Intelligent decision method and system for human resource planning of an input person

Abstract

The present application belongs to the technical field of intelligent decision-making method of human resource planning, in particular to an intelligent decision-making method and system for human resource planning of recruited personnel, which deploys a cross-system data access gateway adapted to the compliance requirements of the medical internal network, combines a standardized terminology dictionary and semantic intelligent matching technology, establishes a dynamic mapping relationship of synonymous fields in multiple systems, realizes unified data grabbing and standardized feature conversion of data scattered in different medical business systems, and through a feature version management mechanism, adapts to the dynamic changes of system field definitions, achieves full-dimensional fusion and consistency modeling of multi-source heterogeneous data; based on medical business hard rules and intelligent algorithms, a full-link data governance system is constructed, through real-time verification and scenario-based intelligent repair, the data integrity and accuracy are improved, with the help of unsupervised anomaly detection, automatic labeling of sudden scene features is realized, and by using causal inference technology, the influence of historical human resource allocation bias is stripped and virtual training samples are supplemented.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent decision-making methods for human resource planning, and particularly relates to an intelligent decision-making method and system for human resource planning involving personnel input. Background Technology

[0002] Currently, medical record entry data is scattered across multiple business systems such as HIS, LIS, and PACS. The field definitions, data formats, and terminology standards of these systems are inconsistent, leading to difficulties in integrating multi-source heterogeneous data, reliance on manual matching of synonymous fields, and poor adaptability to dynamic changes in system fields. At the same time, the medical intranet has stringent compliance requirements, and compatibility between outdated equipment and proprietary protocols is difficult. The standardization, security, and long-term consistency of data collection are hard to guarantee, making it impossible to provide stable and unified data support for human resource planning and modeling for data entry personnel.

[0003] Current human resource planning for medical data entry personnel largely relies on human experience. Time series forecasting using a single model is difficult to adapt to the fluctuation patterns of different medical scenarios such as routine, periodic, and emergency situations. Furthermore, historical data suffers from deficiencies such as biases in human resource allocation, missing data, and reliance on manual anomaly labeling. Data governance and sample debiasing optimization capabilities are also insufficient. In addition, medical sensitive data has high privacy protection requirements. Traditional centralized modeling is prone to information leakage, and simply using privacy protection technologies will reduce model performance. The lack of a dynamic balance mechanism between privacy security and model performance makes it difficult to achieve compliant, accurate, and adaptive intelligent human resource planning decisions. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an intelligent decision-making method for human resource planning of personnel, the method comprising: By deploying a cross-system data access gateway that is adapted to the compliance requirements of the medical intranet, and combining a standardized terminology dictionary and semantic intelligent matching technology, a dynamic mapping relationship of synonymous fields in multiple systems is established to achieve unified data capture and standardized feature conversion of data scattered in different medical business systems. Furthermore, through a feature version management mechanism, dynamic changes in system field definitions are adapted to achieve full-dimensional fusion and consistent modeling of multi-source heterogeneous data. A full-chain data governance system is built based on hard rules for medical business and intelligent algorithms. Data integrity and accuracy are improved through real-time verification and scenario-based intelligent repair. Unsupervised anomaly detection is used to automatically label the characteristics of sudden scenarios. Causal inference technology is used to remove the influence of historical human resource allocation deviations and supplement virtual training samples. Based on the inherent laws of medical scenarios and external influencing factors, the original time series data is divided into scenario layers. Feature enhancement is achieved by introducing medical features and dynamic feature filtering. A hybrid modeling framework with multiple models adapted to different scenarios and a dynamically adjusted sliding window mechanism are adopted. We construct a federated learning architecture for local data storage and encrypted parameter interaction. We protect sensitive features through differential privacy technology adapted to medical scenarios, and use feature distillation method to indirectly infer information from non-sensitive features to compensate for the information loss of desensitized data. We also combine dynamic privacy and performance balance adjustment mechanisms.

[0005] Furthermore, embodiments of the present invention also provide an intelligent decision-making system for human resource planning of personnel, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the aforementioned intelligent decision-making method for human resource planning of personnel entry via executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the aforementioned intelligent decision-making method for human resource planning of data entry personnel.

[0007] Based on the above, through cross-system data access gateways, intelligent medical semantic matching, and feature version management, secure and compliant collection, standardized conversion, and long-term consistency assurance of multi-source heterogeneous medical data are achieved. Relying on the full-link data governance system, real-time data verification, scenario-based intelligent repair, automatic annotation of emergency scenarios, and historical sample de-biasing optimization are completed, improving data quality from the source and eliminating modeling errors caused by human resource allocation bias. Based on scenario layering and splitting, hybrid modeling, and dynamic sliding window mechanisms, the prediction model can accurately adapt to the temporal fluctuation patterns of different medical scenarios such as routine, periodic, and emergency scenarios, greatly improving the accuracy and real-time performance of data entry prediction and human resource demand calculation.

[0008] This invention employs a federated learning architecture combined with scenario-adaptive differential privacy and medical feature distillation technology to achieve local transmission of raw medical data without leaving the domain and encrypted parameter interaction. While strictly meeting medical privacy compliance requirements, it effectively compensates for the information loss of anonymized data. Through a dynamic privacy-performance balance adjustment mechanism, it achieves adaptive synergy between privacy protection strength and model training performance, avoiding performance degradation caused by a single technical means. The overall solution enables intelligent decision-making throughout the entire process of human resource planning for data entry personnel, reducing reliance on manual experience and operating costs. It can flexibly adapt to the dynamic updates of medical business scenarios and compliance policies, significantly improving the scientific nature and scheduling efficiency of hospital human resource allocation. Attached Figure Description

[0009] Figure 1This is a schematic diagram of the execution flow of the intelligent decision-making method for human resource planning of personnel provided in the embodiments of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of the intelligent decision-making system for human resource planning of personnel provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an intelligent decision-making method for human resource planning of personnel input, provided in one embodiment of the present invention. The intelligent decision-making method for human resource planning of personnel input will be described in detail below.

[0012] Step S110: By deploying a cross-system data access gateway that adapts to the compliance requirements of the medical intranet, and combining a standardized terminology dictionary and semantic intelligent matching technology, a dynamic mapping relationship of synonymous fields in multiple systems is established to achieve unified capture and standardized feature conversion of data scattered in different medical business systems. Furthermore, the feature version management mechanism is used to adapt to the dynamic changes in system field definitions, thereby achieving full-dimensional fusion and consistent modeling of multi-source heterogeneous data. A cross-system data access gateway adapted to the Level 3 compliance requirements of the medical intranet was deployed, using the Spring Cloud Gateway architecture and locally deployed on the hospital's intranet server cluster. It supports medical standard protocols such as HL7, DICOM, and FHIR, as well as hospital-defined private protocols. A standardized terminology dictionary was built by integrating SNOMED CT, ICD-11, and the National Medical Service Operation Classification, achieving semantic intelligent matching based on a fine-tuned medical BERT model. The gateway acquires data from systems such as HIS, LIS, and PACS through a combination of scheduled capture and real-time monitoring, completing field matching using a dynamic mapping graph of synonymous fields and converting it into standardized data according to a unified feature paradigm (e.g., uniformly mapping "medical record number" and "medical record ID" from different systems to "unique visit identifier"). A feature version management library was built, adopting a Git-like version control mechanism. When system field definitions change, shadow field creation, historical data recalculation, and version marking are automatically triggered. Change trajectories are recorded through audit logs to ensure that multi-source heterogeneous data maintains consistency even under field change scenarios, providing a unified and standardized dataset for subsequent modeling.

[0013] Step S111: Design the gateway architecture according to compliance requirements, adopt a localized cluster deployment compatible with multiple technology stacks, integrate private protocol parsing and signal conversion capabilities of legacy devices, and carry out secure and compliant collection of medical multi-system data through encrypted transmission, fine-grained access control and offline caching mechanisms. The gateway architecture is designed to meet medical data security compliance requirements, employing a Java+Go hybrid technology stack to build a localized cluster (3 core nodes + 2 backup nodes) supporting horizontal scaling. A private protocol parsing module is integrated, and the custom TCP protocol of the legacy HIS system is adapted through reverse engineering. An RS485-to-TCP / IP signal conversion module is configured to achieve compatible data acquisition from legacy devices. Data transmission uses TLS 1.3 encryption, and fine-grained access control is implemented based on the RBAC model, assigning an independent access account to each access system and limiting the data access scope. A Redis cluster is deployed as an offline cache; when the system disconnects, the collected data is cached (cache validity period is 24 hours, automatically backed up to an encrypted disk after timeout), and batch synchronization occurs upon reconnection. The Quartz framework is used to schedule acquisition tasks, dynamically adjusting the polling frequency according to system load (1 minute / time for core business systems, 30 minutes / time for non-core systems) to ensure secure, efficient, and compliant data acquisition.

[0014] Step S112: Integrate medical dictionaries to build a localized and standardized terminology system. Collect metadata from multiple systems through gateway plugins and form a unified dictionary. Perform semantic completion on missing descriptive fields to solidify the terminology and metadata foundation for unified data integration. A localized, standardized terminology system was built by integrating SNOMED CT terminology sets, drug coding standards, and medical insurance item classification codes. Deployed on a MySQL master-slave cluster, it supports an average of 100,000 query requests per day. A gateway metadata collection plugin (Java interceptor mode) was developed to capture metadata such as field names, data types, lengths, annotations, and business uses from various medical business systems via HTTP / HTTPS interfaces, generating a unified metadata dictionary, stored in a three-level classification by system, department, and business module. For fields with missing descriptions, semantic completion is performed based on a medical knowledge graph: by associating adjacent fields within the same system and matching keywords from the terminology dictionary, standardized annotations are generated (e.g., for the unannotated field "XY001," combined with the department "cardiology" and the associated field "blood pressure value," it is completed as "systolic blood pressure measurement result"). The metadata dictionary is synchronized daily with the latest terminology standards from the National Health Commission, supporting fuzzy search and batch export functions, providing robust terminology and metadata support for the unified integration of data from multiple systems.

[0015] Step S113: Based on the fine-tuned pre-trained model in the medical field, construct semantic intelligent matching capability, establish and store dynamic mapping relationships of synonym fields in the form of a graph, and achieve adaptive updating of the mapping relationship through a combination of timed scanning and manual review. A semantic matching model is built by fine-tuning a pre-trained medical model (BERT-based-chinese), deployed on a GPU server (NVIDIA A100), and provides a RESTful API interface (response time ≤100ms). A medical field mapping knowledge graph is constructed using the Neo4j graph database, with system fields as nodes (including field ID, system identifier, business description, etc.) and synonym / association / hierarchical relationships as edges (including confidence, matching time, scene label, etc.). An XXL-Job distributed scheduling platform is configured to execute metadata scanning tasks at a fixed 24-hour cycle, traversing all metadata interfaces of the connected systems, monitoring for field additions, name modifications, attribute changes, etc., automatically feeding the changed data into the semantic matching model for incremental calculation, and synchronously updating the graph mapping relationships. A web-based manual review interface (Vue3 + Element Plus) is developed, displaying the field source, similarity score, and business context of suspected synonymous fields and low-confidence mappings (confidence 0.6-0.85), supporting confirmation, correction, and rejection operations, with review results written back to the graph in real time. Each quarter, manually corrected data is used as training samples to feed back into the model for incremental fine-tuning. At the same time, change logs are stored in MongoDB, supporting query and backtracking by time, system, and field dimensions, achieving adaptive management of the entire lifecycle of mapping relationships.

