry suffer from three major flaws. First, patient volume sequence data generally exhibits significant nonlinearity and non-stationarity, making it susceptible to the combined effects of multiple factors such as seasonality, holiday effects, marketing activities, and the appeal of physician branding. This makes it difficult for traditional statistical models and mainstream deep learning models to accurately capture the complex interaction patterns of patient volume sequence data, resulting in insufficient robustness. Second, in fine-grained prediction scenarios, the sparsity of patient volume sequence data is particularly prominent, making it difficult for traditional prediction models to distinguish between structured zero values ​​and true zero demand, easily leading to significant prediction bias. Third, the feature engineering of existing prediction solutions largely relies on manual design, requiring manual sorting, definition, and calculation of feature factors, which is inefficient and prone to bias. Summary of the Invention

[0005] In view of this, this disclosure provides a multi-factor diagnosis volume prediction method, apparatus and equipment based on gradient boosting tree in one or more embodiments, which can accurately mine the dynamic time series pattern of diagnosis volume, solve the technical pain points of low accuracy and weak generalization ability of existing diagnosis volume prediction methods, and provide reliable data support for the refined operation of medical service institutions.

[0006] In a first aspect, this disclosure provides a multi-factor diagnosis volume prediction method based on gradient boosting trees. The method includes: acquiring historical diagnosis volume data of medical service institutions, performing data preprocessing to form a historical dataset; automatically extracting features from the historical dataset across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors; using a gradient boosting tree algorithm to filter the multi-dimensional static feature factors and time-series feature factors to generate key feature factors; performing stratified sampling on the historical dataset to generate a target number of mutually exclusive data subsets; independently training a diagnosis volume prediction model based on the gradient boosting tree algorithm using the mutually exclusive data subsets and the key feature factors; using the trained diagnosis volume prediction model to predict the diagnosis volume, obtaining grouped prediction information for the target number, and determining the diagnosis volume prediction information of the medical service institution based on the fusion result of the grouped prediction information.

[0007] Secondly, this disclosure provides a multi-factor patient volume prediction device based on gradient boosting tree. The device includes: a data acquisition unit for acquiring historical patient volume data of medical service institutions, performing data preprocessing to form a historical dataset; a feature extraction unit for automatically extracting features from the historical dataset across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors; a feature filtering unit for using a gradient boosting tree algorithm to filter the multi-dimensional static feature factors and time-series feature factors to generate key feature factors; a hierarchical sampling unit for hierarchically sampling the historical dataset to generate a target number of mutually exclusive data subsets; a group training unit for independently training a patient volume prediction model based on the mutually exclusive data subsets and the key feature factors, wherein the patient volume prediction model is constructed based on the gradient boosting tree algorithm; and an information prediction unit for using the trained patient volume prediction model to predict patient volume, obtaining group prediction information for the target number, and determining the patient volume prediction information of the medical service institution based on the fusion result of the group prediction information.

[0008] Thirdly, this disclosure provides an electronic device, which includes a memory and a processor. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the above-described multi-factor diagnostic prediction method based on gradient boosting tree.

[0009] Fourthly, this disclosure provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the aforementioned multi-factor diagnostic prediction method based on gradient boosting trees.

[0010] This disclosure provides a technical solution through one or more embodiments, focusing on the core needs of efficiently capturing trends and fluctuations in patient volume, and improving prediction accuracy and robustness. It constructs a complete closed-loop prediction scheme for patient volume time-series data, comprising four major innovations: automated feature factor construction, multi-window dynamic time-series feature factor capture, two-stage model training, and multi-model fusion prediction based on stratified sampling. Through the synergistic effect of these four innovations, the prediction results of this disclosed technical solution can effectively guide the refined operational work of medical service institutions, such as doctor scheduling, nursing staff allocation, and clinic resource planning.

[0011] The technical solution provided by one or more embodiments of this disclosure abandons the manual feature construction mode in related technologies and constructs a full-dimensional automated feature factor extraction system of "static dimension + dynamic time-series dimension". Without the need for manual intervention in feature design, sorting, and calculation, the technical solution of this disclosure can automatically derive feature factors covering seven core dimensions, including doctor factors, seasonal / periodic factors, geographic location factors, and four types of time-series related factors. The technical solution of this disclosure significantly reduces the workload of manual labor, avoids feature bias caused by manual design, and solves the problem of excessive reliance on manual labor in related technologies from the source.

[0012] The technical solution provided by one or more embodiments of this disclosure breaks through the limitations of related technologies that only focus on static cross-sectional data, and innovatively designs a dual-window framework of "time dimension + doctor's outpatient visit dimension". This technical solution specifically captures the dynamic temporal characteristics of patient volume—the time dimension can be divided into short-term and long-term windows according to the cyclical pattern of patient volume, capturing short-term fluctuations and long-term trends respectively; the doctor's outpatient visit dimension can simultaneously set short-term and long-term windows to match the recent and long-term outpatient volume patterns of doctors respectively. Based on the multiple trend / fluctuation characteristic factors automatically derived from the above windows, this technical solution can comprehensively and accurately capture the long-term and short-term dynamic temporal patterns of patient volume data.

[0013] This disclosure provides a technical solution with one or more embodiments, innovatively employing a two-stage closed-loop model training strategy. The first stage focuses on important feature identification and redundancy filtering, with the core being the selection of the best from the best. This involves filtering out a batch of key features of optimal quality from automatically generated high-quality features, eliminating redundant interference, thus improving the efficiency of subsequent model training and further enhancing feature quality, laying a data foundation of optimal quality for subsequent model training. The second stage, based on the filtered optimal features, involves stratified sampling and multi-model training, achieving synergy between "feature selection of the best from the best" and "precise model training." This addresses the deficiency of single-model training processes in related technologies that lack targeted optimization, further improving prediction accuracy.

