A disease-specific data management system integrating EDC (Electronic Data Center) and AI digital doctor.

Abstract

This invention discloses a disease-specific data management system integrating EDC (Electronic Data Collection) and AI digital doctor, relating to the field of medical informatics. It manages disease-specific data entered into an electronic data acquisition system, constructs a disease-specific manifestation space, and divides the space into low-density zones based on the density of cases within this space. Within the same time window, it separately calculates the disease subtype coverage rate in the low-density zones for case sets displaying AI digital doctor suggestions and those without, calculates the coverage rate difference, and determines deviations, thereby identifying the impact of AI suggestions on disease subtype coverage. This invention solves the problem that AI digital doctors may cause a gradual decrease in subtype cases and implicit deviations in data distribution during disease-specific management.

Description

Technical Field

[0001] This invention relates to the field of medical informatics, and more specifically, to a disease-specific data management system that integrates EDC (Electronic Data Center) and AI digital doctor. Background Technology

[0002] Electronic data capture (EDC) systems are widely used in clinical research and disease management. Disease-specific data management systems built around a single disease category or a specific disease management theme are gradually becoming an important component of healthcare informatization. Through the systematic collection and management of symptoms, signs, examination and test results, medication administration, and follow-up information, EDC systems provide fundamental data support for disease understanding, treatment decisions, and long-term management. In recent years, artificial intelligence technology, especially AI digital doctor models trained on large-scale medical data, has been introduced into disease management processes to provide reference suggestions to clinicians during case entry or diagnosis and treatment, thereby improving diagnostic and treatment efficiency and reducing human error.

[0003] However, the recommendation capabilities of AI digital doctors are typically built upon existing historical data distributions, and their model training process heavily relies on mainstream case types with sufficient sample size and stable characteristics. In disease management scenarios, diseases often have multiple subtypes or different evolutionary stages. These subtypes account for a low proportion of the total cases, exhibit significant differences in manifestations, and have prominent clinical significance. Due to the imbalance of training data, AI digital doctors tend to cover high-frequency, typical case patterns when generating recommendations, while paying insufficient attention to low-frequency subtypes, atypical manifestations, or transitional states.

[0004] In practical applications, once the suggestions of AI digital doctors are presented to clinicians, they may unintentionally guide the data collection and judgment process, making clinicians more likely to align with the mainstream patterns favored by the model when entering case information, selecting fields, or determining diagnostic pathways. This influence does not manifest as obvious errors, but rather changes the overall distribution structure of disease-specific data in a long-term, gradual manner, leading to a continuous decrease in low-frequency subtype cases in the dataset. Over time, the ability of disease-specific datasets to cover the true diversity of diseases may gradually decline, which in turn further reinforces the AI ​​digital doctor's preference for mainstream subtypes, forming a subtle data bias cycle.

[0005] Therefore, in the process of integrating EDC and AI digital doctors to manage disease-specific data, how to objectively assess the impact of AI suggestions on the distribution of disease-specific subtype data and promptly identify the potential deviation risks has become a problem that needs to be solved in the long-term healthy evolution of disease-specific data. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a disease-specific data management system that integrates EDC and AI digital doctor, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A disease-specific data management system integrating EDC (Electronic Data Center) and AI digital doctor, comprising: The disease-specific field management module is used to receive and store the disease-specific field data entered in the electronic data collection (EDC) of cases, and to form the disease-specific field data into a case record. The AI ​​digital doctor module is used to generate suggestions based on the case disease records and set the display status, including on and off status. When it is on / off, the suggestions are shown / hidden to the clinicians. The disease manifestation space module is used to construct a disease manifestation space based on the disease records of historical cases, and to divide low-density areas based on the density values ​​of historical cases in the disease manifestation space. The low-density areas are areas where the density values ​​are less than a first density threshold. The coverage difference detection module is used to calculate the disease subtype coverage rate for the case set in the open state and the case set in the closed state respectively within the same preset time window. The disease subtype coverage rate is the ratio of the number of cases falling into the low-density area to the total number of cases. The absolute value of the difference between the disease subtype coverage rate corresponding to the open state and the disease subtype coverage rate corresponding to the closed state is used as the coverage difference. The deviation determination module generates a deviation marker when the coverage difference is greater than a preset difference threshold.