[0016] Step S1131: Fine-tune the general pre-trained model based on the text features of the medical field, construct a medical scenario semantic matching model, replace the model input with field names, field annotations, and business purpose text data of the medical business system, remove irrelevant general corpus, complete the model adaptation training by using a few-shot learning method, and optimize the model's semantic understanding ability of medical input scenario terms. BERT-base-chinese was selected as the general pre-trained model, and fine-tuned based on multi-system metadata (including 100,000 field names, annotations, and business-purpose text) from 10 top-tier hospitals. The training data was divided into training and validation sets in an 8:2 ratio, and common stop words such as "field" and "information" were removed. The jieba medical word segmentation extension package was used for text segmentation. Medical domain prompt templates (e.g., "Are {field A} and {field B} synonymous fields in {business scenario}?") were introduced. A few-shot learning framework (Few-Shot BERT) was used to optimize the model, updating only the top-level classification layer parameters while freezing the bottom-level feature extraction layer. The optimizer was AdamW, with an initial learning rate of 0.0001 and 300 iterations. An early stopping mechanism was implemented (training stopped if the validation set accuracy did not improve for 20 consecutive iterations). The fine-tuned model was deployed as a containerized service, achieving elastic scaling through Kubernetes, and supporting batch field matching requests (up to 1000 field pairs processed per request).

[0017] Step S1132: Input the metadata fields of each medical business system into the fine-tuned semantic matching model to generate a high-dimensional semantic feature vector. Calculate the semantic correlation between fields of different systems using the cosine similarity algorithm, preset the grading judgment threshold, and automatically filter out high-confidence synonymous fields, suspected synonymous fields, and unrelated fields to complete the initial intelligent matching. Metadata fields (including name, annotations, and business purpose) from various medical business systems are input into a finely tuned semantic matching model to generate a 768-dimensional high-dimensional semantic feature vector. CUDA 11.4 is used to accelerate cosine similarity calculation, with three preset judgment thresholds: high-confidence synonymous fields (similarity ≥ 0.85), suspected synonymous fields (0.6 ≤ similarity < 0.85), and unrelated fields (similarity < 0.6). Before calculation, the field text is preprocessed: polysemous words are resolved using a medical knowledge graph (e.g., "blood pressure" refers to "clinic blood pressure" in the outpatient system and "dynamic blood pressure" in the monitoring system), and contextual information about the business scenario is supplemented. Preliminary matching results are stored in a temporary MySQL table, labeled with matching type, similarity score, and business scenario tag. High-confidence results are directly used for mapping graph construction, suspected fields are pushed to a manual review queue, and unrelated fields are temporarily stored in a waiting-to-match pool to reserve data for subsequent graph updates, ensuring the accuracy and comprehensiveness of synonym field selection.

[0018] Step S1133: Design a medical field mapping knowledge graph architecture, with each system field as an independent node and synonym, association, and hierarchical relationships as connecting edges. Label each mapping relationship with matching confidence, source system, matching time, and applicable business scenario attributes. Use a graph database to complete the structured storage and association query of the knowledge graph. The architecture for a medical field mapping knowledge graph is designed, with nodes divided into system nodes and field nodes. Field node attributes include field ID, system identifier, field name, data type, business description, and creation time. System node attributes include system name, version number, and deployment department. Edge types are categorized into synonym relationships, association relationships, and hierarchical relationships. Edge attributes include matching confidence, matching time, applicable business scenario (outpatient / inpatient / emergency), and update frequency. A distributed graph database cluster (3 core nodes + 2 read-only nodes) is built using Neo4j Enterprise Edition. Unique indexes are created for field ID and system identifier, and composite indexes are created for matching time and confidence, optimizing the performance of association queries (query response time ≤ 500ms). Cypher statements are supported for multi-dimensional queries (e.g., "query synonymous field pairs between HIS and LIS systems" and "filter inpatient scenario mapping relationships with confidence ≥ 0.9"). The graph data employs a "daily full backup + hourly incremental backup" strategy, with backup data stored on an AES-256 encrypted disk array, complying with medical data backup and disaster recovery compliance requirements to ensure the security and availability of the mapping relationship data.

[0019] Step S1134: Configure a scheduled scanning task to traverse all connected medical business system metadata interfaces at a fixed period, monitor three types of data changes in real time: new fields, field name modifications, and business attribute changes, and automatically input the changed metadata into the semantic matching model to trigger incremental matching calculation and synchronously update the mapping relationship in the knowledge graph. Configure the XXL-Job distributed scheduling platform to create scheduled scanning tasks. The scanning cycle can be flexibly adjusted through the configuration center (default 24 hours / time). During task execution, the metadata version number of each access system is first obtained via HTTP interface and compared with the local cached version. Only data changes in three categories—new fields, modified field names, and business attribute changes—are collected to reduce unnecessary calculations. The changed metadata is categorized by field type (text / numeric / date) and batch-input into the semantic matching model for incremental matching calculations. Only the mapping relationships related to the changed fields are updated, without affecting existing stable mappings in the graph. The Neo4j transaction mechanism is used to execute graph update operations. If an anomaly occurs during the update process (such as network interruption or data conflict), a transaction rollback is automatically triggered, and the administrator is notified via SMS and email. The data synchronization delay is ≤1 hour to ensure that the mapping graph can adapt to system field changes in a timely manner and maintain the timeliness of multi-system data mapping.

[0020] Step S1135: For suspected synonymous fields and low-confidence mapping relationships generated by automatic matching, trigger the manual review process, display the field source, semantic similarity, and business context information on the management terminal, and provide three operation options: confirm, correct, and reject. Write the review results back to the knowledge graph in real time. A web-based manual review management platform was developed, with an interactive interface built on Vue3 and Element Plus. It comprises three main modules: a field matching list, a details pop-up, and an operation panel. The field matching list displays the source system, field name, semantic similarity, and matching status of the field pairs to be reviewed, supporting filtering by similarity, system type, and business scenario. The details pop-up displays field annotations, screenshots of the business context (such as screenshots of the system field configuration page), a list of candidate synonym fields, and the distribution of similarity scores. The operation panel provides three functions: "Confirm," "Correct," and "Reject." Confirmation marks the mapping relationship as "valid" and saves it to the knowledge graph; correction allows manual input of the target matching field, recalculates the similarity, and updates it; rejection requires filling in the rejection reason (such as "inconsistent business scenario" or "difference in field meaning"), marks it as "invalid," and saves it to the history database. Review operations support single-item processing and batch processing (maximum 100 items per batch). Review results are written to the knowledge graph in real time, and the operator, operation time, and operation content are recorded in the audit log, meeting the traceability requirements of medical data operations.

[0021] Step S1136: Establish a model iteration mechanism, use the manually reviewed and corrected mapping data as new training samples, periodically feed them back to the semantic matching model, complete small incremental fine-tuning, and continuously improve the accuracy and confidence of the model's automatic matching. A model iteration and optimization mechanism was established. Monthly, manually reviewed and corrected mapping data (including positive samples: confirmed synonym pairs; negative samples: rejected matching pairs) were collected, with a sample size of no less than 1000 records. The data was organized into a training dataset in the format of "Field A Text - Field B Text - Label". An incremental fine-tuning strategy was adopted, freezing 80% of the model's underlying parameters and training only the top-level classification and semantic adaptation layers. The AdamW optimizer was used, with a learning rate set to 1 / 10 (0.00001) of the initial fine-tuning value, and 100 iterations were performed. An A / B testing mechanism was introduced, deploying the fine-tuned model and the online model in parallel. 10% of the data collection tasks were selected as test traffic to compare the matching accuracy and recall of the two models. When the accuracy of the new model improved by ≥5% compared to the old model, the online model was replaced through blue-green deployment, while the old model version was retained (for 3 months), supporting rollback. Quarterly, training data was supplemented and full fine-tuning was performed based on newly added medical terminology and changes in business scenarios to continuously improve the model's semantic matching ability for new fields and special scenarios.

[0022] Step S1137: Construct a mapping relationship change log and backtracking mechanism to fully record the operation information and mapping content of each automatic update and manual correction. Support querying historical mapping versions by time, system, and field dimensions to achieve traceability and rollback of mapping relationships and complete full lifecycle adaptive update management.

[0023] A mapping relationship change log and backtracking system is constructed, using a MongoDB sharded cluster to store log data. Log fields include version number, operation type (automatic update / manual correction), operator, operation time, mapping relationship before change, mapping relationship after change, affected data range, and remarks. Version numbers are generated according to the "year-month-day-serial number" rule (e.g., 20240520-001), supporting multi-dimensional retrieval by time range, system name, field ID, and version number. The backtracking function is triggered by the version number: after the user selects the target version, the system automatically compares the current graph with historical versions, generates a change list, and, after administrator confirmation, executes a graph data rollback operation (locking relevant nodes and edges during the rollback process and prohibiting write operations). Rollback takes ≤5 minutes, and an SMS notification is sent upon completion, while the rollback log is recorded. Change logs are retained for 3 years, complying with medical data retention regulations, ensuring the entire lifecycle of mapping relationships is traceable and rollback-capable, and providing assurance for data fusion consistency.

[0024] Step S114: Design a multi-mode data capture strategy that combines real-time push and timed polling for different types of medical business systems. After extracting structured / semi-structured data, complete field matching, indicator calculation and data format standardization conversion according to a unified feature paradigm. For core systems such as HIS and LIS, WebSocket long connections are used to achieve real-time push (data latency ≤ 1 second); for legacy PACS and financial systems, timed polling is used (polling cycle configurable from 1 to 60 minutes). A multi-format data parsing module was developed, using regular expressions to parse text-based semi-structured data, XPath to parse XML data, JSONPath to parse JSON data, and POI to parse Excel spreadsheet data. After data extraction, field matching is performed based on a synonym field mapping graph, and standardized transformation is carried out according to a unified feature paradigm: numeric fields are uniformly retained to two decimal places, date fields are converted to the "yyyy-MM-dd HH:mm:ss" format, and categorical fields are uniformly encoded according to a terminology dictionary (e.g., the "gender" field maps "male / female" to "1 / 0"). Core business indicators (such as average daily data entry volume, error rate, and field completeness rate) are calculated. The transformed data is stored in the InfluxDB time-series database, partitioned by "system-date-indicator type," supporting fast queries by time and system dimensions, providing standardized input for subsequent data governance and modeling.

[0025] Step S115: Build a feature version library based on the configuration center. Through field change monitoring, shadow field compatibility, historical data recalculation and audit log recording, establish a feature version management mechanism that adapts to dynamic field changes and ensures long-term consistency of multi-source data.