[0014] The technical solution provided by one or more embodiments disclosed herein innovatively adopts a hierarchical sampling strategy, which can divide the dataset according to multiple dimensions and train multiple gradient boosting tree (LightGBM) models. By jointly predicting with multiple models and fusing the means, the prediction error caused by the randomness of sampling or local data bias of a single model is effectively offset, significantly improving the robustness and accuracy of the prediction results and solving the deficiency of weak generalization ability of a single model in related technologies. Attached Figure Description

[0015] The features and advantages of the embodiments of this disclosure will be more clearly understood by referring to the accompanying drawings, which are illustrative and should not be construed as limiting the present disclosure in any way. In the drawings: Figure 1 This illustration shows a step diagram of a multi-factor diagnostic prediction method based on gradient boosting tree in one embodiment of the present disclosure; Figure 2 This diagram illustrates the functional modules of a multi-factor diagnostic prediction device based on a gradient boosting tree in one embodiment of the present disclosure. Figure 3 A schematic diagram of the structure of an electronic device according to one embodiment of the present disclosure is shown. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0017] Please see Figure 1 The present disclosure provides a multi-factor diagnostic prediction method based on gradient boosting tree, which may include the following steps.

[0018] S1: Obtain historical patient volume data from medical service institutions, perform data preprocessing, and form a historical dataset.

[0019] In this embodiment, the historical consultation volume data can be sourced from electronic medical record systems, appointment booking platforms, and clinic operation management systems of medical service institutions. Preferably, historical consultation volume data for the past three years (covering the complete seasonal cycle and holiday distribution) is selected as the training scope for the model. Preferably, the historical consultation volume data undergoes data preprocessing, which can involve automated identification and correction of missing values, outliers, and other noisy data. Missing values ​​can be automatically filled using linear interpolation, and outliers can be automatically identified using the 3σ principle and then automatically corrected using industry experience values. The preprocessed historical consultation volume data provides a high-quality historical dataset, facilitating subsequent model training.

[0020] S2: For the historical dataset, perform automated feature extraction across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors.

[0021] In this embodiment, a multi-dimensional automated feature engineering strategy can be designed to address the changing characteristics of historical patient volume data. After determining multiple core feature dimensions (e.g., doctor factor dimension, seasonal / periodic factor dimension, geographic location factor dimension, time-related trend factor dimension, time-related fluctuation factor dimension, doctor-related trend factor dimension, doctor-related fluctuation factor dimension, etc.), differentiated automated processing strategies can be configured for different dimensions to achieve automated derivation and extraction of static and time-series feature factors. Static feature factors can be extracted using automated direct feature extraction, while time-series feature factors are calculated using a multi-window automated method to ensure the accuracy of feature representation. Since no manual intervention is required in feature design, dimension determination, and processing rule setting throughout the process, this fully demonstrates the innovative aspect of "automated extraction of time-series special factors" in this disclosure.

[0022] In some implementations, the static feature factors include at least one of a doctor factor, a periodicity factor, and a geographic location factor. The automated feature extraction across multiple feature dimensions for the historical dataset, generating multi-dimensional static feature factors and time-series feature factors, includes at least one of the following: automatically extracting unique doctor identifiers at the individual doctor level and automatically grouping doctors using a hierarchical clustering algorithm to determine the doctor factor; automatically extracting the month information of the target date to generate a month factor; automatically extracting the ordinal number of the target date within the current month to generate a date factor; automatically extracting the ordinal number of the target date within the current week according to the natural week to generate a week factor; automatically determining whether the target sample date is a Saturday or Sunday to generate a weekend / holiday factor; determining the periodicity factor based on the month factor, the date factor, the week factor, and the weekend / holiday factor; and automatically extracting and standardizing the city identifier of the target clinic to determine the geographic location factor.

[0023] Specifically, the doctor factor is used to capture the impact of doctor brand effect on users' medical treatment decisions within healthcare service institutions. The doctor factor can be generated in the following ways: At the individual doctor level, a unique doctor identifier (ID) is automatically extracted to accurately capture differences in patient volume among different doctors; through a hierarchical clustering algorithm, doctors are automatically grouped using factors such as years of service, professional title, past patient satisfaction, and specialty as clustering input features. Doctors with similar work experience, professional abilities, and past records are assigned a unified group label, automatically capturing differences in patient volume between different doctor groups without the need for manual group division.

[0024] Seasonal / periodic factors are used to capture the seasonal / periodic effects of user medical visit behavior. These factors can be generated in the following ways: Automatically extracting the month information of the target date to capture the cyclical characteristics within the natural year, resulting in a month factor (values ​​1-12), thus accurately matching the cyclical patterns of seasonal epidemics and student summer and winter vacations within the natural year; automatically extracting the target date's sequence number within the month to obtain a date factor (values ​​1-31, automatically adjusted according to the actual number of days in the month), thus capturing monthly cyclical characteristics such as peak follow-up visits at the beginning of the month and low visits at the end of the month; automatically extracting the target date's sequence number within the week according to the natural week, generating a week factor (values ​​1-7), thus capturing differences in the distribution of medical visits within the week; and automatically determining whether the target sample date is a Saturday or Sunday to generate a weekend factor (value 1 if yes, 0 otherwise), thus accurately matching the characteristics of users having more free time and higher demand for medical visits on weekends.

[0025] Geographic location factors are used to capture differences in healthcare-seeking behavior patterns among users in different cities. Some healthcare service providers may be located in multiple cities. The climate and socio-cultural differences in these cities may influence user healthcare-seeking behavior. Geographic location factors can be represented as city tags, which can be automatically categorized based on the economic level, climate type (tropical / subtropical / temperate, etc.), and population density of the cities where the healthcare service providers are located. By automatically linking historical patient volume data from different cities, these city tags can accurately capture differences in user behavior patterns such as frequency of visits and preferred times of visit due to geographical differences, without the need for manual city categorization.

