A cardiac rehabilitation nursing case library intelligent teaching system and method

Abstract

The application discloses a kind of heart rehabilitation nursing case library intelligent teaching system and method, belong to the cross field of medical informatization and artificial intelligence medical education.System contains multi-center federal case library construction, causal knowledge graph, double-engine virtual standardized patient, student portrait and intelligent matching, immersive interactive teaching, counterfactual multidimensional evaluation and self iteration seven big modules.The application breaks through data island and sample amount predicament under the premise of privacy compliance through federal learning architecture, realizes high-fidelity dynamic disease deduction relying on physiological-mental double engine, realizes personalized stratified teaching in combination with three-layer matching mechanism, provides causal level depth evaluation feedback by counterfactual reasoning, and constructs double closed loop self iteration mechanism.The application significantly improves the professionalization and intelligent level of heart rehabilitation nursing teaching, and can efficiently cultivate specialized nursing talents meeting clinical needs.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of medical informatics, artificial intelligence medical education, and digital healthcare. Specifically, it relates to an intelligent teaching system and method for cardiac rehabilitation nursing case database based on multi-center federated learning, causal knowledge graphs, physiological-psychological twin virtual patients, and counterfactual reinforcement learning. It is applicable to the full-cycle standardized training, clinical decision-making thinking training, and job competency assessment of cardiovascular specialist nursing staff. Background Technology

[0002] Cardiac rehabilitation is a core component of the comprehensive management of cardiovascular diseases. Through a comprehensive approach integrating exercise, medication, nutrition, psychological intervention, and risk factor control, it can significantly reduce readmission rates and cardiovascular event mortality, while improving quality of life. Data from the International Federation of Rehabilitation Societies shows that cardiac rehabilitation participation rates in developed countries exceed 60%, while in my country it is currently only around 8%, and less than 12% of tertiary hospitals have independently established cardiac rehabilitation departments. One of the core bottlenecks hindering the development of cardiac rehabilitation in my country is the shortage of professional nursing personnel. Existing nurses generally lack specialized knowledge of cardiac rehabilitation, clinical experience, and standardized operational skills, failing to meet the rapidly growing clinical demand for rehabilitation nursing.

[0003] Traditional cardiac rehabilitation nursing teaching models have four inherent flaws:

[0004] First, teaching resources are highly homogenized, and there is a lack of real clinical cases. Existing training mainly focuses on textbook lectures and standardized cases, with few cases, limited disease coverage, and a general lack of complete time-series data on the entire recovery cycle of patients during hospitalization, outpatient visits, and home care. This makes it difficult to recreate complex comorbidities, complications, and individualized intervention scenarios in clinical practice, resulting in a disconnect between theory and practice for trainees.

[0005] Second, the teaching model lacks personalized adaptation and cannot achieve differentiated instruction. Traditional training adopts a rote learning approach with a uniform pace and content, without considering the differences in trainees' educational background, years of work experience, associate degree foundation, and skill gaps, resulting in a low overall return on investment for training.

[0006] Third, there is a lack of practical training scenarios, resulting in insufficient development of clinical decision-making abilities. Traditional practical training relies on one-on-one guidance between mannequins and instructors, which not only has high teacher costs and limited coverage, but also makes it difficult to simulate complex scenarios such as dynamic changes in patients' conditions and sudden adverse events. As a result, trainees' clinical thinking and emergency response capabilities cannot be effectively trained.

[0007] Fourth, the teaching evaluation system is simplistic and lacks a closed-loop feedback mechanism throughout the entire process. Existing assessments are mostly based on summative theoretical examinations, lacking quantitative assessment methods for soft skills such as operational standardization, clinical decision-making logic, communication skills, and humanistic care. At the same time, the assessment results are only used for grade determination and do not drive the dynamic optimization of teaching content and case resources, making it difficult to form a teaching closed loop of continuous improvement.

[0008] With the application of artificial intelligence technology in nursing education, some intelligent teaching systems have been developed. However, for the field of cardiac rehabilitation nursing, existing technologies still have multi-dimensional technical shortcomings:

[0009] First, the construction of case libraries faces the dual dilemma of "data silos" and "privacy barriers." High-quality teaching cases rely on large-sample real-world data from multiple centers, but the strong privacy attributes of medical data mean that data from different medical institutions cannot be directly shared. The sample size of a single hospital's case library is limited, there is a lack of long-tail complication cases, and most cases are static cross-sectional data, lacking full-cycle time-series evolution characteristics, which cannot support long-term clinical thinking training.

[0010] Secondly, the case interaction is still a "pre-scripted" branch, failing to achieve realistic dynamic physiological simulation. The evolution of the disease in the existing system is mostly based on preset fixed branches, which cannot simulate the temporal changes of the patient's physiological indicators in real time and non-linearly according to the intervention measures of the trainees. It also fails to form a closed loop linkage between the patient's psychological state, compliance and physiological recovery trajectory, making it difficult to reproduce the dynamic decision-making process of "intervention-feedback-re-intervention" in clinical practice.

[0011] Third, the accuracy of intelligent matching is insufficient, and the depth of personalized teaching is limited. Most systems only push cases based on keywords or disease tags, without combining multi-dimensional ability profiles of students to achieve adaptive difficulty and semantic-level association of knowledge points, nor do they build long-term personalized learning paths, thus failing to truly achieve individualized instruction.

[0012] Fourth, teaching assessment remains at the level of "comparing standard answers," lacking in-depth causal feedback. Existing systems can only determine whether a decision is right or wrong, but cannot answer counterfactual questions such as "how would the patient's physiological trajectory evolve if an alternative approach were taken." Trainees cannot deeply understand the potential risks and relative advantages of decisions, and their advanced clinical reasoning skills are not adequately developed.

[0013] Fifth, the knowledge system is static and outdated, lacking a self-iterative mechanism. Existing system cases and knowledge bases are mostly statically pre-built, unable to be automatically updated and optimized based on teaching process data, the latest clinical guidelines, and real cases from multiple centers. As guidelines are updated and clinical technologies develop, this system is prone to becoming outdated.