[0026] A feature version repository is built based on the Nacos configuration center to store feature version rules, field change records, and version-related data. A field change monitoring module is deployed to monitor field definition changes in various systems in real time via database triggers. When a field name change, type change, or business meaning adjustment is detected, a shadow field is automatically created (stored in parallel with the original field and marked with a "historical version" identifier). The Spark distributed computing framework is used to recalculate historical data, batch processing the historical data according to the new field rules (the recalculation task is executed in the background, without affecting front-end queries). After recalculation, the version marker is updated, allowing users to choose to use either the historical or the latest version of the data. An audit log system is built to record information such as version change time, operator, change content, and affected data range, providing a version comparison function (visualizing the differences in field changes). The feature version repository supports version switching (switching latency ≤ 10 seconds), and a caching preheating mechanism ensures stable query performance after switching, effectively adapting to scenarios with dynamic changes in system fields and ensuring long-term consistency of multi-source data.

[0027] Step S120: Construct a full-link data governance system based on medical business hard rules and intelligent algorithms, improve data integrity and accuracy through real-time verification and scenario-based intelligent repair, achieve automatic labeling of sudden scene features with the help of unsupervised anomaly detection, and use causal inference technology to remove the influence of historical human resource allocation deviations and supplement virtual training samples. A comprehensive governance system for medical data entry is constructed, integrating hard business rules with machine learning algorithms to form a closed-loop governance architecture. At the data acquisition layer, a real-time verification module is deployed to identify and correct anomalies based on medical business standards. At the data repair layer, generative adversarial networks combined with scenario-based prior knowledge are used to intelligently fill in missing data. At the feature labeling layer, unsupervised anomaly detection algorithms automatically identify and label sudden scenarios such as peak patient visits and medical insurance settlements. At the sample optimization layer, propensity score matching and counterfactual reasoning are used to eliminate data bias caused by insufficient or redundant historical manpower, while generating virtual training samples that conform to business logic. This entire system spans the entire process of data acquisition, cleaning, repair, labeling, and sample optimization, ensuring data integrity and accuracy while eliminating historical data bias. This provides high-quality, unbiased training datasets for subsequent time-series modeling, improving the compliance and accuracy of model predictions.

[0028] Step S121: Organize medical business-related rules and convert them into executable logic, build a rule engine that supports flexible configuration by department and business type, embed a real-time verification module in the data collection link, trigger rule verification synchronously when data is accessed, classify and mark abnormal data and remind supplementary data, and complete the real-time correction of problems that can be automatically corrected by associating with professional dictionaries and historical reasonable data ranges. We streamlined medical business rules regarding timeliness, data format, logical consistency, and medical insurance compliance for medical record entry, transforming them into executable logic such as conditional judgments, numerical ranges, and format validation. We then built a real-time validation framework based on a rule engine. The rule engine supports flexible configuration of rule items by department, business type, and entry scenario, adapting to the differentiated requirements of different departments. A real-time validation module is embedded between the data acquisition gateway and the business system. Rule matching is triggered synchronously upon data access, classifying and marking entry delays and missing fields, and reminding entry personnel to complete them via system pop-ups and push notifications. For correctable issues such as ICD encoding errors and numerical values ​​outside the range, the system automatically correlates with a medical terminology dictionary and historical reasonable data ranges for immediate correction, preventing abnormal data from flowing into subsequent processes and achieving compliance control in the data acquisition stage.

[0029] Step S122: Based on historical valid data, statistically analyze the data distribution characteristics of different diseases, departments, and data entry stages. Combine clinical pathways and treatment guidelines to construct a medical scenario prior library. Use generative adversarial networks to build an intelligent repair model. Train the model with the scenario prior library as a constraint condition so that the generator learns the real distribution pattern of data in different scenarios. For missing data, identify the medical scenario it belongs to through the data context. The model generates fill values ​​that conform to the distribution characteristics of the scenario. After filling, perform secondary verification through the medical business rules. A medical scenario prior library is constructed based on historical quality control qualified data, and a Generative Adversarial Network (GAN) is used to achieve scenario-based intelligent repair of missing data. By statistically analyzing the data distribution characteristics of different diseases, departments, and data entry stages, scenario constraint rules are formed in conjunction with clinical pathways and treatment guidelines. The GAN model uses the scenario prior library as training constraints, enabling the generator to learn the real distribution of data in each scenario and avoiding business distortion caused by general imputation methods. For missing data, the medical scenario to which it belongs is first identified through contextual information, and then the model generates imputation values ​​that conform to the distribution of that scenario. After imputation, a business rule engine is called for secondary verification to ensure that the repaired data meets the format, logic, and compliance requirements. This approach significantly improves the authenticity and business adaptability of missing data repair, outperforming traditional imputation methods such as mean and median.

[0030] Step S1221: Divide medical scenarios into four dimensions: disease type, department of visit, data entry process, and treatment level. Collect historical data without abnormalities from the past three years that have been verified by quality control. Filter data based on data entry duration, number of codes, number of modifications, error type, and major ICD code categories. Use kernel density estimation and association rule mining algorithms to statistically analyze the distribution characteristics of data under each scenario. Combine clinical pathway standards and treatment guidelines to supplement scenario constraints. Store scenario features and constraints in a structured manner to build a weighted medical scenario prior library, supporting quick query and retrieval by scenario dimension. A medical scenario system is constructed based on four dimensions: disease type, department, data entry process, and treatment level. Historical data from the past three years, verified by quality control and free of anomalies, is collected. Core fields such as data entry duration, number of codes, number of modifications, error type, and ICD code category are selected. A kernel density estimation algorithm is used to calculate the distribution characteristics of numerical fields. Business logic relationships between fields are mined through association rules, and scenario constraints such as timeliness and coding standards are supplemented by clinical pathway specifications. Scenario features, distribution parameters, and constraint rules are structured and stored as a weighted scenario prior library. A scenario index and fast query interface are established, supporting precise retrieval of corresponding prior knowledge by department, disease, and data entry process. This provides accurate scenario constraints for the GAN model, ensuring that the generated data closely reflects the actual patterns of medical business.

[0031] Step S1222: The generator uses a fully connected neural network. The input layer contains scene feature vectors and context feature vectors of missing data. The output layer designs output units according to the type of missing field. The discriminator uses a fully connected network. The input layer receives data and scene labels. The output layer outputs the joint probability of data authenticity and scene matching degree. The training adopts a phased strategy: In the first phase, the generator and discriminator are pre-trained based on historical real data of each scene to optimize the basic loss function. In the second phase, scene prior constraint loss is introduced. The deviation between the generated data and the distribution features of the corresponding scene prior library are included in the loss function and weighted and summed. The KL divergence on the validation set is used as an indicator to adopt an early stopping mechanism. A GAN model adapted to structured medical data is constructed, with both the generator and discriminator employing a fully connected neural network architecture. The generator takes scene feature vectors and missing data context vectors as input, and its output layer is designed with adaptation units based on numerical and coded fields. The discriminator takes data and scene labels as input and outputs a joint probability of realism and scene matching degree. The model employs a two-stage training process: the first stage pre-trains with real data from each scene to optimize the basic loss function; the second stage introduces scene prior constraint loss, incorporating a weighted average of the distribution deviation between the generated data and the prior library into the loss function, using KL divergence to measure distribution similarity, and employing an early stopping mechanism to avoid overfitting. The optimal model weights are retained during training to ensure that the generated data conforms to both statistical distribution and the business constraints of the medical scenario, improving the rationality of the data input.

[0032] Step S1223: Extract patient visit type, department code, surgical / treatment item name, and features of entered fields from the context of missing data. Encode them into low-dimensional vectors and input them into a scene classifier optimized by logistic regression. The classifier queries the feature matching rules of the scene prior library, outputs the confidence of the scene, selects scenes with a confidence of ≥0.75 as target scenes, and automatically associates similar scene clusters and uses a scene feature fusion strategy to generate comprehensive prior constraints when the confidence is lower than 0.75. To accurately identify medical scenarios with missing data, features such as visit type, department code, treatment items, and entered fields are extracted from the data context and encoded into low-dimensional vectors, which are then input into a scenario classifier optimized by logistic regression. The classifier queries a scenario prior library for matching rules and outputs the scenario confidence score. Scenarios with a confidence score ≥ 0.75 are identified as target scenarios. For edge scenarios with a confidence score below the threshold, similar scenario clusters are automatically associated, and comprehensive constraints are generated by fusing multiple scenario prior features. This mechanism balances the recognition accuracy of both common and special scenarios, avoids inaccurate filling due to scenario matching deviations, and reduces manual intervention, achieving automated and high-precision identification of missing data scenarios, laying the foundation for subsequent intelligent filling.

[0033] Step S1224: Input the prior feature vector of the target scene and the context feature vector of the missing data into the trained generator to generate initial filling values. When multiple fields are missing, fill them in order of field importance. Verify the logical consistency through the association rules between fields. Generate 3 sets of candidate filling values ​​for each missing field. Calculate the matching score of each set with the prior distribution of the scene. Select the highest score value as the final filling value. Scenario-based missing data imputation is performed by inputting the target scenario prior vector and context vector into a generator to generate initial imputation values. When multiple fields are missing, they are prioritized according to business importance, with core fields such as ICD codes and data entry duration being filled first. The logical relationships between fields are checked simultaneously during the imputation process, and if there are conflicts, the data is regenerated. Three sets of candidate values ​​are generated for each missing field, and the matching score of each set with the scenario prior distribution is calculated. The highest score is selected as the final imputation value. This method avoids the randomness of a single imputation value through multiple candidate selection and logical consistency checks, ensuring that the imputation value conforms to both the scenario distribution and the business relationships between fields, significantly improving the accuracy and business rationality of missing data repair.

[0034] Step S1225: Call the configurable rule engine to perform multi-dimensional verification on the populated data in terms of time dimension, format dimension, logical consistency, and compliance. If the verification passes, it is marked as valid and written to the database. If the verification fails, the reason is recorded and fed back to the generator to regenerate the populated value. If the retry fails, it is marked as pending manual review and pushed to the management terminal to display relevant information for manual correction. The system performs secondary validation on the populated data using multi-dimensional rules, invoking a configurable rule engine to verify time specifications, format standards, logical consistency, and medical insurance compliance. If the validation passes, the data is marked as valid and written to the database; if validation fails, the reason for failure is recorded, and the generator is fed back to regenerate the populated value, with a maximum of two retries; if the retries still fail, the data is marked for manual review and pushed to the management console to display scenario information, the population process, and the validation results. This mechanism forms a closed loop of "generation-validation-retry-review," ensuring both the efficiency of automated repair and the prevention of non-compliant data through manual oversight, ensuring that the repaired data fully complies with medical business and compliance requirements.

[0035] Step S1226: Classify the validated data by scenario, update it to the medical scenario prior library monthly, recalculate the scenario distribution characteristics and optimize the constraint rules, and incrementally fine-tune the GAN model quarterly using newly added historical valid data and manually corrected data. Calculate the data governance effect report for each scenario, including the data filling pass rate, validation pass rate and manual correction rate.

[0036] An iterative optimization mechanism for the scenario-based prior library and the GAN model is established. Monthly, validated fill data is categorized by scenario, updated to the prior library, and its distribution characteristics and constraint rules are recalculated and optimized. Quarterly, newly added qualified data and manually corrected samples are used to incrementally fine-tune the GAN model, improving the accuracy of fill data in edge scenarios. Simultaneously, metrics such as fill pass rate, validation pass rate, and manual correction rate for each scenario are statistically analyzed to generate a data governance performance report, providing a basis for scenario classification and rule optimization. This iterative mechanism allows the model and prior library to continuously adapt to business changes, constantly improving the automation and accuracy of intelligent repair while reducing manual governance costs.