[0026] In some implementations, the time-series feature factors include at least one of time trend factors, time fluctuation factors, doctor trend factors, and doctor fluctuation factors. The step of automatically extracting features from the historical dataset across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors includes at least one of the following: automatically capturing the quantitative features of the patient volume trend of the target sample within a specific time period based on a preset time dimension window, and determining the time trend factor, wherein the time trend factor includes at least one of the patient volume mean, patient volume difference, patient volume difference mean, and patient volume difference mean standard deviation; automatically capturing the quantitative features of the patient volume fluctuation rate of the target sample within a specific time period based on a preset time dimension window, and determining the time trend factor. The time fluctuation factor includes at least one of the following: standard deviation of consultation volume, range of consultation volume, and coefficient of variation of consultation volume over time. Based on a preset doctor's outpatient dimension window, the quantitative characteristics of the consultation volume trend of the doctor corresponding to the target sample within a specific number of outpatient periods are automatically captured to determine the doctor trend factor, which includes at least one of the following: mean of doctor's consultation volume, difference of consultation volume, mean of consultation volume difference, and standard deviation of mean of consultation volume difference. Based on a preset doctor's outpatient dimension window, the quantitative characteristics of the fluctuation rate of consultation volume of the doctor corresponding to the target sample within a specific number of outpatient periods are automatically captured to determine the doctor fluctuation factor, which includes at least one of the following: standard deviation of doctor's consultation volume, range of doctor's consultation volume, and coefficient of variation of doctor's consultation volume over time.

[0027] Specifically, current mainstream methods for predicting patient volume in the industry mostly rely on basic static information such as date, day of the week, consultation time (morning / afternoon / evening), and specific time (hour / minute, or hour / minute / second) as input data. They focus only on basic attributes at a single point in time and completely lack records of sequential changes over time. For example, key time-series information such as the rise and fall of patient volume in the past week, the fluctuation range of patient volume in the past month, and the changing trend of doctors' recent patient visits are not included. This directly leads to the model's inability to effectively identify the core factors driving changes in patient volume—it is difficult to capture the volatile impact caused by short-term sudden factors, and it is also difficult to uncover the trend patterns formed in the long-term development process, ultimately resulting in insufficient accuracy and stability of the prediction results.

[0028] This disclosure innovatively designs a multi-window automated time-series feature capture method, breaking through the limitation of existing technologies that only focus on static cross-sectional data. It can comprehensively and accurately capture the potential impact of the evolution and fluctuation patterns of consultation volume trends on future consultation volume—it can capture both the trend / fluctuation characteristics of short-term dynamic changes in consultation volume and the trend / fluctuation characteristics of long-term consultation volume development trajectory.

[0029] Among these, differentiated time windows and doctor's consultation frequency windows are key technical means to achieve this goal. Short-term time windows (e.g., selecting the most recent 1–4 weeks) have been proven to effectively cover fluctuations in consultation volume caused by statutory holidays and short-term public health emergencies. Long-term time windows (e.g., selecting the most recent 8–12 weeks) can fully cover seasonal changes in consultation volume caused by seasonal transitions such as influenza in spring and summer and respiratory diseases in autumn and winter. Short-term doctor consultation windows (e.g., the most recent 1–10 consultation records) can match the recent reputation dissemination cycle of doctors, while long-term doctor consultation windows (e.g., the most recent 10–50 consultation records) can cover patient follow-up consultation cycles and doctor brand building cycles. The parameter settings and feature derivation for all windows can be completed automatically without manual adjustment.

[0030] To further illustrate the application logic of the multi-window method, we will take two typical characteristics, the average diagnostic volume factor and the diagnostic volume fluctuation factor, as examples.

[0031] Under the time-related trend factor dimension, the dynamic trend of short-term increase or decrease in consultation volume can be automatically captured by the "average consultation volume in the last 2 weeks", and the steady-state development direction of long-term consultation volume can be automatically captured by the "average consultation volume in the last 12 weeks". Under the time-related volatility factor dimension, the volatility amplitude of short-term consultation volume can be automatically measured by the "standard deviation of consultation volume in the last 2 weeks", and the volatility stability of long-term consultation volume can be automatically measured by the "standard deviation of consultation volume in the last 12 weeks".

[0032] Under the doctor-related trend factor dimension, the recent dynamic trend of the doctor's patient traffic can be automatically captured by the "average number of visits in the doctor's last 5 visits", and the long-term patient traffic baseline can be automatically captured by the "average number of visits in the doctor's last 30 / 50 visits". Under the doctor-related fluctuation factor dimension, the recent fluctuation degree of the number of visits can be automatically reflected by the "coefficient of variation of the number of visits in the doctor's last 5 visits", and the long-term fluctuation pattern of the number of visits can be automatically reflected by the "coefficient of variation of the number of visits in the doctor's last 30 / 50 visits".

[0033] The construction method of dimensions and their feature factors based on multi-window construction is as follows (fully automated, no manual intervention required).

[0034] Time-related trend factors: Based on the set time dimension window, the daily diagnosis data of the target sample within the corresponding time period is automatically traced back; the diagnosis difference is automatically calculated on a daily basis, and the mean of the diagnosis difference is the automatic average of the daily differences within the specified window; during the traversal calculation, it is automatically classified and processed by week, weekday, and weekend to eliminate noise interference from different cycle types, and finally automatically forms a set of factors covering long-term and short-term trends without the need for manual calculation and classification.

[0035] Time-related fluctuation factors: Daily diagnostic data is automatically traced based on a time-dimensional window. The coefficient of variation of diagnostic data over time is automatically calculated according to "standard deviation / mean" (eliminating the influence of dimensions). The range is the automatic difference between the maximum and minimum daily diagnostic data within the window. By automatically comparing fluctuation indicators under different windows, the difference between short-term sudden fluctuations and long-term stable fluctuations is accurately characterized, providing a more comprehensive fluctuation representation for the model. All calculation processes are completed automatically.