[0008] Preferably, the system further includes: The system also includes a reinforcement learning correction module, used to perform parameter correction on the AI ​​digital doctor module after the deviation label is generated; The reinforcement learning correction module uses the negative value of the coverage difference as the reward value and the proportion of low-density areas in the disease-specific records in the disease-specific performance space and the deviation marker as the state input to update the model parameters of the AI ​​digital doctor module so that the coverage difference is reduced in the subsequent preset time window.

[0009] Preferably, the disease-specific field management module receives and stores disease-specific field entry data including symptom field values, physical sign field values, examination and test field values, medication and treatment field values, and follow-up field values. The disease-specific field management module sorts the disease-specific field entry data by time according to the field identifier and entry timestamp of each field to form the case-specific disease record.

[0010] Preferably, the disease manifestation space is a multi-dimensional vector space, which is composed of disease manifestation vectors corresponding to historical cases; wherein, each disease record of the case is mapped to a disease manifestation vector, and each dimension of the disease manifestation vector corresponds to a pre-determined disease field feature, which includes the following: symptom field value features, sign field value features, examination and test field value features, medication and treatment field value features, and follow-up field value features; The disease-specific manifestation space module maps the disease-specific manifestation vectors of all historical cases to the same multidimensional coordinate system, forming a disease-specific manifestation space used to characterize the differences in the distribution of disease-specific features among different cases.

[0011] Preferably, the specific disease manifestation spatial module calculates the density value in the following manner: In the disease-specific manifestation space, the disease-specific manifestation vector corresponding to any case is used as the center vector to determine the neighborhood range with a preset neighborhood radius. The number of disease-specific manifestation vectors corresponding to historical cases falling within the neighborhood is counted to obtain the number of neighborhood cases. The neighborhood volume corresponding to the neighborhood range is determined based on the neighborhood radius; The ratio of the number of neighboring cases to the volume of the neighboring area is used as the density value at the position corresponding to the center vector.

[0012] Preferably, the first density threshold is determined in the following manner: The density values ​​of historical cases in the disease manifestation space are statistically analyzed by the disease manifestation space module to obtain a set of historical case density values; The set of historical case density values ​​is sorted according to their numerical values, and target quantiles are determined based on a preset quantile ratio. The density value corresponding to the target quantile is determined as the first density threshold, wherein the target quantile ratio is used to limit the proportion of cases below the first density threshold in the total number of historical cases, so that the low-density area corresponds to the region where the sample distribution in historical cases is relatively sparse.

[0013] Preferably, the target quantile ratio ranges from 0.05 to 0.30.

[0014] Preferably, the AI ​​digital doctor module includes a large language model and a prompt word construction unit. The prompt word construction unit is used to convert the case disease record into structured prompt word text. The structured prompt word text includes disease field identifiers and corresponding field values. The large language model performs inference on the structured prompt word text to generate the suggestion and outputs the suggestion label and suggestion confidence level corresponding to the suggestion. The suggestion confidence level is the output probability of the large language model for the suggestion label.

[0015] Preferably, the case is a case that has been included in the preset disease management scope, and the disease field entry data received and stored by the disease field management module corresponds to the preset disease management scope; The suggestions generated by the AI ​​digital doctor module include performing subtype determination, stage determination, or status determination on cases within the preset disease management scope, and generating corresponding medication or treatment suggestions.

[0016] Preferably, the disease-specific management object is a disease-specific management object set for a single disease category or a single disease management theme. The disease-specific management object includes one or more of the following disease types: malignant tumor diseases, chronic metabolic diseases, cardiovascular and cerebrovascular diseases, autoimmune diseases, genetic diseases, infectious diseases, or rare diseases.