[0037] Step S123: Filter the features of data entry volume, processing time, concurrent data entry users, and field modification frequency. Train an anomaly detection model using the isolated forest algorithm calibrated with medical business thresholds. The calibration thresholds are set in combination with the hospital's historical peak data and the pattern of medical insurance settlement cycles. The model monitors the data flow in real time and automatically identifies sudden scenarios such as peak visits, medical insurance settlement periods, and quality control spot checks. It triggers a weakly supervised labeling mechanism and uses the joint features of sudden increases in data entry volume and decreases in processing time, as well as increases in specific time periods and field modification frequency, as the label generation rules. It automatically labels the sudden scenario data and writes it into the model training set simultaneously. A system for unsupervised detection and automatic labeling of emergency scenarios is constructed. Core features include data entry volume, processing timeliness, concurrent users, and modification frequency. The model is trained using an isolated forest algorithm calibrated with medical business thresholds, which are set in conjunction with historical peak times and medical insurance cycles. The model monitors the data stream in real time, automatically identifying emergency scenarios such as peak patient visits, medical insurance settlements, and quality control spot checks. This triggers a weakly supervised labeling mechanism, automatically labeling data with scenario tags based on joint labeling rules such as "sudden increase in data entry volume + decreased timeliness" and "time period + increased modification frequency," and synchronizing these tags to the training set. This method expands the sample of emergency scenarios without manual labeling, improves the model's adaptability to abnormal fluctuations, and solves the problem of scarce emergency scenario samples.

[0038] Step S124: Use a logistic regression model to screen historical staff shortage / redundancy influencing factors, including department, time period, data entry type, and patient flow. Construct a set of confounding variables, calculate the propensity score for each historical data point based on the propensity score matching algorithm, match a subset of samples from the dataset that are similar to the characteristics of a reasonable staffing configuration scenario, and use them as debiased training data. Build a counterfactual reasoning framework, set reasonable staffing configuration parameters based on the debiased real data, simulate the performance of reasonable staffing configuration under different scenarios, generate virtual training samples containing dimensions of data entry time, error rate, and processing volume, verify the consistency between the virtual samples and the real data distribution through KL divergence, and merge the debiased real samples and virtual samples to form an unbiased model training dataset. To eliminate historical staffing bias, causal inference is used to screen influencing factors such as department, time period, data entry type, and patient traffic, constructing a set of confounding variables. Propensity score matching (PSM) is used to calculate data propensity scores, matching samples similar to those with reasonable staffing scenarios to form a debiased dataset. A counterfactual reasoning framework is built, setting standard staffing parameters to simulate the efficiency performance under reasonable staffing, generating virtual samples containing data entry time, error rate, and processing volume. KL divergence is used to verify distribution consistency. Finally, the debiased real samples and virtual samples are merged to form an unbiased training set, eliminating model bias caused by historical staffing imbalances and improving the fairness and guidance of prediction results.

[0039] Step S125: Real-time statistics of rule verification pass rate, abnormal data repair accuracy rate, and sudden scene identification accuracy rate indicators; review the governance effect according to the preset cycle; for new abnormal data not covered by hard rules, manually label and supplement to the configurable rule engine; periodically feed back the repaired data, labeled sudden scene samples, and bias-corrected samples to the intelligent repair model and the anomaly detection model to complete the incremental training of the model.

[0040] A closed-loop iterative mechanism for data governance is established, with real-time statistics on indicators such as rule validation pass rate, repair accuracy rate, and emergency scenario recognition rate, and governance effectiveness is reviewed weekly. New types of abnormal data not covered by the rules are manually labeled and added to the rule engine; repaired data, labeled samples, and bias-corrected samples are periodically fed back into the intelligent repair and anomaly detection model for incremental training. Through this closed loop of indicator monitoring, rule expansion, and model iteration, the adaptability and accuracy of the governance system are continuously improved, adapting to dynamic changes in medical business processes, coding standards, and policy requirements, ensuring the long-term effectiveness of data governance.

[0041] Step S130: Based on the inherent laws of medical scenarios and external influencing factors, the original time series data is split into scenario layers. Feature enhancement is achieved by introducing medical features and dynamic feature filtering. A hybrid modeling framework that adapts to different scenarios and a dynamically adjusted sliding window mechanism are adopted. A scenario-based modeling system for medical time-series data is constructed. Based on inherent internal patterns such as disease cycles and treatment patterns, as well as external factors such as medical insurance policies and holidays, the original time-series data is divided into three types of scenario sequences: routine, periodic, and sudden. Medical-specific features are introduced and redundant information is eliminated through dynamic filtering to improve feature effectiveness. A hybrid model architecture adapted to different scenarios is adopted, combined with a dynamic sliding window mechanism to adapt to data fluctuations. This ensures both the trend fitting accuracy of routine scenarios and the response speed of periodic and sudden scenarios, achieving accurate modeling of time-series data across all scenarios and providing technical support for manpower demand forecasting.

[0042] Step S131: Analyze the influencing factors of medical time series data, divide them into internal inherent laws and external influencing factors, define quantitative rules for routine scenarios, periodic scenarios and sudden scenarios based on the influencing factors, and transform them into executable logical conditions. Embed the rule engine, integrate the unsupervised anomaly detection algorithm of isolated forest and DBSCAN clustering, monitor the fluctuation amplitude and trend change of the original time series data in real time, automatically identify the start and end boundaries of sudden scenarios, and combine the scenario labels output by the rule engine to split the original time series data into three scenario sequences according to the time axis. Each sequence is labeled with scenario type, driving factor and time range attribute. This study analyzes the influencing factors of medical time-series data, categorizing them into internal inherent patterns and external influencing factors. It defines quantitative rules for three scenarios: routine, cyclical, and sudden, and transforms these rules into executable logic embedded in a rule engine. Integrating the Isolation Forest and DBSCAN clustering algorithms, it monitors data fluctuations and trend abrupt changes in real time, automatically identifying the start and end boundaries of sudden scenarios. Combined with rule engine tags, the time-series data is split into three scenario sequences along the time axis, labeled with scenario type, driving factors, and time range attributes. Scenario splitting allows for independent modeling of data with different fluctuation characteristics, avoiding insufficient fitting by a single model and improving the targeting and accuracy of subsequent modeling.

[0043] Step S132: Introduce multi-dimensional medical features, including external correlation features, hospital internal operation features, policy features, and business correlation features. Standardize all features and construct a dynamic feature screening module. Adopt a two-stage screening strategy. In the first stage, invalid features with an absolute value of correlation coefficient between input volume and processing efficiency < 0.2 are eliminated based on Pearson correlation coefficient and mutual information value. In the second stage, an online learning feature importance evaluator based on gradient boosting tree is introduced. The feature importance ranking is updated every hour, and the low contribution features in the bottom 10% of the ranking are automatically eliminated. At the same time, policy features and business correlation features with a contribution of ≥ 0.1 are forcibly retained. Four categories of medical features—external correlation, hospital operation, policy, and business correlation—are introduced and standardized through normalization and one-hot encoding. A two-stage dynamic screening module is constructed. In the first stage, weakly correlated features are eliminated using Pearson coefficients and mutual information. In the second stage, gradient boosting trees are used to evaluate feature importance online, updating the ranking hourly and eliminating the bottom 10% of low-contribution features while forcibly retaining core policy and business features. Dynamic screening reduces feature dimensionality, improves computational efficiency, retains key business information, avoids redundant features interfering with the model, and ensures the effectiveness and lightweight nature of the modeling input.

[0044] Step S133: Design a model and build a fusion mechanism for the scene sequence. For regular scenes, an improved LSTM model with an attention layer for medical scenes is used. For periodic scenes, an ARIMA model optimized by seasonal decomposition of STL is used and the parameters are adjusted in combination with medical cycle characteristics. For sudden scenes, a graph neural network is used and a system, indicator, and impact path map is built. A scenario-specific independent training + joint optimization strategy is adopted. First, the model is trained separately based on the historical data of each scenario. Then, a gated fusion network is built. The output ratio is dynamically adjusted with the prediction error rate of each scenario as the weight. The final loss function after fusion is the weighted sum of the model losses of each scenario. Medical business constraint loss is introduced during the training process. A scenario-specific hybrid modeling and fusion framework is constructed. For routine scenarios, an LSTM with medical attention is used; for periodic scenarios, an STL-ARIMA model is used; and for sudden events, a GNN graph model is employed. Each scenario is first trained independently, then a gated fusion network dynamically weights the output based on the prediction error rate. A joint loss function integrates the losses from each model with the losses due to business constraints. This hybrid architecture adapts to the data characteristics of different scenarios, and the gated fusion automatically allocates model weights, taking into account trend, periodic, and sudden change predictions. Business constraints ensure that the results comply with medical regulations, improving overall prediction accuracy and compliance.

[0045] Step S1331: A multi-layer LSTM network architecture is adopted, and a medical scene attention layer is embedded between the hidden layer and the output layer of the LSTM. The medical features such as department type, patient condition complexity, and consultation time are encoded into attention vectors. The attention weight of each feature is calculated through the attention mechanism, so that the model pays priority to scene information that is strongly related to the input volume / processing efficiency. The output layer adopts a linear activation function to output the predicted value of the input volume under the normal scene. The model training adopts a basic loss function and optimizer, combined with a learning rate decay strategy. An improved LSTM model is constructed using a multi-layer network architecture. A medical scenario attention layer is embedded between the hidden and output layers, encoding department, disease complexity, and consultation time into attention vectors. This attention mechanism focuses on core features. The output layer uses linear activation to output the predicted input values. Training employs mean squared error loss and the Adam optimizer, combined with a learning rate decay strategy. The attention mechanism enables the model to focus on key factors in medical operations, improving the accuracy of long-term trend fitting in routine scenarios and enabling accurate prediction of stable data.

[0046] Step S1332: Perform STL seasonal decomposition on the time series data of the periodic scenario to separate the trend component, seasonal component and residual component. For each component, adopt the appropriate fitting and smoothing method. Optimize the model parameters in combination with the medical cycle pattern. Determine the parameter range through the autocorrelation function and partial autocorrelation function. Select the optimal parameter combination based on the information criterion and set parameter constraints. During training, aim to minimize the prediction error of the seasonal component. Optimize the parameters using the grid search method. Select the corresponding basic loss function to improve the fitting accuracy of the periodic fluctuation. STL decomposition is performed on cyclical scenario data to separate trend, seasonal, and residual components, which are then fitted and smoothed separately. ARIMA parameters are optimized in conjunction with medical cycle analysis, with ranges determined by ACF and PACF, and the optimal combination selected using the AIC criterion. Parameters are constrained to be related to cycle length. With the goal of minimizing seasonal error, a grid search is used to optimize parameters, employing the mean absolute error loss. This decomposition and optimization improves the fitting accuracy of cyclical fluctuations, making it suitable for predictions of fixed-cycle scenarios such as medical insurance settlement and seasonal disease incidence.