[0036] Doctor-related trend factors: Based on the doctor's outpatient visit dimension window, the system automatically tracks the outpatient visit data of the target doctor (automatically excluding invalid outpatient visit records such as temporary substitutes); the outpatient visit difference is automatically calculated according to the outpatient visit order, and the mean difference is the automatic average of all adjacent differences within the window; by automatically comparing short-term (1-10 times) and long-term (10-50 times) trends, the system captures the trend changes brought about by the doctor's recent patient traffic fluctuations and long-term reputation, without the need for manual intervention in data tracking and calculation.

[0037] Fluctuation factors related to doctors: Based on the doctor's outpatient volume window, the effective outpatient volume data is automatically traced. The range of doctor's outpatient volume is the difference between the maximum and minimum outpatient volume within the window. The coefficient of variation is also automatically calculated according to "standard deviation / mean" (automatically excluding extreme cases with a mean of 0). By quantifying the degree of fluctuation in different outpatient cycles, the model is provided with a basis for distinguishing between "stable high outpatient volume doctors" and "fluctuating high outpatient volume doctors". All calculation processes are completed automatically.

[0038] In summary, the multi-window factor construction method, by designing the window length for time and outpatient frequency in a hierarchical manner, can automatically extract dynamic temporal features of patient volume. This provides more comprehensive and discriminative feature inputs for patient volume prediction models, improving the model's explanatory power and prediction accuracy for changes in patient volume. At the same time, it completely solves the shortcomings of existing technologies that only focus on static cross-sectional data and cannot capture dynamic temporal patterns. Together with the innovative points of automated feature factor construction, it forms a high-quality feature input system.

[0039] S3: The gradient boosting tree algorithm is used to filter the multi-dimensional static feature factors and the time-series feature factors to generate key feature factors.

[0040] In this embodiment, corresponding to the first stage of the innovative "two-stage model training" of this disclosure, a gradient boosting tree algorithm can be used to perform secondary screening of static feature factors and time-series feature factors to identify key feature factors. This stage does not involve training the actual diagnostic prediction model, and the entire screening process is automated. The core objective is to select the best from the best, identifying the highest quality key features from the automatically generated high-quality features, eliminating redundant interference, and further strengthening the innovative advantage of "no human intervention required." Simultaneously, this stage significantly improves the concentration of feature quality, enhancing the efficiency and quality of subsequent model training. This lays a high-quality data foundation for the stratified sampling multi-model training in the second stage, overcoming the deficiency of existing technologies that lack targeted screening after automated feature generation.

[0041] It's important to note that the core reason for choosing the LightGBM gradient boosting tree algorithm is its superior ability to handle high-dimensional sparse features. This allows it to adapt to the massive amounts of multi-dimensional, multi-window features generated by automated time-series factor extraction steps, avoiding the computational efficiency degradation caused by the curse of dimensionality. Furthermore, the algorithm's built-in node splitting gain quantization evaluation mechanism accurately characterizes the contribution of each feature factor to the diagnostic prediction target, providing a reliable basis for subsequent feature selection. Moreover, the entire evaluation and screening process requires no manual intervention, achieving automated feature refinement.

[0042] In some implementations, the step of using the gradient boosting tree algorithm to filter the multi-dimensional static and temporal feature factors to generate key feature factors includes: using all the temporal feature factors and the historical dataset as input, and employing the gradient boosting tree algorithm to build an initial screening and prediction model; during the training of the screening and prediction model, no validation dataset is set, and iterative training is performed only based on the training set data until the model loss function value no longer decreases and the model reaches a convergent state, at which point the training is terminated; after the screening and prediction model converges, the node split gain value of each candidate feature factor during the training of the gradient boosting tree algorithm is extracted, and the total split gain of each candidate feature factor in all decision trees is accumulated; all candidate feature factors are sorted from high to low according to the total split gain value, and the gain ratio of each candidate feature factor after sorting is accumulated sequentially to generate a cumulative contribution percentage curve of the candidate feature factors; the slope change trend of the cumulative contribution percentage curve is monitored, and the critical point where the slope of the curve changes from steep to gentle and tends to be stable is selected as the feature cutoff point; the candidate feature factors before the feature cutoff point are retained and determined as the key feature factors.

[0043] Specifically, this stage adopts a training mode without a validation set, which is essentially a further selection of the best of the automatically extracted high-quality feature factors. This maximizes the use of the information value of historical datasets, avoids interference from the validation set division on feature importance evaluation, and ensures the objectivity of the screening results.

[0044] The first stage, corresponding to the innovation of the two-stage model training, uses fully automatically extracted time-series feature factors and high-quality historical datasets as input. The LightGBM algorithm is employed to automatically build the framework for the selection and prediction model. The initial parameters of the selection and prediction model can be automatically set to a learning rate of 0.1, a maximum tree depth of 6, and 31 leaf nodes per tree. The input data needs to undergo automated feature encoding (e.g., one-hot encoding of doctor IDs and city tags) and normalization (e.g., standardizing trend and volatility factors to the 0-1 range) in advance to ensure the stability of model training, eliminating the need for manual data preprocessing and parameter setting.

[0045] The first stage of the two-stage model training innovation utilizes only the training set data without a validation set to maximize data information utilization, ensuring the objectivity of feature importance assessment. Mean squared error (MSE) is used as the loss function, and model parameters are automatically updated iteratively using gradient descent. When the change in the loss function value over five consecutive iterations is less than 1e-5, the model is automatically considered to have reached convergence, terminating training and avoiding feature evaluation bias caused by overtraining. No manual monitoring or intervention is required during the training process, ensuring the accuracy of feature selection throughout.