[0014] Therefore, there is an urgent need for an intelligent teaching system for cardiac rehabilitation nursing that can simultaneously break through data privacy barriers, achieve high-fidelity time-series dynamic extrapolation, provide in-depth causal assessment, support personalized teaching throughout the entire lifecycle, and possess self-iterative capabilities. This system would systematically address many pain points of existing teaching models and significantly improve the quality and efficiency of cardiac rehabilitation nursing talent training. Summary of the Invention

[0015] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent teaching system and method for cardiac rehabilitation nursing case databases based on federated learning and dual twin engines. This invention constructs a full-cycle structured case database through a multi-center federated learning architecture while adhering to privacy and compliance. It combines causal knowledge graphs to build virtual standardized patients driven by a physiological-psychological dual engine. Relying on multi-dimensional student profiles, it achieves three-layer precise matching and adaptive learning path planning. Through counterfactual reinforcement learning, it achieves multi-dimensional in-depth assessment and constructs a dual-closed-loop self-iterative mechanism for knowledge and cases. This systematically improves the specialization, personalization, and intelligence of cardiac rehabilitation nursing education, efficiently cultivating cardiac rehabilitation nursing professionals who meet clinical needs.

[0016] To achieve the above objectives, the present invention provides an intelligent teaching system for cardiac rehabilitation nursing case databases, comprising a multi-center federated case feature extraction and full-cycle case database construction module, a causal-driven cardiac rehabilitation nursing knowledge graph module, a dual-engine multimodal virtual standardized patient generation module, a multi-dimensional student profile and intelligent case matching and push module, an immersive full-cycle scenario-based interactive teaching module, a multi-dimensional teaching evaluation module based on counterfactual reinforcement learning, and a knowledge graph and case database self-iteration module. The modules are interconnected and collaboratively invoked through a data bus.

[0017] 1. Multi-center Federated Case Feature Extraction and Full-Lifecycle Case Library Construction Module

[0018] This module is used to construct a multi-dimensional, full-lifecycle structured case library by integrating real-world data from multiple centers, while ensuring data privacy. The module includes a federated learning feature extraction unit and a structured storage unit for the case library.

[0019] Federated Learning Feature Extraction Unit: Employing a Federated Proximal Optimization (FedProx) architecture, this unit connects local servers across multiple healthcare institutions. Without allowing the original patient data to remain within the local domain, it extracts multidimensional temporal features of cardiac rehabilitation and nursing intervention features, aggregating them to generate a global feature representation. Each local node encodes the features of its heterogeneous local data and only uploads the encrypted model parameter updates to the central server for weighted aggregation, physically preventing patient privacy leaks. The FedProx local loss function is defined as:

[0020]

[0021] in, For the empirical risk of the k-th local node, This is the proximal penalty coefficient, used to limit the local model parameters from deviating too far from the global model, adapting to the characteristics of non-independent and identically distributed data; These are the global model parameters. The central server updates the global model using a weighted aggregation strategy:

[0022]

[0023] Where K is the total number of nodes participating in federated learning. Let n be the number of samples at the k-th node, and n be the total number of samples. This represents the parameter update amount for the k-th node.

[0024] The case library's structured storage unit: Based on global features and local typical cases, a five-dimensional, full-cycle structured case library is constructed. Each case contains five data dimensions: ① Patient basic information layer: includes fields such as demographic characteristics, past medical history, cardiovascular risk factors, and social support; ② Clinical diagnosis and treatment data layer: includes fields such as diagnostic results, auxiliary examination indicators, medication regimens, surgical information, and risk stratification results; ③ Full-cycle rehabilitation care layer: stores time-stamped time-series data such as nursing assessments, nursing problems, nursing interventions, dynamic data of vital signs, patient compliance behavior, and changes in psychological state for each of the three rehabilitation stages: inpatient (Phase I), outpatient (Phase II), and home-based (Phase III); ④ Outcome and follow-up layer: includes long-term outcome data such as rehabilitation effect indicators, adverse event occurrence, readmission status, and quality of life scores; ⑤ Extended tag layer: includes search and matching tags such as case difficulty coefficient, core knowledge point tags, complication types, and key nursing directions.

[0025] The difficulty level of the case study is calculated by weighting five factors and normalizing to a score of 1-5:

[0026]

[0027] in, As a factor for the number of knowledge points, For comorbidity complex factors, As the adverse event quantity factor, The number of decision branches is a factor. This is a factor representing the rate of change in the disease condition.

[0028] 2. Causal-Driven Knowledge Graph Module for Cardiac Rehabilitation Nursing

[0029] This module constructs a multi-level knowledge graph that integrates relational and causal logic, providing a logical foundation for virtual patient evolution, decision-making reasoning, and counterfactual assessment. Graph nodes cover seven categories of entities: disease entities, symptom manifestations, rehabilitation assessment tools, interventions, complication risks, medications, and psychological states. Edge relationships include four categories: temporal associations, causal effects, taboo constraints, and guideline recommendation levels. The module integrates structural causal models and Bayesian inference algorithms to mine potential causal relationships between variables from multi-center anonymized data, calculate the average causal effect (ACE), and supplement it to the knowledge graph, elevating it from a relational graph to a causal graph. For example, it quantifies the causal effect strength of "increased exercise intensity → increased myocardial oxygen consumption → angina attack," providing causal evidence for subsequent physiological evolution and counterfactual reasoning. The graph supports incremental updates and version control, ensuring the traceability of the knowledge system.

[0030] 3. Dual-engine multimodal virtual standardized patient generation module

[0031] This module is the core simulation engine of the system. Based on a dynamic case library and causal knowledge graph, it generates multimodal virtual standardized patients (VSPs) that include physiological evolution engines and behavioral psychology engines, enabling high-fidelity dynamic disease simulation.

[0032] The temporal physiological state evolution engine employs a hybrid modeling architecture driven by physiological mechanisms and data. The white-box layer, based on a lumped-parameter cardiovascular circulation model, describes the dynamic relationships of blood pressure, blood flow, and volume in the systemic-pulmonary circulation through a system of ordinary differential equations. The gray-box layer, based on a time fusion converter (TFT) architecture, corrects the residuals of the physiological model by fitting them, balancing physiological interpretability with individual fitting accuracy. The TFT model input includes a static covariate c and historical temporal physiological data. Intervention vectors of trainees Output the probability distribution of physiological indicators for the next k time steps:

[0033]

[0034] in, and The first The model measures the mean and variance of target physiological indicators (heart rate, blood pressure, myocardial oxygen consumption, BNP, etc.) at any given time. It automatically selects key features through a variable selection network, filters irrelevant noise through a gating mechanism, and captures long-term temporal dependencies through a multi-head attention layer.