[0047] Step S1333: Construct a heterogeneous graph of system, indicators, and impact paths. The graph nodes include system nodes, indicator nodes, and impact factor nodes. The edges include system-indicator related edges, indicator-indicator linkage edges, and impact factor-system trigger edges. Initialize the edge weights according to the corresponding logic. Use a graph convolutional network as the architecture. The input layer is the real-time feature vector of each node. Learn the linkage and impact weights between nodes through multiple GCN layers. The output layer outputs the predicted value of the input volume fluctuation under sudden scenarios. The model training adopts a loss function and optimizer adapted to outlier handling. A heterogeneous graph of "system-indicator-impact path" is constructed, with nodes containing systems, indicators, and influencing factors, and edges containing associations, linkages, and triggering relationships, with initialized weights. A two-layer GCN is used to learn the linkage weights of nodes, outputting the fluctuation value of input volume in sudden scenarios, and Huber loss is used during training to reduce the impact of outliers. Graph modeling captures the linkage effect of multiple systems, accurately fits the sudden change trend of sudden scenarios, and improves the predictive response capability under abnormal fluctuations.

[0048] Step S1334: Divide the split scene sequence data into a training set and a validation set. The training set adopts a rolling window training method, and the window size is dynamically set according to the scene characteristics. The improved LSTM, STL-ARIMA, and GNN models are trained independently. After each round of training, the prediction error is calculated on the validation set. An early stopping mechanism is used to control the training process and save the optimal model weights for each scene. The scene sequences were divided into training and validation sets in a 7:3 ratio, and a rolling window was used for training, with the window dynamically set according to the scene. LSTM, ARIMA, and GNN models were trained separately, and the validation error was calculated in each round. An early stopping mechanism was used to prevent overfitting and to save the optimal weights. Independent training for each scene ensured model adaptability, while the rolling window and early stopping improved training efficiency and generalization ability, avoiding prediction bias caused by overfitting.

[0049] Step S1335: Construct a gated fusion unit. The input is the prediction results of the three scenario models and their respective validation set prediction error rates. Calculate the gate weights based on the prediction error rates. The lower the error rate, the higher the weight of the model. Obtain the fusion output result by weighted summation of the weights. In the joint optimization stage, merge the training set data of the three scenarios and construct a joint loss function. The joint loss function includes the weighted sum of the basic losses of each scenario model and the medical business constraint loss. Fine-tune the parameters of each scenario model and the gate weights through joint training to make the fusion prediction result conform to the overall time series pattern. A gated fusion unit is constructed, taking the prediction results and error rates of the three models as input, calculating weights inversely proportional to the error rates, and then summing them to obtain the fused output. Joint optimization merges the training sets, constructing a joint loss including model loss and business constraints, and fine-tuning the models and gate weights. Dynamic weight allocation ensures that the superior model dominates the output, and joint optimization improves the overall fit, ensuring that the predictions conform to global temporal patterns.

[0050] Step S1336: Transform the hard rules of medical business into mathematical constraints, use the slack variable method to handle inequality constraints, construct a constraint loss function, and integrate the constraint loss function into the joint loss function.

[0051] Medical business rules are transformed into mathematical constraints, and the slack variable method is used to handle inequality constraints. A constraint loss function is constructed and incorporated into the joint loss. The constraints cover business specifications such as timeliness, manpower, and error rate, ensuring that model training balances accuracy and compliance, avoids outputting non-compliant results, and guarantees that the prediction results meet the requirements of medical business management.

[0052] Step S134: Using the coefficient of variation as a quantitative evaluation index of data fluctuation intensity, set corresponding basic window thresholds in combination with scenario type, configure differentiated initial windows and dynamic adjustment rules for each scenario. For regular scenarios, the window size is adapted according to the fluctuation intensity, for periodic scenarios, it is flexibly adjusted according to the business cycle length, and for sudden scenarios, an adaptation logic of emergency contraction followed by gradual expansion is adopted. The fluctuation coefficient of the current data segment is dynamically updated through real-time calculation mode, and the window size is automatically iterated in combination with scenario labels to ensure the accuracy of time series feature extraction and resource utilization efficiency.

[0053] The coefficient of variation is used to quantify the intensity of data fluctuations, and window thresholds are set according to the scenario: the size is adjusted according to the fluctuation in normal scenarios, adapted to the period length in periodic scenarios, and first contracted and then expanded in sudden scenarios; the fluctuation coefficient is calculated in real time every 5 minutes, and the window is automatically iterated in combination with the scenario label. The dynamic window ensures the efficiency of large windows in stable scenarios and improves the accuracy of small windows in sudden scenarios, balancing the accuracy of feature extraction and the utilization of computing resources, and optimizing modeling performance.

[0054] Step S140: Construct a federated learning architecture for local data storage and parameter encryption interaction. Protect sensitive features through differential privacy technology adapted to medical scenarios. Use feature distillation method to indirectly infer information using non-sensitive features to compensate for the information loss of desensitized data. Combine dynamic privacy and performance balance adjustment mechanism.

[0055] A federated learning architecture adapted to medical scenarios is constructed, adopting a distributed deployment model of "local nodes + coordination center." Lightweight training models are deployed on each medical business system node, and raw medical data is stored and trained only locally, avoiding cross-node data transfer to meet compliance requirements. Adaptive differential privacy technology for medical scenarios dynamically adjusts the noise injection intensity according to the sensitivity level of the patient's condition, achieving precise protection of sensitive features. A knowledge distillation architecture is employed, using a teacher model trained on complete local data as a benchmark to guide student models on anonymized data to learn feature mapping relationships, compensating for information loss caused by anonymization. A dynamic privacy-performance balance adjustment mechanism is established, monitoring three core indicators in real time: model prediction accuracy, training convergence speed, and privacy leakage risk. Based on preset strategies, noise intensity and encryption levels are automatically adjusted to ensure both patient privacy and model training performance, achieving synergistic optimization of "usable but invisible" medical data and model performance.

[0056] Step S141: Deploy lightweight local models on each medical business system node. The original medical data is stored locally and is not transmitted across nodes. Each node only outputs model training parameters. A federated coordination center is built to be responsible for parameter aggregation and distribution. A federated averaging algorithm is used to complete the global aggregation of multi-node model parameters. Homomorphic encryption is used throughout the parameter interaction process. At the same time, an identity authentication and permission verification module is configured for local nodes. Only authorized nodes participate in federated training. Lightweight LSTM or CNN local models are deployed on various medical business system nodes, such as HIS, LIS, and medical record systems. The model parameter size is controlled within 1 million to adapt to the node hardware resources. Raw medical data is stored in a local encrypted database, and model parameters are only output after training is completed, without any raw data being transmitted across nodes. A federated coordination center is built, and a federated averaging algorithm is used to perform parameter aggregation: after each node completes 10 rounds of local training, the parameters are uploaded to the coordination center through an encrypted channel. The center calculates global parameters based on the proportion of data volume of each node, aggregates them, and then distributes them to each node to complete the model update. The parameter interaction uses the BFV homomorphic encryption algorithm throughout, with a key length of 2048 bits to ensure that the parameters cannot be decrypted during transmission. Each local node is configured with a digital certificate-based identity authentication module and an OAuth2.0 permission management system. Only authorized nodes that have passed compliance review can access the federated network. When a node accesses the network, it must verify the device fingerprint, IP whitelist, and business authorization scope to prevent unauthorized access and parameter tampering.

[0057] Step S142: Classify and label sensitive features such as patient condition and diagnosis results, divide the protection intensity range according to the severity of the condition and the level of privacy sensitivity, embed an adaptive noise injection module in the forward propagation and gradient calculation stage of local model training, dynamically adjust the noise intensity according to the medical scenario and the level of sensitivity, establish a privacy budget allocation mechanism, split the total privacy budget according to the training round and node permissions, monitor the privacy consumption progress in real time and dynamically adjust the noise intensity or the training status of high-sensitivity features. Features are categorized into three levels based on the sensitivity of medical data: Level 1 (high sensitivity) includes patient diagnosis and gene testing results; Level 2 (medium sensitivity) includes treatment plans and medication records; and Level 3 (low sensitivity) includes visit time and department information, with corresponding protection strength ranges of [0.8, 1.2], [0.4, 0.8], and [0.1, 0.4], respectively. An adaptive noise injection module is embedded in the forward propagation feature layer and gradient calculation stage of the local model training. Based on the current data scenario (e.g., emergency room, inpatient) and sensitivity level, Gaussian noise is dynamically generated and superimposed. A total privacy budget allocation mechanism of ε=5 is established, with each round's budget of ε=0.05 divided according to the total training rounds (100 rounds), and a 1.2 times budget weight assigned to tertiary hospital nodes. The privacy budget monitoring module tracks the consumption progress in real time. When the remaining budget is less than 20% of the total budget, the noise intensity is automatically increased by 30%, or the gradient update of Level 1 sensitive features is paused, ensuring that the risk of privacy leakage is controlled within compliance limits.

[0058] Step S143: Select non-sensitive business features as distillation input, and use disease complexity and treatment level business information as distillation output. Use knowledge distillation method, and use the teacher model trained on local complete data as the benchmark to guide the student model on desensitized data to learn the feature mapping relationship. Reduce the output difference through gradient matching and distribution alignment, and combine the feedback data of the quality control system to back-optimize the distillation model and update the feature mapping weights. Non-sensitive business features such as data entry duration, number of codes, number of modifications, and ICD code categories are selected as distillation inputs, while core business information such as disease complexity, treatment level, and quality control error rate are used as distillation outputs. A dual-model architecture of "teacher-student" is adopted: the teacher model is trained on local, un-anonymized historical valid data, while the student model is deployed on a local node of the federated learning system and trained on anonymized data. Through a knowledge distillation loss function, the student model learns the intermediate layer feature mappings and output distributions of the teacher model. Gradient matching uses MSE loss to align the gradients of the teacher and student models, and distribution alignment achieves output distribution consistency by minimizing KL divergence. After each training round, error types and processing efficiency feedback data from the quality control system are collected, and the feature mapping weights of the student model are updated through backpropagation. Monthly incremental fine-tuning of the distillation model is performed based on new data to ensure that the inference accuracy of non-sensitive features for key business information is ≥90%, compensating for the information loss from the anonymized data.

[0059] Step S144: Build a real-time monitoring module to collect indicators such as model prediction accuracy, training convergence speed, and privacy leakage risk value. Preset three adjustment strategies: privacy priority, performance priority, and balance mode. Automatically adjust the differential privacy noise intensity and federated parameter encryption level according to the indicator monitoring results. The adjustment process adopts a smooth transition strategy. A distributed real-time monitoring module was built, deploying lightweight data acquisition probes on each local node and a global indicator aggregation module at the federal coordination center to simultaneously collect three core indicators: model prediction accuracy, training convergence speed, and privacy leakage risk value. Three preset adjustment strategies were implemented: a privacy-first mode with a noise intensity benchmark of 0.8 and a default high encryption level, allowing a 30% reduction in training convergence speed; an efficiency-first mode with a noise intensity benchmark of 0.3 and a default medium encryption level, requiring prediction accuracy ≥88%; and a balanced mode with an intermediate benchmark, a noise intensity of 0.5-0.7, and a medium-high encryption level, constraining the privacy risk value to ≤0.6 and prediction accuracy ≥85%. Indicator fluctuations were calculated based on a sliding window (window size 10 training rounds). When an indicator deviates from the current strategy threshold by more than 10%, a smoothing adjustment mechanism was triggered, gradually adjusting the noise intensity and encryption level using an exponentially weighted average algorithm to avoid parameter mutations causing model training oscillations.