[0046] In the first stage of the two-stage model training innovation, after model convergence, the node splitting gain value of each feature factor in all decision trees can be automatically extracted. The total splitting gain is the sum of the gains of all splitting nodes in which the feature participates. Feature factors are automatically sorted from high to low according to the total splitting gain value, and the gain ratio of each feature (total gain of a single feature / sum of total gains of all features) is automatically calculated, which lays the foundation for subsequent cumulative contribution analysis and optimal feature identification. No manual calculation or sorting is required, further strengthening the innovative advantage of the first stage of "selecting the best from the best features," which differs from the traditional mode of existing technologies that require manual feature selection and do not achieve the selection of the best from the best.

[0047] In the first stage of the two-stage model training innovation, the gain percentage is automatically accumulated according to the sorted feature order, generating a cumulative contribution percentage curve. The curve slope (the difference between the cumulative percentage of the subsequent feature and the previous feature) is automatically calculated using the "adjacent feature gain difference method." When the slope of three consecutive features is less than 0.01, the curve is automatically determined to have entered a stable interval. Taking the starting feature of this interval as the cutoff point automatically retains the feature factors before the cutoff point, forming a key feature set (achieving feature selection from the best, eliminating redundant interference, and improving subsequent training efficiency).

[0048] S4: Perform stratified sampling on the historical dataset to generate a target number of mutually exclusive data subsets.

[0049] In this embodiment, corresponding to the second stage of the innovative "two-stage model training" of this disclosure, multiple diagnostic prediction models can be independently trained based on mutually exclusive data subsets obtained by stratified sampling, ultimately achieving multi-model fusion prediction. Stratified sampling ensures that the feature distribution of each training data subset is highly consistent with the feature distribution of the overall dataset, avoiding insufficient model generalization ability caused by sampling bias.

[0050] In some implementations, the step of performing stratified sampling on the historical dataset to generate a target number of mutually exclusive data subsets includes: performing stratified sampling on the historical dataset proportionally according to five dimensions: city, time, clinic, department, and doctor, to ensure that the distribution characteristics of each group of mutually exclusive data subsets are consistent with the overall distribution characteristics of the historical dataset, thereby obtaining the target number of mutually exclusive data subsets.

[0051] In a practical application example, the system automatically stratifies data according to five dimensions: city, time (year / quarter / month), clinic, department, and doctor. Within each dimension, it can automatically sample according to the proportion of different categories (e.g., the sample proportion of first-tier cities is consistent with the overall dataset). The error between the sample sizes of the final 10 mutually exclusive subsets is controlled within ±5%, ensuring that each data subset can represent the overall data distribution characteristics and improving the model's generalization ability.

[0052] S5: Using the mutually exclusive data subset and the key feature factors, independently train the diagnostic prediction model for the target number of cases. The diagnostic prediction model is constructed based on the gradient boosting tree algorithm.

[0053] In this embodiment, corresponding to the second stage of the innovative "two-stage model training" of this disclosure, based on the mutually exclusive data subset obtained by stratified sampling, the LightGBM algorithm is selected to automatically build multiple sets of diagnostic volume prediction models, which can balance model training efficiency and nonlinear fitting ability.

[0054] In some implementations, the independent training of the diagnostic prediction model for the target quantity using the mutually exclusive data subsets and the key feature factors includes: using the key feature factors and the historical dataset as input, employing the gradient boosting tree algorithm to build an initial diagnostic prediction model for the target quantity; performing iterative training for the target quantity rounds, where each round of training automatically selects a set of mutually exclusive data subsets as a validation set and the remaining mutually exclusive data subsets as a training set, and monitors the loss function value of the validation set in real time during training; when the loss function value of the validation set no longer decreases for several consecutive rounds, the current round of training is terminated; after completing the training for the target quantity rounds, a stable diagnostic prediction model for the target quantity is obtained.

[0055] In a practical application example, using a carefully selected set of key feature factors (optimal quality automated features) and a high-quality historical dataset as input, the LightGBM algorithm can be used to automatically build a diagnostic prediction model. Leveraging the key feature factor set allows for more accurate multi-model training. The key feature factors are automatically associated and concatenated with the original dataset according to sample IDs to form the input data for model training. Simultaneously, initial regularization parameters (L1=0.01, L2=0.1) can be automatically set to prevent overfitting.

[0056] The system automatically performs 10 rounds of iterative training for the diagnostic volume prediction model, selecting one set as the validation set and nine sets as the training set in each round. During training, the MSE value of the validation set is monitored in real time. Training automatically terminates when the MSE of the validation set stops decreasing after five consecutive rounds. Model parameters are automatically saved after each training round. After 10 rounds of training, 10 stable and independent LightGBM models (covering feature patterns from different subsets of data) are obtained.

[0057] S6: Use the trained consultation volume prediction model to predict consultation volume, obtain grouped prediction information of the target number, and determine the consultation volume prediction information of the medical service institution based on the fusion result of the grouped prediction information.

[0058] In this embodiment, by using multiple models to make predictions together, the prediction error of a single model can be effectively offset, significantly improving the stability, noise resistance and accuracy of the prediction results.

[0059] In some implementations, the step of using the trained diagnosis volume prediction model to predict the diagnosis volume, obtaining grouped prediction information for a target number, and determining the diagnosis volume prediction information for the medical service institution based on the fusion result of the grouped prediction information includes: removing abnormal results that exceed a reasonable range from the grouped prediction information; calculating the arithmetic mean of the remaining valid results of the grouped prediction information, and using the arithmetic mean as the diagnosis volume prediction information.

[0060] Specifically, during prediction, the automated feature extraction method in step S2 can automatically extract feature factors of the object to be predicted (e.g., a doctor, a city, and a date), ensuring that the dimensions, encoding methods, and normalization standards of the predicted features are completely consistent with those of the training features (e.g., the extraction rules for the month factor and week factor of the predicted date are the same as those in the training stage), thus avoiding prediction bias caused by feature distribution drift.