[0035] Behavioral Psychology and Emotional Dialogue Engine: Based on a finely tuned large language model in the medical field, a multi-turn emotional dialogue sub-model is constructed. The Prompt template embeds virtual patient psychological profile tags (anxious, depressive, denial, optimistic, etc.) and current physiological discomfort indicators. The output includes interaction results containing natural language responses and emotional state tags, and simultaneously outputs compliance correction coefficients. The compliance correction coefficient will directly affect the execution weight of the exercise prescription in the physiological evolution engine, forming a closed-loop linkage of "communication strategy → psychological state → compliance → physiological rehabilitation trajectory": if the trainee's communication is inappropriate and causes the patient to resist, even if the exercise prescription is reasonable, the actual rehabilitation effect will be discounted due to insufficient execution, which highly restores the real clinical logic.

[0036] 4. Multi-dimensional student profiling and intelligent case matching and recommendation module

[0037] This module is used to build a comprehensive competency profile of learners and to achieve precise adaptive push of case studies and learning path planning through a three-layer matching mechanism.

[0038] The trainee multi-dimensional profile unit comprises five dimensions: basic attribute tags, knowledge mastery tags, operational ability tags, clinical reasoning tags, and learning behavior tags. Each dimension has multiple sub-categories and is quantitatively scored from 0 to 100 points. Basic attributes include static information such as education level, years of work experience, department, and specialist training experience; knowledge mastery covers basic theories of cardiac rehabilitation, key nursing points for various diseases, and complication management; operational ability includes skills such as vital sign monitoring, motion monitoring, and emergency procedures; clinical reasoning includes dimensions such as risk assessment, decision-making logic, and emergency response; and learning behavior includes behavioral characteristics such as learning duration, case completion rate, and error distribution. The profile uses a decay update mechanism, with higher weighting for recent assessment data, and a forgetting coefficient is set to reflect real-world memory patterns. The trainee's comprehensive ability score is normalized to 1-5 points.

[0039]

[0040] in, The average score for knowledge mastery. The average score for clinical reasoning. The average score is based on operational ability.

[0041] The three-layer intelligent matching unit sequentially selects the optimal cases through three mechanisms: rule filtering, difficulty adaptive matching, and semantic precision matching. The first layer, rule filtering, selects candidate cases that match the disease type, rehabilitation stage, and knowledge point range based on the student's current learning stage and training syllabus. The second layer, difficulty adaptive matching, selects a subset of cases where the absolute value of the difference between the case difficulty coefficient and the student's comprehensive ability score is ≤0.8 to ensure difficulty matching. The third layer, semantic precision matching, calculates the semantic correlation between the core knowledge points of the case and the student's weak knowledge points based on a causal knowledge graph, and combines this with a weighted ranking of historical wrong answers to output the top 3-5 optimal matching cases.

[0042] Adaptive learning path planning unit: Based on student profiles and assessment results, it uses deep reinforcement learning algorithms to dynamically plan subsequent learning paths, adjust case difficulty, complication types and intervention scenarios, and achieve gradual ability improvement.

[0043] 5. Immersive, full-cycle interactive teaching module

[0044] This module is used to transform static cases into interactive, full-cycle, contextualized teaching scenarios, including case scenario restoration units, branch decision-making interaction units, dynamic disease evolution units, and real-time prompting and guidance units.

[0045] The case scenario reconstruction unit presents the patient's condition background and nursing scenarios throughout the entire recovery cycle of stages I / II / III in a timeline format. The left side is the patient information panel, and the right side is the operation function panel.

[0046] The branch decision-making interaction unit sets 3-5 decision options at each key nursing node, covering types such as assessment, intervention, communication, and emergency treatment. Complex nodes set up multi-level decision chains.

[0047] The dynamic disease evolution unit calls the dual-engine VSP to extrapolate the patient's physiological indicators, psychological state and disease trend in real time based on the trainee's decisions, and presents the clinical outcomes corresponding to different decisions, including multiple paths such as stable condition, symptom relief, adverse event triggering, and disease deterioration.

[0048] The real-time prompting and guidance unit has a three-level prompting mechanism: level 1 is knowledge point prompting, level 2 is thought process guidance prompting, and level 3 is explicit correction prompting. The prompting level can be automatically adjusted according to the learner's ability.

[0049] 6. Multi-dimensional teaching assessment module based on counterfactual reinforcement learning

[0050] This module is used to conduct a comprehensive quantitative assessment and in-depth causal feedback of the learners' learning process, and to construct a Markov decision process with the virtual patient's physiological state and psychological compliance as the state space and nursing intervention as the action space.

[0051] The five-dimensional basic assessment unit is based on quantitative scoring of five core dimensions: knowledge mastery (weight 20%), operational standardization (weight 20%), clinical decision-making ability (weight 30%), communication and humanistic care ability (weight 15%), and emergency response ability (weight 15%). The weighted average is then used to obtain the comprehensive score.

[0052] Counterfactual depth assessment unit: Performs counterfactual reasoning based on a structural causal model. The specific process is as follows: ① Obtain the trainee's status at key state nodes. actual actions Compared with the actual state trajectory Calculate the actual cumulative reward ② Keeping the background noise variable constant, the action intervention is the optimal action recommended by the guideline. Or common mistakes made by students Using a time-series physiological evolution engine to deduce counterfactual state trajectories Calculate the cumulative counterfactual reward ③ Calculate the counterfactual therapeutic effect (CTE):

[0053]

[0054] The reward function is defined as a multi-dimensional weighted sum:

[0055]

[0056] in, For the degree of improvement in cardiac function, For patient compliance, Punishment for adverse events For violations of the guidelines, These are the corresponding weighting coefficients.

[0057] If the CTE is lower than the preset threshold, a counterfactual warning is triggered, showing trainees a comparison chart of the physiological evolution of the actual trajectory and the counterfactual trajectory, intuitively presenting the clinical differences between different decisions, and helping trainees establish causal decision-making thinking.