[0060] Step S1441: Clarify the medical scenario adaptation calculation rules for the three types of indicators, deploy lightweight collection probes synchronously on local nodes and the federal coordination center, collect local training-related data and global aggregated related data respectively, summarize and report to the time series monitoring library at preset time intervals, form a traceable view of the whole process indicators, and solidify the data foundation for dynamic adjustment of indicators. The following calculation rules are defined for the medical scenario adaptation of three types of indicators: Prediction accuracy is calculated using mean absolute error (MAE), with a baseline value of ≥90% for internal medicine outpatient scenarios and ≥85% for surgical scenarios; Training convergence speed is measured by the number of iterations required for the model loss to decrease to a stable value, with ≤500 iterations for routine scenarios and ≤300 iterations for emergency scenarios; Privacy leakage risk is calculated by weighting the privacy budget consumption rate (current consumption / total budget) and the model reverse attack success rate, with weights of 0.7 and 0.3 respectively, and a security threshold of ≤0.7. Local node probes are deployed to collect training loss, prediction error, and privacy budget consumption data in real time; the coordination center probes aggregate global parameter consistency and cross-node training progress synchronization rate data. All indicators are collected at 5-minute intervals and reported to an AES-256 encrypted time-series monitoring library. The monitoring library stores data in three dimensions: "node-training round-indicator type," supporting the tracing of indicator changes along a timeline and providing complete data support for dynamic adjustment.

[0061] Step S1442: Set differentiated privacy protection baselines for the three modes of privacy priority, performance priority, and balance. The privacy priority mode strengthens the protection intensity and strictly constrains privacy risks, while the performance priority mode reduces the protection intensity and prioritizes the model performance. The balance mode adopts an intermediate baseline to simultaneously constrain the two types of indicators. Bind the corresponding noise intensity range, encryption level, and indicator trigger threshold to each mode. The privacy-first mode sets the noise intensity range to [0.8, 1.2], with a high encryption level (BFV algorithm, 2048-bit key length), a privacy leakage risk threshold ≤ 0.5, and a training convergence speed tolerance threshold ≤ 800 rounds. The performance-first mode sets the noise intensity range to [0.2, 0.5], with a medium encryption level (Paillier algorithm, 1024-bit key length), a prediction accuracy threshold ≥ 88%, and a privacy risk tolerance threshold ≤ 0.8. The balanced mode sets the noise intensity range to [0.5, 0.8], with a medium-high encryption level, and synchronization constraints of privacy risk ≤ 0.6, prediction accuracy ≥ 85%, and convergence speed ≤ 600 rounds. Trigger thresholds are assigned to each mode: in privacy-first mode, a privacy risk value > 0.5 triggers a noise intensity increase of 0.1; in performance-first mode, a prediction accuracy < 88% triggers a noise intensity decrease of 0.05; in balanced mode, any indicator exceeding the threshold triggers the corresponding adjustment to ensure accurate matching between the mode and parameters.

[0062] Step S1443: Real-time evaluation of the deviation between the collected indicators and the strategy baseline, classifying them into three levels of status, and automatically matching the direction of protection strength adjustment based on the judgment results of privacy risk exceeding the standard or model performance being insufficient; if there are no abnormalities, the parameters are kept stable. The evaluation engine calculates the deviation rate between the collected metrics and the current strategy baseline in real time. The deviation rate is calculated as |Actual Metric - Baseline Metric| / Baseline Metric, and is divided into three levels: a deviation rate < 5% is normal, maintaining the current parameters; 5% ≤ Deviation Rate ≤ 10% is a warning state, marking an anomaly but not triggering adjustment; and a deviation rate > 10% triggers adjustment. When the deviation rate for privacy leakage risk exceeds 10%, it is considered an excessive privacy risk, and the adjustment direction is to increase noise intensity and encryption level. When the prediction accuracy deviation rate is < -10% or the convergence speed deviation rate is > 10%, it is considered insufficient model performance, and the adjustment direction is to reduce noise intensity and encryption level. The evaluation engine performs deviation calculations every 5 minutes, outputting adjustment instructions. When there are no anomalies, the parameters remain stable, ensuring the accuracy and timeliness of adjustments.

[0063] Step S1444: Calculate the target noise intensity by combining the strategy and indicator deviation, adopt a progressive step size adjustment, dynamically adjust the step size in conjunction with the privacy budget consumption progress and prioritize the protection of highly sensitive features, and ensure the uniformity of noise parameters across all nodes. Based on the current strategy baseline and indicator deviation rate, the target noise intensity is obtained through a linear calculation model: Target value = Current value + Deviation rate × Adjustment coefficient (adjustment coefficient is set to 0.1). A progressive step size adjustment is adopted, with each adjustment controlled within ±0.05, and only one adjustment is performed per training round, gradually approaching the target value. The step size is dynamically adjusted in conjunction with the privacy budget consumption progress: when the remaining budget > 50%, the normal step size is used; when 20% ≤ remaining budget ≤ 50%, the step size is reduced to 0.03; when the remaining budget < 20%, the step size is reduced to 0.01, and the noise intensity of the first-level sensitive feature is not lower than 0.8. The adjustment command is synchronized to all local nodes in real time through an encrypted channel. After receiving the command, the nodes enable the new noise parameters in the next training round, ensuring parameter uniformity across the entire cluster and avoiding inconsistent data distribution.

[0064] Step S1445: Divide the encryption levels into different security strengths and overheads, switch the levels step by step according to the adjustment direction, and ensure uninterrupted parameter transmission through the parallel transition mechanism of dual encryption protocols; The encryption levels are low (AES-128, suitable for non-sensitive parameters), medium (Paillier-1024, suitable for general parameters), and high (BFV-2048, suitable for core gradient parameters), with corresponding increases in security strength and computational costs of 1, 2.5, and 4 times the baseline, respectively. The encryption level is switched incrementally according to the adjustment direction: for privacy enhancement, the level is increased in the order of low → medium → high, upgrading only one level at a time; for performance optimization, the level is decreased in the order of high → medium → low, also switching incrementally. A parallel transition mechanism using dual encryption protocols is employed during the switching process: first, the target level encryption protocol is synchronously enabled at the coordination center and local nodes. After all nodes confirm receipt and feedback, the dual protocols continue to run for one round of training. Once parameter transmission is confirmed to be correct, the original level encryption protocol is disabled, ensuring uninterrupted and unlost parameter interaction while balancing security and continuity.

[0065] Step S1446: Use exponential weighted filtering to smooth the adjustment parameters, set adjustment intervals and amplitude limits, lock the core training process, and avoid training oscillations caused by frequent parameter switching; An exponentially weighted filtering algorithm is used to smooth the adjustment parameters, with the filtering coefficient α set to 0.8. The calculation formula is: smoothed value = α × current adjusted value + (1-α) × historical smoothed value, eliminating erroneous adjustments caused by instantaneous index fluctuations. The minimum adjustment interval is set to 30 minutes, and the maximum adjustment range per instance is ±0.2, avoiding frequent parameter switching within a short period. After the adjustment command is issued, the core training process is locked using a distributed lock, pausing parameter updates for one round of training. Once all nodes have completed parameter synchronization and confirmed stability, the process is unlocked and resumes full-speed training. When two consecutive adjustments are in opposite directions, a fluctuation suppression mechanism is automatically triggered, pausing adjustments for one hour and using historically optimal parameters to prevent model training oscillations and ensure the overall stability of federated learning.

[0066] Step S1447: After adjustment, continuously monitor the recovery status of the indicators, make secondary fine-tuning if necessary, record information throughout the process and write it into the audit log, and regularly optimize relevant parameters to improve the accuracy of the balance between privacy and performance.

[0067] After adjustment, the recovery status of the indicators is continuously monitored, with an observation window of 5 training rounds. If the indicators still exceed the policy threshold by more than 5% after the window ends, a second fine-tuning is initiated, with the fine-tuning step size being 50% of the initial adjustment step size. All adjustment information is recorded throughout the process, including adjustment time, trigger indicator deviation, target parameter value, actual adjustment value, and indicator change results, and written to the blockchain audit log to ensure traceability. Quarterly, based on historical adjustment data, a linear regression algorithm is used to optimize the policy threshold, adjustment step size, and smoothing coefficient. For example, when a department experiences frequent triggers of adjustments due to unforeseen scenarios, the performance threshold tolerance for that scenario is specifically lowered to improve adjustment accuracy and achieve continuous optimization that balances privacy and performance.

[0068] Step S145: Generate operation logs for the entire process of local data storage, encrypted parameter transmission, noise injection, and feature distillation. Record node information, training rounds, and privacy parameter adjustment records to meet audit requirements. Regularly evaluate the privacy protection effect and model performance indicators, optimize noise adjustment strategies, feature distillation weights, and balance adjustment thresholds, and adapt to medical business scenarios and compliance policy updates.

[0069] Operation logs are generated for the entire federated learning process. These logs include local data storage encryption methods, parameter transmission encryption levels, noise injection strength and type, feature distillation weight update records, node authentication information, and details of training round parameter adjustments. Logs are stored on an encrypted disk array using a "daily backup + monthly archiving" strategy, with a retention period of at least three years, meeting medical data auditing and compliance requirements. A comprehensive evaluation is conducted quarterly. Privacy protection effectiveness is measured by privacy budget consumption compliance rate and model reverse attack success rate, while model performance is evaluated by prediction accuracy and convergence speed. Optimization is based on evaluation results: if new anomalous data increases, the noise adjustment strategy is adjusted; if distillation accuracy decreases, the feature mapping weights are updated; if policy requirements escalate, the balance adjustment threshold is tightened, ensuring the technical solution continuously adapts to changes in medical business scenarios and updates to data security regulations.

[0070] Based on the same inventive concept, please refer to Figure 2 This paper shows a schematic block diagram of the structure of an intelligent decision-making system 100 for human resource planning of personnel, which is provided in an embodiment of this application for executing the intelligent decision-making method for human resource planning of personnel described above. The intelligent decision-making system 100 for human resource planning of personnel may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0071] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the intelligent decision-making system 100 for human resource planning of personnel data entry, and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the intelligent decision-making system 100 for human resource planning of personnel data entry, and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate and interact with external systems through the communication unit 110.

[0072] The processor 130 is the control center of the intelligent decision-making system 100 for human resource planning of personnel. It connects to various parts of the system via various interfaces and lines. By running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, it performs various functions and processes data of the intelligent decision-making system 100, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to realize the intelligent decision-making method for human resource planning of input personnel provided in the aforementioned method embodiments.