[0061] By simultaneously inputting predictive features into multiple (e.g., 10) independent patient volume prediction models, multiple sets of initial prediction information can be automatically obtained. First, abnormal results that exceed reasonable ranges (such as predicted values ​​less than 0 or greater than 1.5 times the historical maximum daily patient volume) are automatically removed from the multiple sets of initial prediction information. Then, the arithmetic mean of the remaining valid results can be calculated to determine the final patient volume prediction result.

[0062] Multi-model fusion prediction can offset the sampling bias of a single model, improving the stability and reliability of the final diagnostic prediction. The entire inference and fusion process requires no human intervention, achieving a dual improvement in prediction robustness and accuracy.

[0063] This disclosure provides a technical solution through one or more embodiments, focusing on the core needs of efficiently capturing trends and fluctuations in patient volume, and improving prediction accuracy and robustness. It constructs a complete closed-loop prediction scheme for patient volume time-series data, comprising four major innovations: automated feature factor construction, multi-window dynamic time-series feature factor capture, two-stage model training, and multi-model fusion prediction based on stratified sampling. Through the synergistic effect of these four innovations, the prediction results of this disclosed technical solution can effectively guide the refined operational work of medical service institutions, such as doctor scheduling, nursing staff allocation, and clinic resource planning.

[0064] The technical solution provided by one or more embodiments of this disclosure abandons the manual feature construction mode in related technologies and constructs a full-dimensional automated feature factor extraction system of "static dimension + dynamic time-series dimension". Without the need for manual intervention in feature design, sorting, and calculation, the technical solution of this disclosure can automatically derive feature factors covering seven core dimensions, including doctor factors, seasonal / periodic factors, geographic location factors, and four types of time-series related factors. The technical solution of this disclosure significantly reduces the workload of manual labor, avoids feature bias caused by manual design, and solves the problem of excessive reliance on manual labor in related technologies from the source.

[0065] The technical solution provided by one or more embodiments of this disclosure breaks through the limitations of related technologies that only focus on static cross-sectional data, and innovatively designs a dual-window framework of "time dimension + doctor's outpatient visit dimension". This technical solution specifically captures the dynamic temporal characteristics of patient volume—the time dimension can be divided into short-term and long-term windows according to the cyclical pattern of patient volume, capturing short-term fluctuations and long-term trends respectively; the doctor's outpatient visit dimension can simultaneously set short-term and long-term windows to match the recent and long-term outpatient volume patterns of doctors respectively. Based on the multiple trend / fluctuation characteristic factors automatically derived from the above windows, this technical solution can comprehensively and accurately capture the long-term and short-term dynamic temporal patterns of patient volume data.

[0066] This disclosure provides a technical solution with one or more embodiments, innovatively employing a two-stage closed-loop model training strategy. The first stage focuses on important feature identification and redundancy filtering, with the core being the selection of the best from the best. This involves filtering out a batch of key features of optimal quality from automatically generated high-quality features, eliminating redundant interference, thus improving the efficiency of subsequent model training and further enhancing feature quality, laying a data foundation of optimal quality for subsequent model training. The second stage, based on the selected optimal features, conducts stratified sampling multi-model training, achieving synergy between "feature selection of the best from the best" and "precise model training." This addresses the deficiency of single-model training processes in related technologies that lack targeted optimization, further improving prediction accuracy.

[0067] The technical solution provided by one or more embodiments disclosed herein innovatively adopts a hierarchical sampling strategy, which can divide the dataset according to multiple dimensions and train multiple gradient boosting tree (LightGBM) models. By jointly predicting with multiple models and fusing the means, the prediction error caused by the randomness of sampling or local data bias of a single model is effectively offset, significantly improving the robustness and accuracy of the prediction results and solving the deficiency of weak generalization ability of a single model in related technologies.

[0068] Please see Figure 2 This disclosure also provides a multi-factor diagnostic prediction device based on gradient boosting trees, the device comprising: The data acquisition unit 100 is used to acquire historical patient volume data of medical service institutions, and after data preprocessing, form a historical dataset; The feature extraction unit 200 is used to automatically extract features from the historical dataset in multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors. The feature filtering unit 300 is used to filter the multi-dimensional static feature factors and the temporal feature factors using the gradient boosting tree algorithm to generate key feature factors. The hierarchical sampling unit 400 is used to perform hierarchical sampling on the historical dataset to generate a target number of mutually exclusive data subsets; Group training unit 500 is used to independently train a diagnostic prediction model for a target number of cases using the mutually exclusive data subset and the key feature factors. The diagnostic prediction model is constructed based on the gradient boosting tree algorithm. The information prediction unit 600 is used to predict the number of patients using the trained diagnosis prediction model, obtain group prediction information of the target number, and determine the diagnosis prediction information of the medical service institution based on the fusion result of the group prediction information.

[0069] In one embodiment, the static feature factor includes at least one of a doctor factor, a periodic factor, and a geographic location factor, and the feature extraction unit 200 is specifically used for: Using individual doctors as the granularity, the system automatically extracts unique identifiers for each doctor and uses a hierarchical clustering algorithm to automatically group doctors and determine the doctor factors. Automatically extract the month information of the target date to generate a month factor; automatically extract the ordinal number of the target date in the current month to generate a date factor; automatically extract the ordinal number of the target date in the current week according to the natural week to generate a week factor; automatically determine whether the target sample date is a Saturday or Sunday to generate a weekend factor; determine the periodicity factor based on the month factor, the date factor, the week factor, and the weekend factor. The city identifier of the target clinic is automatically extracted and standardized and coded to determine the geographical location factor.