[0017] The advantage of this invention over existing technologies lies in its system architecture, which addresses the potential hidden bias issues caused by AI-powered digital doctors in disease-specific data management. This architecture enables quantitative monitoring and deviation identification of disease subtype coverage. By introducing a contrast mechanism between displayed and hidden AI suggestions in disease management scenarios, the system distinguishes between cases with displayed and hidden AI suggestions. Within the same time window, it calculates the coverage of each suggestion in low-density subtype areas, thus obtaining the coverage difference. This allows for the objective quantification of the impact of AI suggestions on disease-specific data distribution. This mechanism avoids reliance on subjective judgment or post-hoc statistical analysis, enabling continuous monitoring of disease-specific data bias during system operation.

[0018] By constructing a multidimensional vector space based on disease-specific fields and dividing low-density areas according to the distribution density of historical cases in this space, this invention can characterize the sparse regions of disease-specific subtypes at the overall data structure level. This allows subtype coverage to no longer be limited to the number of labels or simple proportions, but to reflect the true distribution in the disease-specific feature space. Combined with a mechanism for determining coverage differences and preset thresholds, deviation markers can be generated before deviations evolve into significant data distortion, thereby enabling early warning of the health status of disease-specific datasets.

[0019] Building upon this foundation, a reinforcement learning correction mechanism is introduced, using changes in coverage discrepancies as feedback signals to adaptively update the model parameters of the AI ​​digital doctor. This allows AI suggestions to gradually reduce their negative impact on disease subtype coverage during subsequent operation. This mechanism transforms the AI ​​digital doctor from a potential source of bias into a part of the self-regulating distribution of disease-specific data, helping to maintain the continuous coverage of low-frequency subtypes while ensuring clinical reference value. Through these technical means, this invention enhances the stability, objectivity, and disease diversity representation capabilities of the disease-specific data management system during long-term operation. Attached Figure Description

[0020] Figure 1 This is a diagram of the overall module architecture of the present invention; Figure 2 This is a schematic diagram illustrating the spatial distribution characteristics of the disease manifestations of the present invention, the division of low-density areas, and the calculation of local density. Figure 3 This is a schematic diagram illustrating the comparison of coverage difference and deviation determination logic in the on and off states of the present invention; Figure 4 This is a schematic diagram of the reinforcement learning correction module of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] For ease of understanding, the system described in this invention can be deployed in a hospital information system or a clinical research data platform. It can consist of a server cluster, a database, terminal devices, and a network communication module. Clinical personnel can complete the entry of case information and view suggestions through the electronic data collection (EDC) interface. The system continuously monitors the distribution of case data and the impact of AI digital doctor suggestions in the background, and triggers model parameter correction when necessary.

[0023] like Figure 1 As shown, the overall module architecture of this invention includes a disease-specific field management module, an AI digital doctor module, a disease-specific performance space module, a coverage difference detection module, and a deviation judgment module, and optionally includes a reinforcement learning correction module.

[0024] In one embodiment, the cases are those already included in the preset disease-specific management scope. The preset disease-specific management scope is used to define the boundaries of the objects managed by this system, avoiding the mixing of different disease management themes into the same statistical space and causing distribution distortion. The preset disease-specific management scope can be configured by medical institutions before the system goes live, and the configuration can be based on clinical pathways, research protocols, or disease-specific center management requirements. Diseases can be set for a single disease category or a single disease management theme, including malignant tumors, chronic metabolic diseases, cardiovascular and cerebrovascular diseases, autoimmune diseases, genetic diseases, infectious diseases, or rare diseases. The motivation for this setting is that disease-specific data has scarce and highly individualized characteristics. Only by first defining the management scope can the subsequent low-density areas and coverage differences have a consistent statistical caliber and medical interpretation basis.