[0073] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

[0074] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for intelligent decision making in workforce planning for an organization, the method comprising: Includes the following steps: ​ By deploying a cross-system data access gateway that is adapted to the compliance requirements of the medical intranet, and combining a standardized terminology dictionary and semantic intelligent matching technology, a dynamic mapping relationship of synonymous fields in multiple systems is established to achieve unified data capture and standardized feature conversion of data scattered in different medical business systems. Furthermore, through a feature version management mechanism, dynamic changes in system field definitions are adapted to achieve full-dimensional fusion and consistent modeling of multi-source heterogeneous data. A full-chain data governance system is built based on hard rules for medical business and intelligent algorithms. Data integrity and accuracy are improved through real-time verification and scenario-based intelligent repair. Unsupervised anomaly detection is used to automatically label the characteristics of sudden scenarios. Causal inference technology is used to remove the influence of historical human resource allocation deviations and supplement virtual training samples. Based on the inherent laws of medical scenarios and external influencing factors, the original time series data is divided into scenario layers. Feature enhancement is achieved by introducing medical features and dynamic feature filtering. A hybrid modeling framework with multiple models adapted to different scenarios and a dynamically adjusted sliding window mechanism are adopted. We construct a federated learning architecture for local data storage and encrypted parameter interaction. We protect sensitive features through differential privacy technology adapted to medical scenarios, and use feature distillation method to indirectly infer information from non-sensitive features to compensate for the information loss of desensitized data. We also combine dynamic privacy and performance balance adjustment mechanisms.

2. The intelligent decision-making method for human resource planning of personnel input as described in claim 1, characterized in that: The process involves deploying a cross-system data access gateway adapted to the compliance requirements of the medical intranet, combining a standardized terminology dictionary and semantic intelligent matching technology to establish a dynamic mapping relationship between synonymous fields across multiple systems. This enables unified data capture and standardized feature conversion across different medical business systems. Furthermore, a feature version management mechanism adapts to dynamic changes in system field definitions, achieving full-dimensional fusion and consistent modeling of multi-source heterogeneous data. This includes: The gateway architecture is designed in accordance with compliance requirements, and a localized cluster deployment compatible with multiple technology stacks is adopted. It integrates private protocol parsing and signal conversion capabilities of legacy devices, and performs secure and compliant collection of medical multi-system data through encrypted transmission, fine-grained access control and offline caching mechanisms. By integrating medical dictionaries to build a localized and standardized terminology system, and by collecting metadata from multiple systems through gateway plugins to form a unified dictionary, semantic completion of missing descriptive fields is performed, thus consolidating the terminology and metadata foundation for unified data integration. Based on a fine-tuned pre-trained model in the medical field, semantic intelligent matching capabilities are built. The dynamic mapping relationship of synonym fields in multiple systems is established and stored in the form of a graph. The mapping relationship is adaptively updated by combining timed scanning and manual review. For different types of medical business systems, a multi-mode data capture strategy combining real-time push and timed polling is designed. After extracting structured / semi-structured data, field matching, indicator calculation and data format standardization conversion are completed according to a unified feature paradigm. A feature version library is built based on the configuration center. Through field change monitoring, shadow field compatibility, historical data recalculation, and audit log recording, a feature version management mechanism that adapts to dynamic field changes is established to ensure long-term consistency of multi-source data.

3. The intelligent decision-making method for human resource planning of personnel input as described in claim 2, characterized in that: The semantic intelligent matching capability is built based on a fine-tuned pre-trained model in the medical field. It establishes and stores dynamic mapping relationships of synonymous fields across multiple systems in the form of a graph. Through a combination of timed scanning and manual review, it achieves adaptive updates of the mapping relationships, including: Based on the text features in the medical field, a general pre-trained model is fine-tuned to construct a semantic matching model for medical scenarios. The model input is replaced with field names, field annotations, and business purpose text data from the medical business system. Irrelevant general corpus is removed, and model adaptation training is completed using a few-shot learning method to optimize the model's semantic understanding ability of medical input scenario terms. The metadata fields of each medical business system are input into the fine-tuned semantic matching model to generate a high-dimensional semantic feature vector. The semantic correlation between fields of different systems is calculated by the cosine similarity algorithm. The grading judgment threshold is preset, and high-confidence synonymous fields, suspected synonymous fields and unrelated fields are automatically filtered out to complete the initial intelligent matching. Design a medical field mapping knowledge graph architecture, with each system field as an independent node and synonym, association, and hierarchical relationships as connecting edges. Label each mapping relationship with matching confidence, source system, matching time, and applicable business scenario attributes. Use a graph database to complete the structured storage and association query of the knowledge graph. Configure a scheduled scanning task to traverse all connected medical business system metadata interfaces at a fixed period, monitor three types of data changes in real time: new fields, field name modifications, and business attribute changes, automatically input the changed metadata into the semantic matching model, trigger incremental matching calculations, and synchronously update the mapping relationships in the knowledge graph. For suspected synonymous fields and low-confidence mapping relationships generated by automatic matching, a manual review process is triggered. The management terminal displays the field source, semantic similarity, and business context information, and provides three operation options: confirm, correct, and reject. The review results are written back to the knowledge graph in real time. Establish a model iteration mechanism, using manually reviewed and corrected mapping data as new training samples, periodically feed them back to the semantic matching model, complete small-scale incremental fine-tuning, and continuously improve the accuracy and confidence of the model's automatic matching. Construct a mapping relationship change log and backtracking mechanism to fully record the operation information and mapping content of every automatic update and manual correction. Support querying historical mapping versions by time, system, and field dimensions to achieve traceability and rollback of mapping relationships and complete full lifecycle adaptive update management.

4. The intelligent decision-making method for human resource planning of personnel input as described in claim 1, characterized in that: The aforementioned end-to-end data governance system, built upon hard rules for medical operations and intelligent algorithms, improves data integrity and accuracy through real-time verification and scenario-based intelligent repair. It utilizes unsupervised anomaly detection to automatically label features of sudden scenarios and employs causal inference techniques to remove the influence of historical human resource allocation biases and supplement virtual training samples. This includes: We sort out the relevant rules for medical business and transform them into executable logic. We build a rule engine that supports flexible configuration by department and business type. We embed a real-time verification module in the data collection link. When data is accessed, rule verification is triggered synchronously. Abnormal data is marked and supplemented. For problems that can be automatically corrected, we associate them with professional dictionaries and reasonable historical data ranges to complete the real-time correction. Based on historical valid data, the data distribution characteristics of different diseases, departments, and data entry stages are statistically analyzed. A medical scenario prior library is constructed in combination with clinical pathways and treatment guidelines. A generative adversarial network is used to build an intelligent repair model. The model is trained with the scenario prior library as a constraint, so that the generator learns the real distribution pattern of data in different scenarios. For missing data, the medical scenario to which it belongs is identified through the data context. The model generates a fill value that conforms to the distribution characteristics of the scenario. After filling, a second verification is performed through the medical business rules. The model selects data based on features such as data entry volume, processing time, concurrent data entry users, and field modification frequency. It then trains an anomaly detection model using an isolated forest algorithm calibrated with medical business thresholds. These calibration thresholds are set in conjunction with historical peak data from hospitals and patterns in medical insurance settlement cycles. The model monitors the data stream in real time and automatically identifies sudden scenarios such as peak patient visits, medical insurance settlement periods, and quality control spot checks. This triggers a weakly supervised labeling mechanism, using the combined features of a sudden increase in data entry volume and a decrease in processing time, as well as an increase in the frequency of field modifications during specific time periods, as label generation rules. The model automatically labels the data for sudden scenarios and writes them into the model training set simultaneously. A logistic regression model was used to screen historical factors affecting staff shortages / redundancy, including department, time period, data entry type, and patient flow. A set of confounding variables was constructed, and a propensity score was calculated for each historical data point based on a propensity score matching algorithm. A subset of samples similar to the characteristics of a reasonable staffing configuration scenario was matched from the dataset and used as debiased training data. A counterfactual reasoning framework was built, and reasonable staffing configuration parameters were set based on the debiased real data. The performance of reasonable staffing configuration under different scenarios was simulated, and virtual training samples containing dimensions such as data entry time, error rate, and processing volume were generated. The consistency between the virtual samples and the real data distribution was verified by KL divergence. The debiased real samples and virtual samples were merged to form an unbiased model training dataset. The system performs real-time statistical analysis of rule validation pass rate, abnormal data repair accuracy rate, and sudden scene identification accuracy rate, and reviews the governance effect according to a preset cycle. For new abnormal data not covered by hard rules, manual annotation is used to supplement the configurable rule engine. The repaired data, annotated sudden scene samples, and bias-corrected samples are periodically fed back to the intelligent repair model and the anomaly detection model to complete incremental training of the models.

5. The intelligent decision-making method for human resource planning of personnel input as described in claim 4, characterized in that: The intelligent repair model is built using a generative adversarial network (GAN). The model is trained with a scenario prior library as constraints, enabling the generator to learn the true distribution patterns of data in different scenarios. For missing data, the model identifies the medical scenario it belongs to through data context, generates filler values ​​that conform to the distribution characteristics of that scenario, and then performs secondary verification according to the aforementioned medical business rules. This includes: Medical scenarios are categorized into four dimensions: disease type, department, data entry process, and treatment level. Historical data without abnormalities from the past three years that have undergone quality control verification are collected. Data entry time, number of codes, number of modifications, error type, and ICD code category fields are selected. Kernel density estimation and association rule mining algorithms are used to statistically analyze the distribution characteristics of data under each scenario. Clinical pathway standards and treatment guidelines are combined to supplement scenario constraints. Scenario features and constraints are stored in a structured manner to build a weighted medical scenario prior library, which supports quick query and retrieval by scenario dimension. The generator uses a fully connected neural network. The input layer contains scene feature vectors and context feature vectors for missing data. The output layer designs output units according to the type of missing field. The discriminator uses a fully connected network. The input layer receives data and scene labels. The output layer outputs the joint probability of data authenticity and scene matching degree. The training adopts a phased strategy: the first phase pre-trains the generator and discriminator based on historical real data of each scene and optimizes the basic loss function. The second phase introduces scene prior constraint loss, incorporates the deviation between the generated data and the distribution features of the corresponding scene prior library into the loss function and sums them in a weighted manner, and adopts an early stopping mechanism with KL divergence on the validation set as the indicator. The patient visit type, department code, surgical / treatment item name, and entered field features are extracted from the context of missing data. After being encoded into low-dimensional vectors, they are input into a scene classifier optimized by logistic regression. The classifier queries the feature matching rules of the scene prior library and outputs the confidence of the scene. Scenes with a confidence of ≥0.75 are selected as target scenes. When the confidence is lower than 0.75, similar scene clusters are automatically associated and a scene feature fusion strategy is used to generate comprehensive prior constraints. The prior feature vector of the target scene and the context feature vector of the missing data are input into the trained generator to generate initial fill values. When multiple fields are missing, they are filled in order of importance. Logical consistency is verified by the association rules between fields. Three sets of candidate fill values ​​are generated for each missing field. The matching score of each set with the prior distribution of the scene is calculated, and the highest score value is selected as the final fill value. The configurable rule engine is invoked to perform multi-dimensional checks on the populated data, including time, format, logical consistency, and compliance. If the check passes, the data is marked as valid and written to the database. If the check fails, the reason is recorded and fed back to the generator to regenerate the populated value. If the retry fails, the data is marked as pending manual review and pushed to the management end to display relevant information for manual correction. The validated data is categorized by scenario and updated monthly to the medical scenario prior library. The scenario distribution characteristics are recalculated and the constraint rules are optimized. Every quarter, the GAN model is incrementally fine-tuned using newly added historical valid data and manually corrected data. The data governance effect report is generated by statistically analyzing the data filling pass rate, validation pass rate, and manual correction rate for each scenario.