[0070] In one embodiment, the time-series feature factor includes at least one of a time trend factor, a time fluctuation factor, a doctor trend factor, and a doctor fluctuation factor, and the feature extraction unit 200 is specifically used for: Based on a preset time dimension window, the quantitative characteristics of the diagnosis volume trend of the target sample within a specific time period are automatically captured, and the time trend factor is determined. The time trend factor includes at least one of the following: the diagnosis volume mean, the diagnosis volume difference, the diagnosis volume difference mean, and the standard deviation of the diagnosis volume difference mean. Based on a preset time dimension window, the quantitative characteristics of the fluctuation rate of the target sample in the diagnosis volume within a specific time period are automatically captured, and the time fluctuation factor is determined. The time fluctuation factor includes at least one of the diagnosis volume standard deviation, diagnosis volume range, and diagnosis volume time series variation coefficient. Based on a preset doctor outpatient visit dimension window, the system automatically captures the quantitative characteristics of the outpatient visit trend of the corresponding doctor in the target sample within a specific outpatient visit period, and determines the doctor trend factor. The doctor trend factor includes at least one of the following: doctor outpatient visit mean, outpatient visit difference, outpatient visit difference mean, and outpatient visit difference mean standard deviation. Based on a preset doctor outpatient visit dimension window, the system automatically captures the quantitative characteristics of the fluctuation rate of the doctor's outpatient visits within a specific outpatient visit period for the target sample, and determines the doctor fluctuation factor. The doctor fluctuation factor includes at least one of the following: doctor outpatient visit standard deviation, doctor outpatient visit range, and doctor outpatient visit time series variation coefficient.

[0071] In one embodiment, the feature selection unit 300 is specifically used for: taking all the time-series feature factors and the historical dataset as input, and using the gradient boosting tree algorithm to build an initial selection and prediction model; during the training process of the selection and prediction model, no validation dataset is set, and iterative training is performed only based on the training set data until the model loss function value no longer decreases and the model reaches a convergent state, at which point the training is terminated; after the selection and prediction model converges, the node split gain value of each candidate feature factor during the training process of the gradient boosting tree algorithm is extracted, and the total split gain of each candidate feature factor in all decision trees is accumulated; all candidate feature factors are sorted from high to low according to the total split gain value, and the gain ratio of each candidate feature factor after sorting is accumulated in turn to generate a cumulative contribution percentage curve of the candidate feature factors; the slope change trend of the cumulative contribution percentage curve is monitored, and the critical point where the slope of the curve changes from steep to gentle and tends to be stable is selected as the feature cutoff point; the candidate feature factors before the feature cutoff point are retained and determined as the key feature factors.

[0072] In one embodiment, the stratified sampling unit 400 is specifically used to: perform stratified sampling of the historical dataset according to five dimensions: city, time, clinic, department, and doctor, to ensure that the distribution characteristics of each group of mutually exclusive data subsets are consistent with the overall distribution characteristics of the historical dataset, thereby obtaining a target number of mutually exclusive data subsets.

[0073] In one implementation, the group training unit 500 is specifically used to: take the key feature factors and the historical dataset as input, and use the gradient boosting tree algorithm to build an initial diagnostic prediction model for the target number; perform target number rounds of iterative training, in which each round of training automatically selects a set of mutually exclusive data subsets as a validation set and the remaining mutually exclusive data subsets as a training set, and monitors the loss function value of the validation set in real time during the training process. When the loss function value of the validation set no longer decreases for several consecutive rounds, the current round of training is terminated; after completing the target number rounds of training, a stable diagnostic prediction model for the target number is obtained.

[0074] In one embodiment, the information prediction unit 600 is specifically used to: remove abnormal results that exceed a reasonable range from the grouped prediction information; calculate the arithmetic mean of the remaining valid results of the grouped prediction information, and use the arithmetic mean as the diagnostic prediction information.

[0075] The various units described in the above embodiments can be implemented by computer chips or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a smartphone, a tablet computer, a wearable device, or any combination of these devices.

[0076] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0077] Please see Figure 3 This disclosure also provides an electronic device, which includes a memory and a processor. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the above-described multi-factor diagnostic prediction method based on gradient boosting tree.

[0078] This disclosure also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements the above-described multi-factor diagnostic prediction method based on gradient boosting trees.

[0079] The processor can be a central processing unit (CPU). It can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations thereof.

[0080] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods in the above-described embodiments.

[0081] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0083] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, embodiments of apparatus, devices, and storage media are basically similar to method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0084] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0085] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-factor diagnostic prediction method based on gradient boosting trees, characterized in that, The method includes: Historical patient volume data from medical service institutions is obtained, and after data preprocessing, a historical dataset is formed. For the historical dataset, automated feature extraction is performed across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors. The gradient boosting tree algorithm is used to filter the multi-dimensional static feature factors and the temporal feature factors to generate key feature factors. The historical dataset is subjected to stratified sampling to generate a target number of mutually exclusive data subsets; Using the mutually exclusive data subset and the key feature factors, the diagnostic prediction model for the target number of patients is independently trained, and the diagnostic prediction model is constructed based on the gradient boosting tree algorithm. The trained patient volume prediction model is used to predict patient volume, and group prediction information of the target number is obtained. Based on the fusion result of the group prediction information, the patient volume prediction information of the medical service institution is determined.

2. The method according to claim 1, characterized in that, The static feature factors include at least one of doctor factors, periodic factors, and geographic location factors. For the historical dataset, automated feature extraction is performed across multiple feature dimensions to generate multi-dimensional static feature factors and time-series feature factors, including at least one of the following: Using individual doctors as the granularity, the system automatically extracts unique identifiers for each doctor and uses a hierarchical clustering algorithm to automatically group doctors and determine the doctor factors. Automatically extract the month information of the target date to generate a month factor; automatically extract the ordinal number of the target date in the current month to generate a date factor; automatically extract the ordinal number of the target date in the current week according to the natural week to generate a week factor; automatically determine whether the target sample date is a Saturday or Sunday to generate a weekend factor; determine the periodicity factor based on the month factor, the date factor, the week factor, and the weekend factor. The city identifier of the target clinic is automatically extracted and standardized and coded to determine the geographical location factor.