[0025] In this embodiment, the disease-specific field management module receives and stores the disease-specific field data entered into the electronic data collection (EDC) for cases, and forms a case-specific disease record from this data. The disease-specific field data includes values ​​for symptoms, signs, examinations, medications, and follow-up. To ensure the data can be stably used by the model and statistical modules, the system assigns a field identifier to each field and writes an entry timestamp when data is entered or updated. The disease-specific field management module sorts the entered data by time based on the field identifier and entry timestamp to form the case-specific disease record. This design aims to preserve the chronological order of clinical information, facilitating the integration of follow-up and treatment processes into the same disease representation, and also providing a reproducible input structure for constructing prompts for the large language model.

[0026] Furthermore, the AI ​​digital doctor module generates suggestions based on case-specific disease records and sets a display status, including an on and off state. When on, the suggestions are displayed to clinicians; when off, they are hidden. The purpose of setting the display status is to create a control condition within the same system, allowing the system to compare the differences in disease subtype coverage under the two scenarios of displayed and hidden suggestions, thereby quantifying the guiding effect of suggestions on data distribution. The display status can be pre-assigned by the research protocol, randomly assigned by the system according to a preset ratio, or controlled by clinical role permissions. To minimize interference with clinical workflow, the system can determine the display status when a case is created and maintain it within a preset time window. The preset time window can be 7 days, 14 days, or 30 days, with the specific value determined based on the number of cases and follow-up frequency; a common range is 7 to 30 days.

[0027] In another embodiment, the AI ​​digital doctor module includes a large language model and a prompt word construction unit. The large language model can be an existing general-purpose large language model, fine-tuned for specific disease scenarios. Existing general-purpose large language models include ChatGPT. The prompt word construction unit converts case-specific disease records into structured prompt word text. The structured prompt word text includes disease field identifiers and corresponding field values, organized according to a fixed order of field identifiers, and includes an entry timestamp or follow-up timestamp to distinguish baseline and follow-up information. The large language model performs inference on the structured prompt word text to generate suggestions and outputs suggestion labels and suggestion confidence levels corresponding to the suggestions. The suggestion confidence level is the output probability of the large language model for the suggestion label. To enable the suggestions to be used by the coverage difference detection module, the system maps the suggestions to a preset suggestion label set. This suggestion label set is pre-configured by the disease management team and establishes a correspondence with the disease field set.

[0028] In the disease-specific manifestation space module, the system constructs a disease-specific manifestation space based on historical case records. This space is a multi-dimensional vector space, composed of disease-specific manifestation vectors corresponding to historical cases. Each case record is mapped to a disease-specific manifestation vector, with each dimension corresponding to pre-defined disease-specific field features. These features include symptom field values, physical sign field values, examination and test field values, medication and treatment field values, and follow-up field values.

[0029] To ensure that fields with different dimensions can perform distance and neighborhood operations within the same vector space, the system performs normalization on the disease-specific performance vectors. Normalization can be achieved using either min-max normalization or Z-score normalization: min-max normalization maps each dimension to 0 to 1, suitable for indicators with stable value ranges; Z-score normalization standardizes each dimension according to the mean and standard deviation, suitable for indicators with significant differences across institutions or batches.

[0030] For discrete values ​​such as symptom and sign fields, one-hot encoding or target encoding can be used; for medication treatment fields, they can be encoded as discrete dimensions according to drug category, dosage range, or treatment item; for follow-up fields, the stage status can be encoded as a discrete dimension, and a follow-up timestamp sequence can be added to form sequence features, and then a fixed-length vector can be generated through statistical summarization. By mapping the disease-specific manifestation vectors of all historical cases to the same multidimensional coordinate system, a disease-specific manifestation space is formed to characterize the differences in the distribution of disease-specific features among different cases. The motivation for this design is to elevate the differences in disease-specific subtypes from single-label differences to differences in feature distribution, so that low-frequency subtypes can be identified in the form of spatially sparse regions.