[0058] 7. Knowledge Graph and Case Library Self-Iteration Module

[0059] This module is designed to continuously optimize and iterate the case study library and knowledge system, forming a closed-loop teaching system. The module includes a case study quality assessment unit, a knowledge point association and update unit, and a guide synchronization and update unit.

[0060] The case quality assessment unit uses data such as student completion rate, error rate, and teaching effectiveness feedback to quantitatively score the teaching value of cases, eliminating low-quality cases and expanding high-quality cases;

[0061] The knowledge point association update unit automatically optimizes the association mapping between cases and knowledge points based on the distribution of students' wrong questions and learning difficulties, and increases the priority of pushing cases corresponding to weak knowledge points.

[0062] The guideline synchronization and update unit supports importing the latest clinical guidelines and expert consensus, automatically extracting key updates and synchronizing them to the knowledge graph, while also performing compliance verification and batch updates of intervention protocols in the case library. Simultaneously, the federated learning architecture regularly incorporates the latest clinical data from multiple centers, iterating the global feature model and VSP evolution engine to achieve continuous self-evolution of the case library.

[0063] Another aspect of the present invention provides an intelligent teaching method for a cardiac rehabilitation nursing case database, implemented based on the above-mentioned system, comprising the following steps:

[0064] S1. Multi-center federated feature extraction and full-cycle case library construction: Multi-center features are aggregated under the premise of privacy compliance through the FedProx federated learning architecture, and a five-dimensional full-cycle case library is constructed by combining structured processing. At the same time, a causal-driven knowledge graph of cardiac rehabilitation nursing is constructed.

[0065] S2. Initialization of multi-dimensional student profile: Collect basic student information, generate an initial ability profile through an introductory test, and identify the student's knowledge gaps and ability levels.

[0066] S3. Intelligent Case Matching and Learning Path Planning: Based on the student's current profile and learning objectives, suitable cases are pushed through a three-layer matching mechanism, and a phased personalized learning path is generated.

[0067] S4. Immersive full-cycle scenario-based interactive teaching: Trainees participate in the nursing care of patients throughout the entire recovery cycle as responsible nurses, make decisions at key points, and the system uses the dual-engine VSP to simulate the patient's condition in real time and provide graded prompts.

[0068] S5. Multi-dimensional teaching assessment and profile update: After the case study, quantitative scoring is carried out from five basic dimensions, and counterfactual reasoning is performed to generate an in-depth assessment report. The student profile is updated based on the assessment results.

[0069] S6. Knowledge System and Case Library Self-Iteration: Regularly optimize case quality, update knowledge point associations, and synchronize clinical guidelines based on teaching data and multi-center clinical data to achieve continuous improvement of system capabilities.

[0070] Compared with the prior art, the present invention has the following significant technical advancements and beneficial effects:

[0071] 1. Federated Learning Empowers a Full-Cycle Case Library, Overcoming the Dual Dilemmas of Data Privacy and Sample Size. This invention adopts the FedProx federated learning architecture to achieve feature-level fusion of multi-center real-world data without data leaving the domain. This not only strictly protects patient privacy and compliance but also solves the problems of insufficient long-tail cases and insufficient sample size in single hospitals. At the same time, it combines the characteristics of cardiac rehabilitation to construct a five-dimensional full-cycle structured case library, covering the entire process from inpatient to outpatient to home care, highly restoring real clinical scenarios, and achieving a dual improvement in privacy security and the depth of specialized cases.

[0072] 2. Causal Knowledge Graph + Dual-Engine VSP for High-Fidelity Dynamic Temporal Deduction: This invention constructs a causal-driven specialty knowledge graph, breaking through the logical limitations of traditional relational graphs. Simultaneously, it proposes a physiological evolution engine driven by physiological mechanisms and data-driven hybrid modeling, and a behavioral psychology engine driven by emotion computing, achieving closed-loop linkage simulation of "physiology-psychology-behavior." The trajectory of the virtual patient's condition is not a pre-set branch, but rather a real-time deduction based on intervention, realistically reproducing the long-cycle, slow-feedback, and multi-factor coupled clinical nature of cardiac rehabilitation, significantly enhancing the immersion and realism of clinical thinking training.

[0073] 3. Multi-dimensional profiling + three-layer matching + path planning enables deeply personalized tiered teaching. This invention constructs a five-dimensional student ability profile and combines a three-layer matching mechanism of rule filtering, difficulty adaptation, and semantic precision matching to accurately push teaching cases that are highly adapted to students' abilities and knowledge gaps. At the same time, it realizes dynamic planning of long-term learning paths through reinforcement learning, which avoids both learning frustration and waste of efficiency, truly achieves individualized teaching, and significantly improves teaching efficiency and learning experience.

[0074] 4. Five-dimensional assessment + counterfactual reasoning for causal-level in-depth teaching feedback: This invention integrates a five-dimensional quantitative assessment system with a counterfactual reinforcement learning assessment mechanism. It can not only conduct granular assessment of trainees' abilities from multiple dimensions, but also intuitively demonstrate the physiological trajectory differences of different decisions by deduce the clinical outcomes of alternative decisions. This helps trainees understand the causal logic and risk-reward relationship behind decisions, upgrading the traditional "right-wrong assessment" to "causal-level in-depth feedback" and effectively cultivating advanced clinical decision-making abilities.

[0075] 5. Dual-loop self-iterative mechanism to ensure the timeliness of content and continuous evolution of the system. This invention constructs a dual-loop self-iterative mechanism of "teaching feedback-content optimization" and "federated data-model iteration". It can dynamically optimize the association between cases and knowledge points based on teaching process data, and continuously absorb multi-center clinical experience and synchronize with the latest clinical guidelines through federated learning, so as to ensure that the teaching content is always consistent with the clinical frontier and the system capabilities continue to improve with the accumulation of data. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the overall architecture of the intelligent teaching system for cardiac rehabilitation nursing described in this invention;

[0077] Figure 2 This is a flowchart illustrating the multi-center federated learning and full-cycle case library construction process described in this invention.