3. The method according to claim 1 or 2, characterized in that, The time-series feature factors include at least one of time trend factors, time fluctuation factors, doctor trend factors, and doctor fluctuation factors. The automated feature extraction process, performed on the historical dataset across multiple feature dimensions, generates multi-dimensional static feature factors and time-series feature factors, including at least one of the following: Based on a preset time dimension window, the quantitative characteristics of the diagnosis volume trend of the target sample within a specific time period are automatically captured, and the time trend factor is determined. The time trend factor includes at least one of the following: the diagnosis volume mean, the diagnosis volume difference, the diagnosis volume difference mean, and the standard deviation of the diagnosis volume difference mean. Based on a preset time dimension window, the quantitative characteristics of the fluctuation rate of the target sample in the diagnosis volume within a specific time period are automatically captured, and the time fluctuation factor is determined. The time fluctuation factor includes at least one of the diagnosis volume standard deviation, diagnosis volume range, and diagnosis volume time series variation coefficient. Based on a preset doctor outpatient visit dimension window, the system automatically captures the quantitative characteristics of the outpatient visit trend of the corresponding doctor in the target sample within a specific outpatient visit period, and determines the doctor trend factor. The doctor trend factor includes at least one of the following: doctor outpatient visit mean, outpatient visit difference, outpatient visit difference mean, and outpatient visit difference mean standard deviation. Based on a preset doctor outpatient visit dimension window, the system automatically captures the quantitative characteristics of the fluctuation rate of the doctor's outpatient visits within a specific outpatient visit period for the target sample, and determines the doctor fluctuation factor. The doctor fluctuation factor includes at least one of the following: doctor outpatient visit standard deviation, doctor outpatient visit range, and doctor outpatient visit time series variation coefficient.

4. The method according to claim 3, characterized in that, The gradient boosting tree algorithm is used to filter the multi-dimensional static feature factors and the temporal feature factors to generate key feature factors, including: Using all the aforementioned time-series feature factors and the aforementioned historical dataset as input, an initial screening and prediction model is built using the gradient boosting tree algorithm. During the training process of the screening and prediction model, no verification dataset is set. Iterative training is performed only based on the training set data until the model loss function value no longer decreases and the model reaches a convergent state, at which point the training is terminated. After the screening and prediction model converges, the node split gain value of each candidate feature factor during the training process of the gradient boosting tree algorithm is extracted, and the total split gain of each candidate feature factor in all decision trees is calculated cumulatively. Sort all candidate feature factors from high to low according to their total splitting gain value, and then sum up the gain percentage of each candidate feature factor after sorting to generate a cumulative contribution percentage curve of the candidate feature factors. Monitor the slope change trend of the cumulative contribution percentage curve, and select the critical point where the slope of the curve changes from steep to gentle and tends to be stable as the characteristic cutoff point; Candidate feature factors before the feature cutoff point are retained and determined as the key feature factors.

5. The method according to claim 1, characterized in that, The step of performing stratified sampling on the historical dataset to generate a target number of mutually exclusive data subsets includes: The historical dataset is stratified and sampled proportionally according to five dimensions: city, time, clinic, department, and doctor. This ensures that the distribution characteristics of each mutually exclusive data subset are consistent with the overall distribution characteristics of the historical dataset, thereby obtaining the target number of mutually exclusive data subsets.

6. The method according to claim 5, characterized in that, The step of independently training a diagnostic volume prediction model for a target number of patients using the mutually exclusive data subset and the key feature factors includes: Using the key feature factors and the historical dataset as input, a gradient boosting tree algorithm is used to build an initial diagnostic prediction model for the target number of patients. The training process is performed in a target number of rounds. In each round, a set of mutually exclusive data subsets is automatically selected as the validation set and the remaining mutually exclusive data subsets are selected as the training set. The loss function value of the validation set is monitored in real time during the training process. When the loss function value of the validation set no longer decreases for several consecutive rounds, the training is terminated. After completing the training for the target number of rounds, a stable diagnostic prediction model for the target number of cases is obtained.

7. The method according to claim 1, characterized in that, The step of using the trained patient volume prediction model to predict patient volume, obtaining grouped prediction information for the target number, and determining the patient volume prediction information for the medical service institution based on the fusion result of the grouped prediction information includes: Remove abnormal results from the grouped prediction information that exceed a reasonable range; The arithmetic mean of the remaining valid results of the grouped prediction information is calculated, and the arithmetic mean is used as the diagnosis volume prediction information.

8. A multi-factor diagnostic prediction device based on gradient boosting tree, characterized in that, The device includes: The data acquisition unit is used to acquire historical patient volume data from medical service institutions, and after data preprocessing, form a historical dataset. The feature extraction unit is used to automatically extract features from the historical dataset in multiple feature dimensions, generating multi-dimensional static feature factors and time-series feature factors. The feature filtering unit is used to filter the multi-dimensional static feature factors and the temporal feature factors using the gradient boosting tree algorithm to generate key feature factors. A stratified sampling unit is used to perform stratified sampling on the historical dataset to generate a target number of mutually exclusive data subsets; A group training unit is used to independently train a target number of diagnostic prediction models using the mutually exclusive data subsets and the key feature factors. The diagnostic prediction models are constructed based on the gradient boosting tree algorithm. The information prediction unit is used to predict the number of patients using the trained patient volume prediction model, obtain group prediction information for the target number of patients, and determine the patient volume prediction information of the medical service institution based on the fusion result of the group prediction information.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images