[0031] In the disease-specific manifestation space module, it is also necessary to divide the area into low-density regions based on density values. Low-density regions are areas where the density value is less than the first density threshold. For example... Figure 2 As shown, the density value is calculated as follows: In the disease manifestation space, the disease manifestation vector corresponding to any case is used as the center vector, and a neighborhood range with a preset neighborhood radius is determined; the number of disease manifestation vectors corresponding to historical cases falling within the neighborhood range is counted to obtain the number of neighborhood cases; the neighborhood volume corresponding to the neighborhood range is determined based on the neighborhood radius; the ratio of the number of neighborhood cases to the neighborhood volume is used as the density value at the position corresponding to the center vector. The neighborhood range here can be a Euclidean distance spherical neighborhood or a Mahalanobis distance ellipsoidal neighborhood.

[0032] The selection of the preset neighborhood radius directly affects the stability of density estimation. To avoid excessive density fluctuations due to an overly small neighborhood, or the smoothing out of low-density areas due to an overly large neighborhood, the system can adopt one of the following two approaches. The first approach uses the quantile value of the pairwise distance distribution of historical case-specific disease manifestation vectors as the preset neighborhood radius. The quantile value can be between 0.05 and 0.20, ensuring the neighborhood covers a local rather than global area. The second approach uses the number of cases in the target neighborhood as a constraint to inversely deduce the preset neighborhood radius, ensuring that the number of cases in the neighborhood of each central vector falls between 10 and 100, thus achieving sufficient statistical stability. Both approaches can be combined with normalized vector distances to reduce interference from differences in dimensions.

[0033] The determination of the first density threshold can be achieved as follows: The system first calculates the density value for each historical case. Specifically, in the disease manifestation space, the disease manifestation vector corresponding to the historical case is used as the center vector. A neighborhood range is determined with a preset neighborhood radius. The number of disease manifestation vectors corresponding to historical cases falling within this neighborhood range is counted to obtain the number of neighborhood cases. Then, the neighborhood volume is determined according to the preset neighborhood radius. Finally, the ratio of the number of neighborhood cases to the neighborhood volume is used as the density value of the historical case in the disease manifestation space. The system repeats the above density value calculation process for all historical cases to obtain a set of historical case density values ​​that correspond one-to-one with each historical case. Subsequently, the set of historical case density values ​​is sorted by numerical value, and a target quantile is determined with a preset quantile ratio. The density value corresponding to the target quantile is used as the first density threshold. The preset quantile ratio ranges from 0.05 to 0.30. The reason for using the quantile method is that the case size varies for different diseases and different institutions. Directly setting an absolute density threshold would result in the low-density area being divided too wide or too narrow. The quantile threshold, on the other hand, can keep the sparsity of the low-density area relatively stable when the case size changes, so that the low-density area can cover the low-frequency subtype area and ensure that the number of cases falling into the area meets the statistical requirements.

[0034] like Figure 3As shown, in the coverage difference detection module, the system calculates the disease subtype coverage rate for the case set in the open state and the case set in the closed state within the same preset time window. The disease subtype coverage rate is the ratio of the number of cases falling into the low-density area to the total number of cases. Then, the absolute value of the difference between the disease subtype coverage rate corresponding to the open state and the disease subtype coverage rate corresponding to the closed state is calculated as the coverage difference.

[0035] Furthermore, the deviation detection module generates a deviation flag when the coverage difference exceeds a preset difference threshold. The preset difference threshold distinguishes between acceptable random fluctuations and systemic deviations requiring attention; its value can be determined based on the distribution of coverage differences over historical windows. One feasible approach is to collect coverage differences from multiple preset time windows during the initial system deployment phase, taking the mean plus 1.5 times the standard deviation as the preset difference threshold. Another feasible approach is to directly set the preset difference threshold to a fixed value between 0.02 and 0.10 and differentiate it according to disease type. The purpose of generating the deviation flag is not to determine clinical entry errors, but rather to suggest and indicate a potential statistically significant risk of deviation in disease subtype coverage, providing triggering conditions for subsequent corrections.