6. The intelligent decision-making method for human resource planning of personnel input as described in claim 1, characterized in that: Based on the inherent patterns of medical scenarios and external influencing factors, the original time-series data is segmented into scenario layers. Feature enhancement is achieved by introducing medical features and dynamic feature filtering. A hybrid modeling framework adapting to different scenarios and a dynamically adjusted sliding window mechanism are adopted, including: The influencing factors of medical time-series data are analyzed and divided into internal inherent laws and external influencing factors. Based on these influencing factors, quantitative rules for routine scenarios, periodic scenarios, and sudden scenarios are defined and transformed into executable logical conditions. A rule engine is embedded, and unsupervised anomaly detection algorithms of isolated forest and DBSCAN clustering are integrated to monitor the fluctuation amplitude and trend changes of the original time-series data in real time, automatically identify the start and end boundaries of sudden scenarios, and combine the scenario labels output by the rule engine to split the original time-series data into three scenario sequences according to the time axis. Each sequence is labeled with scenario type, driving factor, and time range attribute. Multi-dimensional medical features are introduced, including external correlation features, hospital internal operation features, policy features, and business correlation features. All features are standardized and a dynamic feature screening module is constructed. A two-stage screening strategy is adopted. In the first stage, invalid features with an absolute value of correlation coefficient between input volume and processing efficiency <0.2 are eliminated based on Pearson correlation coefficient and mutual information value. In the second stage, an online learning feature importance evaluator based on gradient boosting tree is introduced to update the feature importance ranking every hour and automatically eliminate the low contribution features in the bottom 10% of the ranking, while forcibly retaining policy features and business correlation features with a contribution of ≥0.

1. Models and fusion mechanisms are designed for scenario sequences. For routine scenarios, an improved LSTM model with an attention layer for medical scenarios is used. For periodic scenarios, an ARIMA model optimized by seasonal decomposition of STL is used and parameters are adjusted in combination with medical cycle characteristics. For sudden scenarios, a graph neural network is used and a system, indicator, and impact path map is constructed. A scenario-specific independent training + joint optimization strategy is adopted. First, the model is trained separately based on historical data of each scenario. Then, a gated fusion network is constructed. The output ratio is dynamically adjusted with the prediction error rate of each scenario as the weight. The final loss function after fusion is the weighted sum of the model losses of each scenario. Medical business constraint loss is introduced during the training process. The coefficient of variation is used as a quantitative evaluation indicator of data fluctuation intensity. A corresponding basic window threshold is set according to the scenario type. Differentiated initial windows and dynamic adjustment rules are configured for each scenario. The window size is adapted according to the fluctuation intensity in regular scenarios, flexibly adjusted according to the business cycle length in periodic scenarios, and an adaptation logic of emergency contraction followed by gradual expansion is adopted in sudden scenarios. The fluctuation coefficient of the current data segment is dynamically updated through real-time calculation mode, and the window size is automatically iterated in combination with scenario labels to ensure the accuracy of time series feature extraction and resource utilization efficiency.

7. The intelligent decision-making method for human resource planning of personnel input as described in claim 6, characterized in that: The design of the model and the construction of the fusion mechanism for the scenario sequence are as follows: For routine scenarios, an improved LSTM model with an attention layer introduced for medical scenarios is used; for periodic scenarios, an ARIMA model optimized by STL seasonal decomposition and parameters adjusted in conjunction with medical cycle characteristics are used; for sudden scenarios, a graph neural network is used and a system, indicator, and impact path map is constructed. A scenario-specific independent training + joint optimization strategy is adopted. First, the model is trained separately based on historical data for each scenario, and then a gated fusion network is constructed. The output proportion is dynamically adjusted using the prediction error rate of each scenario as weight. The final loss function after fusion is the weighted sum of the losses of the models in each scenario. Medical business constraint loss is introduced during training, including: A multi-layer LSTM network architecture is adopted, and a medical scene attention layer is embedded between the hidden layer and the output layer of the LSTM. The medical features such as department type, patient condition complexity, and consultation time are encoded into attention vectors. The attention weight of each feature is calculated through the attention mechanism, so that the model pays priority to scene information that is strongly related to the input volume / processing efficiency. The output layer adopts a linear activation function to output the predicted value of the input volume under normal scenarios. The model training adopts a basic loss function and optimizer, combined with a learning rate decay strategy. STL seasonal decomposition is performed on the time series data of the periodic scenario to separate the trend component, seasonal component and residual component. For each component, an appropriate fitting and smoothing method is adopted. The model parameters are optimized in combination with the medical cycle pattern. The parameter range is determined by the autocorrelation function and partial autocorrelation function. The optimal parameter combination is selected based on the information criterion and parameter constraint relationship is set. During training, the goal is to minimize the prediction error of the seasonal component. The grid search method is used to optimize the parameters and the corresponding basic loss function is selected to improve the fitting accuracy of the periodic fluctuation. A heterogeneous graph of systems, indicators, and impact paths is constructed. The graph nodes include system nodes, indicator nodes, and impact factor nodes. The edges include system-indicator association edges, indicator-indicator linkage edges, and impact factor-system trigger edges. The edge weights are initialized according to the corresponding logic. A graph convolutional network is used as the architecture. The input layer is the real-time feature vector of each node. The linkage and impact weights between nodes are learned through multiple GCN layers. The output layer outputs the predicted value of the input volume fluctuation under sudden scenarios. The model training adopts a loss function and optimizer adapted to outlier handling. The split scene sequence data is divided into a training set and a validation set. The training set adopts a rolling window training method, and the window size is dynamically set according to the scene characteristics. The improved LSTM, STL-ARIMA and GNN models are trained independently. After each round of training, the prediction error is calculated on the validation set. An early stopping mechanism is used to control the training process and save the optimal model weights for each scene. A gated fusion unit is constructed, with the input being the prediction results of three scenario models and their respective validation set prediction error rates. Gating weights are calculated based on the prediction error rates, with higher weights for models with lower error rates. The fusion output is obtained by weighted summation of the weights. In the joint optimization stage, the training set data of the three scenarios are merged, and a joint loss function is constructed. The joint loss function includes the weighted sum of the basic losses of each scenario model and the medical business constraint loss. The parameters of each scenario model and the gate weights are fine-tuned through joint training to make the fusion prediction results conform to the overall time series pattern. The hard rules of medical operations are transformed into mathematical constraints, the slack variable method is used to handle inequality constraints, a constraint loss function is constructed, and the constraint loss function is integrated into the joint loss function.

8. The intelligent decision-making method for human resource planning of personnel input as described in claim 1, characterized in that: The constructed federated learning architecture for local data storage and parameter encryption interaction protects sensitive features through differential privacy technology adapted to medical scenarios, compensates for information loss in desensitized data by using feature distillation methods that indirectly infer information from non-sensitive features, and combines dynamic privacy and performance balance adjustment mechanisms, including: Lightweight local models are deployed on each medical business system node. Raw medical data is stored locally and not transmitted across nodes. Each node only outputs model training parameters. A federated coordination center is built to be responsible for parameter aggregation and distribution. A federated averaging algorithm is used to complete the global aggregation of multi-node model parameters. Homomorphic encryption is used throughout the parameter interaction process. At the same time, identity authentication and permission verification modules are configured for local nodes, and only authorized nodes participate in federated training. Sensitive features such as patient condition and diagnosis results are graded and labeled, and protection intensity ranges are divided according to the severity of the condition and the level of privacy sensitivity. An adaptive noise injection module is embedded in the forward propagation and gradient calculation of local model training. The noise intensity is dynamically adjusted according to the medical scenario and the level of sensitivity. A privacy budget allocation mechanism is established, and the total privacy budget is split according to the training round and node permissions. The privacy consumption progress is monitored in real time and the noise intensity or training status of high-sensitivity features is dynamically adjusted. Non-sensitive business features are selected as distillation inputs, and business information such as disease complexity and treatment level are used as distillation outputs. A knowledge distillation method is adopted, using a teacher model trained on local complete data as a benchmark to guide the student model on desensitized data to learn the feature mapping relationship. Gradient matching and distribution alignment are used to reduce output differences. The distillation model is optimized in reverse by combining feedback data from the quality control system, and the feature mapping weights are updated. A real-time monitoring module is built to collect indicators such as model prediction accuracy, training convergence speed, and privacy leakage risk value. Three adjustment strategies are preset: privacy priority, efficiency priority, and balance mode. The differential privacy noise intensity and federated parameter encryption level are automatically adjusted according to the indicator monitoring results, and a smooth transition strategy is adopted in the adjustment process. The system generates operation logs for the entire process of local data storage, encrypted parameter transmission, noise injection, and feature distillation. It records node information, training rounds, and privacy parameter adjustment records to meet audit requirements. It regularly evaluates the effectiveness of privacy protection and model performance indicators, optimizes noise adjustment strategies, feature distillation weights, and balance adjustment thresholds, and adapts to medical business scenarios and compliance policy updates.

9. The intelligent decision-making method for human resource planning of personnel input as described in claim 8, characterized in that: The aforementioned real-time monitoring module collects indicators such as model prediction accuracy, training convergence speed, and privacy leakage risk value. It presets three adjustment strategies: privacy priority, performance priority, and a balanced mode. Based on the indicator monitoring results, it automatically adjusts the differential privacy noise intensity and the encryption level of the federated parameters. The adjustment process employs a smooth transition strategy, including: The calculation rules for medical scenarios of the three types of indicators are clearly defined. Lightweight collection probes are deployed synchronously at local nodes and the federal coordination center to collect local training-related data and global aggregated related data respectively. The data is summarized and reported to the time-series monitoring library at preset time intervals to form a traceable view of the whole process indicators, and to solidify the data foundation for dynamic adjustment of indicators. Differentiated privacy protection baselines are set for three modes: privacy priority, performance priority, and balance. The privacy priority mode strengthens protection and strictly constrains privacy risks, while the performance priority mode reduces protection and prioritizes model performance. The balance mode uses an intermediate baseline to simultaneously constrain two types of indicators. Each mode is bound to a corresponding noise intensity range, encryption level, and indicator trigger threshold. The system assesses the deviation between the collected indicators and the policy baseline in real time, classifies them into three levels, and automatically matches the direction of protection strength adjustment based on the judgment results of excessive privacy risks or insufficient model performance. If there are no abnormalities, the parameters are kept stable. The target noise intensity is calculated by combining the strategy and indicator deviations. A gradual step size adjustment is adopted, and the step size is dynamically adjusted in conjunction with the privacy budget consumption progress, while prioritizing the protection of highly sensitive features. The noise parameters are synchronized across all nodes to ensure uniformity. The encryption levels are divided into different security strengths and overheads, and the levels are switched step by step according to the adjustment direction. The parameter transmission is ensured to be uninterrupted through a parallel transition mechanism of dual encryption protocols. Exponential weighted filtering is used to smooth the adjustment parameters, and adjustment intervals and amplitude limits are set to lock the core training process and avoid training oscillations caused by frequent parameter switching; After adjustment, continuously monitor the recovery status of indicators, make secondary fine-tuning if necessary, record information throughout the process and write it into the audit log, and regularly optimize relevant parameters to improve the accuracy of the balance between privacy and performance.

10. An intelligent decision-making system for human resource planning involving personnel data entry, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the intelligent decision-making method for human resource planning of personnel as described in any one of claims 1 to 9 by executing the machine-executable instructions.

Images