[0078] Figure 3 This is a flowchart illustrating the multidimensional evaluation logic based on counterfactual reinforcement learning as described in this invention.

[0079] Figure 4 This is a flowchart illustrating the overall steps of the intelligent teaching method for cardiac rehabilitation nursing described in this invention. Detailed Implementation

[0080] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0081] Example 1: Overall System Architecture and Detailed Implementation of Each Module

[0082] like Figure 1 As shown, this embodiment provides an intelligent teaching system for cardiac rehabilitation nursing case databases based on federated learning and dual twin engines. The system adopts a layered architecture, consisting of a data layer, model layer, teaching business layer, and interaction layer from bottom to top. Each core module is deployed on a cloud server and federated edge nodes. Students and teachers access the system via desktop, tablet, or VR terminal. The system specifically includes seven core modules, and the implementation details of each module are as follows:

[0083] 1. Multi-center Federated Case Feature Extraction and Full-Lifecycle Case Library Construction Module

[0084] This module forms the system's data foundation. It achieves compliant data fusion across multiple centers through federated learning, and simultaneously builds a specialized, structured case library. The construction process is as follows: Figure 2As shown, the federated learning component consists of a central parameter server and eight local edge nodes from partner tertiary hospitals. Each node deploys a local training copy of the feature extraction network and the TFT evolution model. Data access covers hospital electronic medical record systems, cardiac rehabilitation management systems, follow-up systems, and wearable device data. All data undergoes local de-identification and anonymization processing, ensuring the original data never leaves the domain. Local training employs the FedProx algorithm, with a proximal penalty coefficient (mu) set to 0.01 to accommodate the Non-IID distribution characteristics of different hospital patient groups. After local training, only the encrypted model parameter gradients are uploaded to the central server. The server weights and aggregates the data according to sample size, updates the global model, and then distributes it to each local node, completing a full iteration every two weeks. Through this mechanism, the system can incorporate clinical case features from hospitals in different regions and at different levels, significantly improving the simulation accuracy of long-tail complication scenarios.

[0085] The case library's structured storage unit is built based on global features and desensitized typical cases provided by various hospitals. It includes 300 pre-set standardized cases covering common cardiac rehabilitation conditions such as post-PCI coronary artery disease, acute myocardial infarction, chronic heart failure, post-valvular heart disease, and hypertensive heart disease. These cases are stratified into low, medium, and high-risk groups, and include complex cases with multiple comorbidities such as diabetes, COPD, and renal insufficiency. Each case is stored strictly according to a five-dimensional structure.

[0086] (1) Patient basic information layer: includes demographic fields such as gender, age, height and weight, BMI, occupation, education level, and social support score, as well as risk factor fields such as history of comorbidities such as hypertension and diabetes, and smoking and drinking;

[0087] (2) Clinical diagnosis and treatment data layer: includes primary and secondary diagnoses, coronary angiography results, surgical information, echocardiography EF value, BNP, blood lipids and blood glucose and other key examination indicators, medication regimens, GRACE risk stratification results, etc.

[0088] (3) Full-cycle rehabilitation nursing layer: Time-series nursing data is stored in stages according to I / II / III stages, with each time node marked with a corresponding timestamp, including a complete nursing timeline such as daily nursing assessment, nursing problems, intervention measures, vital signs records, patient complaints, compliance records, etc.

[0089] (4) Outcomes and follow-up layer: including discharge indicators such as 6-minute walk test distance, SF-36 quality of life score, HADS anxiety and depression score, as well as follow-up data and adverse event records at 1, 3, 6 and 12 months after discharge;

[0090] (5) Extended tag layer: includes difficulty level 1-5, core knowledge point tags, nursing key points tags, complication tags, and applicable training stage tags, which are used for case retrieval and intelligent matching.

[0091] 2. Causal-Driven Knowledge Graph Module for Cardiac Rehabilitation Nursing

[0092] This module uses the Neo4j graph database for storage, containing 7 major entity categories and 28 entity relationships, covering 126 tertiary specialty knowledge points. The graph construction is divided into two steps: First, 5 cardiac rehabilitation nursing experts construct a prior causal graph, defining known causal relationships between core variables, such as "increased blood volume → worsening heart failure" and "β-blockers → decreased heart rate." Second, the causal reasoning submodule uses a PC algorithm to perform causal discovery on multi-center desensitized data, mining potential causal relationships and calculating average causal effects, which are then added to the graph after expert review. For example, the algorithm mines a chain causal relationship from the data: "high preoperative anxiety score → low postoperative exercise compliance → prolonged Phase I rehabilitation time," and quantifies the strength of the causal effect at each stage. This causal relationship will simultaneously affect the VSP's behavioral psychology engine and physiological evolution engine, making the psychological-behavioral-physiological linkage of virtual patients more consistent with real clinical patterns. The graph supports version snapshot management, generating a new version after each guideline update or model iteration, supporting historical version retrospection and comparative teaching.

[0093] 3. Dual-engine multimodal virtual standardized patient generation module

[0094] This module is the core simulation engine of the system, generating and interacting with timing diagrams. Each virtual patient instance contains physiological state vectors and psychological state vectors, driven by two separate engines and linked bidirectionally.

[0095] The temporal physiological state evolution engine employs a hybrid modeling architecture. The white-box component is a 5-chamber lumped-parameter cardiovascular circulation model, describing hemodynamic changes in the left and right ventricles and systemic-pulmonary circulation through 12 sets of ordinary differential equations. The gray-box component is a TFT temporal prediction model, with input dimensions including 24 static covariates, 36 temporal physiological indicators, and 12 intervention action features. It outputs the probability distributions of 8 core indicators, including heart rate, systolic blood pressure, myocardial oxygen consumption, and BNP, for the next 24 hours. The model automatically filters key influencing factors through a variable selection network, extracts nonlinear features through a gated residual network, and captures long-term temporal dependencies through multi-head attention. During runtime, the white-box model outputs the physiological baseline response, and the gray-box model outputs the individual bias correction; the two are superimposed to obtain the final physiological evolution result. This architecture ensures that the evolution result conforms to basic physiological laws while accurately fitting individual differences among patients, achieving simulation accuracy superior to single data-driven models. When physiological indicators exceed preset safety thresholds, adverse events such as arrhythmia, angina pectoris, and hypotension are automatically triggered, initiating an emergency response scenario.