[0036] In embodiments that include a reinforcement learning correction module, such as Figure 4 As shown, the reinforcement learning correction module is used to correct the parameters of the AI ​​digital doctor module after generating deviation labels. The module uses the negative value of the coverage difference as the reward value and the proportion of low-density areas in the disease-specific case record space and the deviation label as state inputs to update the model parameters of the AI ​​digital doctor module, thereby reducing the coverage difference within a preset time window. The proportion of low-density areas can be the proportion of low-density areas in the case set in the open state, or the difference between the low-density areas in the open and closed states. The system can choose one according to implementation needs, but it must remain consistent within the same deployment version. Using the negative value of the coverage difference as the reward value means that as the coverage difference between suggested display and hidden decreases, the reward value increases, thereby driving the strategy to converge in the direction of reducing deviation.

[0037] Reinforcement learning can be implemented using policy gradient or proximal policy optimization methods. The input to the policy network is the state input, and the output is used to adjust the parameter increments of the AI ​​digital doctor module model. To maintain suggestion stability in clinical scenarios, model parameter updates can use small step sizes, with a learning rate of 1e-5 to 1e-3, and the update cycle consistent with the preset time window. The model parameters can be selected from some trainable layers of a large language model, or from the prompt word weight parameters or label mapping layer parameters in the prompt word construction unit. If using some trainable layers of a large language model, the system can freeze most layers and only update the last few layers or adapter layers to reduce the risk of catastrophic forgetting. If using prompt word weight parameters, the system can adjust the presentation weights of different disease-specific fields in the structured prompt word text, allowing the model to focus more on field combinations related to low-density areas when generating suggestions, thereby improving coverage of low-frequency subtypes.

[0038] In another embodiment, to prevent a sudden drop in suggestion quality caused by reinforcement learning updates, the system can introduce a safety constraint mechanism. After each preset time window, the reinforcement learning correction module first evaluates the changing trend of coverage difference on offline playback data and sets an upper limit for the allowed parameter change range. The upper limit can be between 0.01 and 0.10 of the original parameter norm. If the upper limit is exceeded, the parameter increment is pruned. The system can also set entropy constraints on the output distribution of suggestion labels to prevent the model from outputting overly scattered suggestions due to excessive pursuit of reducing coverage difference, thereby maintaining a balance between clinical usability and data diversity.

[0039] Through the above implementation methods, this invention, without relying on manual annotation of disease subtypes, uses the density structure of the disease manifestation space to characterize low-frequency subtype regions, and quantifies the impact of suggestions on disease subtype coverage using a comparative statistical method of suggestion display and hiding. When the deviation exceeds a threshold, a deviation marker is generated, and the AI ​​digital doctor module can be adaptively corrected through a reinforcement learning correction module, thereby reducing the coverage difference in subsequent time windows and suppressing the risk of disease data distribution excessively converging to high-frequency subtypes in long-term evolution.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A disease-specific data management system integrating EDC (Electronic Data Center) and AI digital doctor, characterized in that, include: The disease-specific field management module is used to receive and store the disease-specific field data entered in the electronic data collection (EDC) of cases, and to form the disease-specific field data into a case record. The AI ​​digital doctor module is used to generate suggestions based on the case disease records and set the display status, including on and off status. When it is on / off, the suggestions are shown / hidden to the clinicians. The disease manifestation space module is used to construct a disease manifestation space based on the disease records of historical cases, and to divide low-density areas based on the density values ​​of historical cases in the disease manifestation space. The low-density areas are areas where the density values ​​are less than a first density threshold. The coverage difference detection module is used to calculate the disease subtype coverage rate for the case set in the open state and the case set in the closed state respectively within the same preset time window. The disease subtype coverage rate is the ratio of the number of cases falling into the low-density area to the total number of cases. The absolute value of the difference between the disease subtype coverage rate corresponding to the open state and the disease subtype coverage rate corresponding to the closed state is used as the coverage difference. The deviation determination module generates a deviation marker when the coverage difference is greater than a preset difference threshold.