[0096] Behavioral Psychology and Emotional Dialogue Engine: Based on the Qwen-7B large language model, fine-tuned for medical psychology, the training corpus includes 5000 samples of patient-nurse dialogues in cardiac rehabilitation and psychological intervention cases. The Prompt template incorporates three types of contextual variables: patient psychological profile, current physiological state, and disease stage. The output includes three parts: patient's natural language response, emotional state label, and compliance correction coefficient. The compliance correction coefficient directly affects the weighting of the physiological engine's exercise prescription execution rate, for example... This indicates a high level of patient cooperation, with an exercise prescription adherence rate exceeding 95%. Patients who resisted the treatment often only achieved 50% of the prescribed exercise intensity, resulting in a reduced rehabilitation effect. This mechanism creates a quantifiable closed loop between humanistic communication skills and physiological rehabilitation outcomes, addressing the shortcomings of traditional teaching methods that emphasize operation over communication.

[0097] 4. Multi-dimensional student profiling and intelligent case matching and recommendation module

[0098] This module is responsible for building learner competency profiles and delivering precise case studies. The learner profile includes 5 primary dimensions, 18 secondary tags, and 126 tertiary knowledge point tags, all scored from 0 to 100. The initial profile is generated through an introductory test, including 30 theoretical questions and 2 simple case analysis questions. Subsequent profiles are dynamically updated based on assessment results after each case study session. Updates employ a time decay mechanism, with assessment data more recent to the current time carrying higher weight; the decay coefficient is set to 0.95 per month. A forgetting factor is also included: if a corresponding knowledge point is not studied for 30 consecutive days, the mastery score automatically decreases by 5%, aligning with human memory patterns.

[0099] The intelligent case matching system employs a three-layer mechanism. Taking a beginner student with a comprehensive ability score of 2.4 selecting the topic "Post-PCI Phase I Rehabilitation Nursing" as an example: The first layer is rule filtering: 36 candidate cases are selected based on the disease being post-PCI coronary heart disease, the rehabilitation stage being Phase I, and the applicable stage being beginner. The second layer is adaptive difficulty matching: 21 cases are selected based on the difficulty coefficient between 1.6 and 3.2. The third layer is precise semantic matching: The student's weak knowledge points are extracted as "post-operative complication identification" and "symptom management during exercise." The semantic similarity between the knowledge points and weak points of the candidate cases is calculated based on the knowledge graph. Combined with the weighted ranking of historical incorrect answers, the top 3 ranked cases are finally pushed to the student for selection.

[0100] The adaptive learning path planning unit uses the Deep Q-Network (DQN) algorithm, taking the learner's ability status as input, the case recommendation sequence as the action space, and the ability improvement efficiency as the reward function, to dynamically plan weekly and monthly learning paths for learners, achieving gradual ability advancement.

[0101] 5. Immersive, full-cycle interactive teaching module

[0102] This module transforms case studies into interactive teaching scenarios, supporting both desktop and VR presentation modes. The desktop version uses a 2D interface with a timeline, while the VR version presents a 3D immersive virtual ward. After entering a case study, learners first view the imported case and basic patient information. The teaching progresses step-by-step according to the rehabilitation stage timeline. Each key nursing node presents a decision-making question with 3-5 options, categorized as correct, partially correct, incorrect, and high-risk incorrect. After a learner selects an option, the system doesn't immediately indicate right or wrong; instead, it uses the dual-engine VSP to simulate the subsequent condition, leading to the next scenario and the final decision. For example, in the ward walking training scenario on the third postoperative day, if a learner starts walking without assessing vital signs and without setting monitoring frequency, the VSP will simulate adverse events such as chest tightness, chest pain, decreased blood pressure, and rapid heart rate after 5 minutes of exercise, triggering an emergency response. If the learner assesses first, gradually increases intensity, and monitors regularly according to guidelines, the patient's condition stabilizes, and rehabilitation progresses smoothly. The system features a three-tiered prompting mechanism, which students can use independently or the system can automatically adapt to their abilities: Level 1 prompts only provide the knowledge point name, Level 2 prompts offer directions for thought, and Level 3 prompts clearly state the key operational points. More capable students receive lower-level prompts by default, encouraging independent thinking.

[0103] 6. Multi-dimensional teaching assessment module based on counterfactual reinforcement learning

[0104] This module conducts a multi-dimensional assessment based on all student interaction data. The basic assessment includes five dimensions: knowledge mastery, scored by the accuracy of answering knowledge-related questions and the accuracy of knowledge application; operational standardization, scored by the completeness of steps, the correctness of the sequence, and compliance with details; clinical decision-making ability, scored by the similarity between the decision path and the optimal path, the accuracy of problem identification, and the appropriateness of intervention timing; communication ability, scored by the completeness of health education, the degree of humanistic care, and the effectiveness of patient compliance guidance; and emergency response ability, scored by the speed of adverse event identification, the correctness of the handling process, and the effectiveness of risk prevention and control. The five dimensions are weighted to obtain a comprehensive score.

[0105] The core innovation of this module is the in-depth counterfactual assessment, which involves performing counterfactual reasoning on 3-5 key decision points in the case study. Taking the decision point of "experiencing chest tightness during exercise" as an example:

[0106] Record the trainee's actual decision: immediately stop exercising, take nitroglycerin sublingually, and rest in a supine position. Corresponding actual trajectory: the patient's symptoms were relieved after 10 minutes, and no serious adverse events occurred. Actual cumulative reward. ;

[0107] Constructing Counterfactual Scenario 1 (Guideline Optimal Action): Stop exercise, administer oxygen, monitor vital signs, administer nitroglycerin sublingually, and extrapolate the counterfactual trajectory: The patient's symptoms subside after 6 minutes, myocardial injury markers do not significantly increase, and the counterfactual cumulative reward... Calculated This indicates that there is room for improvement in the trainees' decision-making process;

[0108] Constructing a counterfactual scenario 2 (common incorrect actions): Instruct the patient to continue walking slowly and completing the training volume, and extrapolate the counterfactual trajectory: After 3 minutes, the patient's chest pain worsens, and premature ventricular contractions occur, requiring emergency resuscitation. This is the counterfactual cumulative reward. Calculated This demonstrates that the trainees avoided serious risks in their decision-making.