2. The system according to claim 1, characterized in that, The system also includes a reinforcement learning correction module, used to perform parameter correction on the AI ​​digital doctor module after the deviation label is generated; The reinforcement learning correction module uses the negative value of the coverage difference as the reward value and the proportion of low-density areas in the disease-specific records in the disease-specific performance space and the deviation marker as the state input to update the model parameters of the AI ​​digital doctor module so that the coverage difference is reduced in the subsequent preset time window.

3. The system according to claim 1, characterized in that, The disease-specific field management module receives and stores disease-specific field entry data including symptom field values, physical sign field values, examination and test field values, medication and treatment field values, and follow-up field values. The disease-specific field management module sorts the disease-specific field entry data by time according to the field identifier and entry timestamp of each field to form the case-specific disease record.

4. The system according to claim 1 or 3, characterized in that, The disease-specific manifestation space is a multi-dimensional vector space, which is composed of disease-specific manifestation vectors corresponding to historical cases. Each disease-specific record of a case is mapped to a disease-specific manifestation vector, and each dimension of the disease-specific manifestation vector corresponds to a pre-determined disease-specific field feature. The disease-specific field features include the following: symptom field value features, physical sign field value features, examination and test field value features, medication and treatment field value features, and follow-up field value features. The disease-specific manifestation space module maps the disease-specific manifestation vectors of all historical cases to the same multidimensional coordinate system, forming a disease-specific manifestation space used to characterize the differences in the distribution of disease-specific features among different cases.

5. The system according to claim 1 or 3, characterized in that, The specific disease manifestation spatial module calculates the density value in the following manner: In the disease-specific manifestation space, the disease-specific manifestation vector corresponding to any case is used as the center vector to determine the neighborhood range with a preset neighborhood radius. The number of disease-specific manifestation vectors corresponding to historical cases falling within the neighborhood is counted to obtain the number of neighborhood cases. The neighborhood volume corresponding to the neighborhood range is determined based on the neighborhood radius; The ratio of the number of neighboring cases to the volume of the neighboring area is used as the density value at the position corresponding to the center vector.

6. The system according to claim 5, characterized in that, The first density threshold is determined in the following manner: The density values ​​of historical cases in the disease manifestation space are statistically analyzed by the disease manifestation space module to obtain a set of historical case density values; The set of historical case density values ​​is sorted according to their numerical values, and target quantiles are determined based on a preset quantile ratio. The density value corresponding to the target quantile is determined as the first density threshold, wherein the target quantile ratio is used to limit the proportion of cases below the first density threshold in the total number of historical cases, so that the low-density area corresponds to the region where the sample distribution in historical cases is relatively sparse.

7. The system according to claim 6, characterized in that, The target quantile ratio ranges from 0.05 to 0.

30.

8. The system according to claim 1, characterized in that, The AI ​​digital doctor module includes a large language model and a prompt word construction unit. The prompt word construction unit is used to convert the case disease record into structured prompt word text. The structured prompt word text includes disease field identifiers and corresponding field values. The large language model performs inference on the structured prompt word text to generate the suggestion and outputs the suggestion label and suggestion confidence level corresponding to the suggestion. The suggestion confidence level is the output probability of the suggestion label by the large language model.

9. The system according to claim 1, characterized in that, The cases are those already included in the preset disease management scope, and the disease field entry data received and stored by the disease field management module corresponds to the preset disease management scope; The suggestions generated by the AI ​​digital doctor module include performing subtype determination, stage determination, or status determination on cases within the preset disease management scope, and generating corresponding medication or treatment suggestions.

10. The system according to claim 1 or 9, characterized in that, The term "specialized disease" refers to a disease management object set up for a single disease category or a single disease management theme. The specialized disease includes one or more of the following disease types: malignant tumor diseases, chronic metabolic diseases, cardiovascular and cerebrovascular diseases, autoimmune diseases, hereditary diseases, infectious diseases, or rare diseases.

Images