[0109] The assessment report displays side-by-side comparison curves of physiological indicators for the actual trajectory and the two counterfactual trajectories, visually presenting the clinical differences between different decisions. It also marks cognitive bias points to help trainees deeply understand the causal logic and risk-reward of decision-making.

[0110] 7. Knowledge Graph and Case Library Self-Iteration Module

[0111] This module automatically runs an iterative optimization once a month:

[0112] Case quality assessment: Teaching value score is calculated based on indicators such as case completion rate, average score, student negative review rate, and knowledge point coverage. Cases with a score below 60 are marked as needing optimization and will be modified or eliminated after review by the administrator; the top 20% of cases are marked as classic cases and given priority for recommendation.

[0113] Knowledge point association optimization: Calculate the knowledge point error rate of all students. For knowledge points with an error rate higher than 60% and fewer than 3 corresponding cases, generate supplementary case prompts and automatically increase the weight of the knowledge point in the matching algorithm.

[0114] Guideline updates are synchronized: The system supports importing new versions of clinical guideline documents. Using natural language processing technology, it automatically extracts key updates, compares them with existing knowledge graphs and case content, identifies interventions that do not conform to the latest guidelines, generates an updated list, and then makes batch corrections after expert review. Simultaneously, federated learning completes a global model iteration every two weeks, continuously incorporating the latest multi-center clinical data to optimize VSP evolution accuracy and case coverage.

[0115] Example 2: The Complete Process of Intelligent Teaching Methods

[0116] This embodiment, based on the above system, uses the training process of a newly hired cardiology nurse as an example to explain the complete teaching process in detail:

[0117] Step S1: Initialization of the Case Library and Knowledge Graph Construction System. After the system deployment is complete, it connects to the federated nodes of 8 partner hospitals and completes the first global feature aggregation using the FedProx algorithm. Simultaneously, it imports 250 real de-identified cases and 50 expert-edited typical cases to construct an initial full-cycle structured case library of 300 cases. A causal-driven cardiac rehabilitation nursing knowledge graph is also constructed concurrently, containing 5 primary knowledge points, 28 secondary knowledge points, and 126 tertiary knowledge points, with core knowledge point tags labeled for all cases.

[0118] Step S2: Student Registration and Initial Profile Generation. Student Li, a bachelor's degree graduate, is a newly hired nurse in the cardiology department, with 8 months of work experience and no prior training in cardiac rehabilitation. After initial login, she fills in basic information and completes an introductory test. The system generates an initial competency profile: average knowledge mastery score of 46, including 40 for heart failure nursing and 36 for complication management; initial operational ability score of 43; initial clinical reasoning score of 39; and a comprehensive ability score of 2.2, classifying her as an entry-level student.

[0119] Step S3: Intelligent Case Matching and Learning Path Planning. When a student selects the learning topic "Post-PCI Phase I Rehabilitation Nursing," the system activates a three-layer matching mechanism, ultimately pushing three suitable cases: "Post-PCI Phase I Rehabilitation Nursing for Low-Risk Patients (No Complications)," "Post-PCI Phase I Rehabilitation Nursing for Patients with Mild Hypertension," and "Post-PCI Bedridden Abdominal Distension Nursing Case." Simultaneously, the system generates a learning path for the first month: starting with low-risk cases without complications, gradually transitioning to cases involving complications and comorbidities, progressively improving skills.

[0120] Step S4: Immersive Full-Cycle Scenario Interactive Teaching. The learner selects the first case to begin learning, covering a 5-day timeline from post-operative day 1 to discharge. Post-operative day 1 involves bed rest assessment and guidance on bedside activities; day 2 involves bedside sitting training; day 3 involves ward walking training; and days 4-5 involve increasing activity levels and discharge guidance. During the learning process, in the day 3 walking training scenario, the learner did not assess the patient's resting heart rate beforehand and began walking. The VSP simulation showed that the patient experienced chest tightness and a heart rate rising to 108 beats / min after 4 minutes of exercise. The learner immediately stopped the activity and assisted the patient to lie flat, and the symptoms gradually subsided. The learner used only two level-one knowledge point prompts throughout the process and did not use advanced prompts. Final outcome: The patient was discharged on day 5 post-operatively, having walked 310 meters in 6 minutes, with no serious complications.

[0121] Step S5: Multi-dimensional Assessment and Profile Update. After the case is completed, the system generates an assessment report: overall score of 59 points, including 62 points for knowledge mastery, 56 points for operational standardization, 54 points for clinical decision-making ability, 68 points for communication ability, and 58 points for emergency response ability. The report points out the core weaknesses: non-standard pre-exercise assessment process and insufficient control over the timing of handling discomfort symptoms during exercise. The counterfactual assessment section presents two contrasting scenarios: Counterfactual 1 (exercise after standardized assessment): the patient has no discomfort, recovery is smoother, and the score increases by 8 points; Counterfactual 2 (continued forced exercise): angina pectoris is induced, requiring emergency treatment, and the score decreases by 32 points. Based on the assessment results, the system updates the trainee profile: mastery of post-PCI stage I nursing knowledge points increases to 56 points, clinical reasoning score increases to 44 points, and comprehensive ability score increases to 2.5 points. At the same time, "standardized pre-exercise assessment" is marked as a key weak knowledge point.

[0122] Step S6: Knowledge Iteration and Path Optimization. After the system had been running for a full month, the self-iteration module found that the overall error rate for the knowledge point "Management of Hypotension During Exercise" was 65%, with only 4 corresponding cases. Therefore, optimization suggestions were generated to supplement it with 2 introductory-level cases. Simultaneously, the new version of the "Evidence-Based Practice Guidelines for Coronary Artery Disease Rehabilitation in China" was released. After the administrator imported it, the system automatically identified 6 cases where the exercise prescription intensity recommendations were inappropriate. After expert review, batch updates were completed. Based on the learners' progress in the first month, the system automatically adjusted the learning path for the following month, increasing the proportion of cases related to complication management to specifically address weaknesses.

[0123] Example 3: Application of Standardized Assessment Scenarios

[0124] This system is also applicable to the standardized assessment of cardiac rehabilitation nurses. In the assessment mode, the instructor specifies the difficulty level and knowledge scope, and the system randomly selects matching cases to form the assessment paper. Hints are disabled throughout the process, and the system records the trainees' decision-making and thinking time, increasing the assessment weight of emergency response capabilities. After the assessment, the system automatically generates an assessment report, including an overall score, scores for each dimension, detailed error analysis, and a counterfactual depth report, which can be directly used as a basis for training assessment. Assessment data is synchronously updated to trainee profiles, promoting learning through assessment. It also supports the generation of overall class learning reports to identify common weaknesses and assist instructors in optimizing offline teaching focus.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent teaching system for cardiac rehabilitation nursing case database, characterized in that: It includes a multi-center federated case feature extraction and full-cycle case library construction module, a causal-driven cardiac rehabilitation nursing knowledge graph module, a dual-engine multimodal virtual standardized patient generation module, a multi-dimensional student profile and intelligent case matching and push module, an immersive full-cycle scenario interactive teaching module, a multi-dimensional teaching evaluation module based on counterfactual reinforcement learning, and a knowledge graph and case library self-iteration module. The modules are interconnected and collaboratively invoked through a data bus. The multi-center federated case feature extraction and full-cycle case library construction module is used to aggregate multi-center clinical features through federated learning and construct a multi-dimensional, full-cycle structured cardiac rehabilitation nursing case library without leaving the original patient data domain. The causal-driven cardiac rehabilitation nursing knowledge graph module is used to construct a cardiac rehabilitation specialty knowledge graph that combines correlation and causal logic, providing a logical basis for virtual patient evolution, decision reasoning, and counterfactual assessment. The dual-engine multimodal virtual standardized patient generation module is used to generate virtual standardized patients that include a temporal physiological state evolution engine and a behavioral psychological and emotional dialogue engine, so as to realize dynamic disease prediction with physiological-psychological linkage. The student multi-dimensional profile and intelligent case matching and push module is used to construct a student multi-dimensional ability profile and push suitable teaching cases and plan personalized learning paths through a multi-layer matching mechanism. The immersive full-cycle scenario-based interactive teaching module is used to transform cases into interactive scenario-based teaching scenarios covering the entire rehabilitation cycle, receive student decision input, and provide real-time graded prompts. The multi-dimensional teaching evaluation module based on counterfactual reinforcement learning is used to conduct multi-dimensional quantitative evaluation of the student's learning process and generate causal-level in-depth teaching feedback through counterfactual reasoning. The knowledge graph and case library self-iteration module is used to continuously optimize and iterate the case library and knowledge system based on teaching process data and the latest clinical guidelines.

2. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The multi-center federated case feature extraction and full-cycle case library construction module includes a federated learning feature extraction unit and a case library structured storage unit; the federated learning feature extraction unit adopts a federated proximal optimization algorithm architecture, and the local loss function is: in For the empirical risk of the k-th local node, This is the proximal penalty coefficient. These are global model parameters; the central server updates the global model using a weighted aggregation strategy. Where K is the total number of participating nodes. Let n be the number of samples at the k-th node, and n be the total number of samples. The parameter update amount for the k-th node; the case library structured storage unit constructs a five-dimensional structured case library, each case includes a patient basic information layer, a clinical diagnosis and treatment data layer, a rehabilitation full-cycle nursing layer, an outcome and follow-up layer, and an extended tag layer; the case difficulty coefficient is calculated by weighting five factors: the number of knowledge points, the complexity of comorbidities, the number of adverse events, the number of decision branches, and the rate of change of the condition, and normalized to 1-5 points.

3. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The causal-driven cardiac rehabilitation nursing knowledge graph module's nodes cover seven categories of entities: disease entities, symptom manifestations, rehabilitation assessment tools, intervention measures, complication risks, medications, and psychological states. The edge relationships include four categories: temporal association, causal effect, contraindication constraints, and guideline recommendation levels. The module integrates a structural causal model and a Bayesian inference algorithm to mine potential causal relationships between variables from multi-center desensitized data and calculate the average causal effect, upgrading the graph from an association graph to a causal graph, and supporting incremental updates and version control.

4. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The dual-engine multimodal virtual standardized patient generation module includes a temporal physiological state evolution engine and a behavioral psychology and emotion dialogue engine. The temporal physiological state evolution engine adopts a physiological mechanism-data-driven hybrid modeling architecture. The white-box layer is a lumped parameter cardiovascular circulation model, and the gray-box layer is based on a time fusion converter architecture. The input includes static covariates, historical temporal physiological data, and student intervention action vectors, and the output is the probability distribution of physiological indicators for multiple future time steps. The behavioral psychology and emotion dialogue engine is built based on a fine-tuned large language model in the medical field, and the output includes natural language responses, emotional state labels, and compliance correction coefficients. The compliance correction coefficient directly affects the execution weight of the exercise prescription in the physiological evolution engine, forming a closed-loop linkage between communication strategy, psychological state, compliance, and physiological rehabilitation trajectory.

5. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The multi-dimensional student profile and intelligent case matching and push module includes a multi-dimensional student profile unit, a three-layer intelligent matching unit, and an adaptive learning path planning unit. The multi-dimensional student profile unit constructs a quantitative profile encompassing five dimensions: basic attributes, knowledge mastery, operational ability, clinical reasoning, and learning behavior. The student's comprehensive ability score is normalized to 1-5 points, calculated using the following formula: in The average score for knowledge mastery. The average score for clinical reasoning. The average score is calculated based on operational ability. The three-layer intelligent matching unit sequentially filters cases through three mechanisms: rule filtering, difficulty adaptive matching, and semantic precise matching. Among them, the difficulty adaptive matching filters a subset of cases where the absolute value of the difference between the difficulty coefficient of the case and the student's comprehensive ability score is ≤0.

8. The adaptive learning path planning unit dynamically plans the student's subsequent learning path based on a deep reinforcement learning algorithm.

6. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The multi-dimensional teaching assessment module based on counterfactual reinforcement learning includes a five-dimensional basic assessment unit and a counterfactual in-depth assessment unit. The five-dimensional basic assessment unit calculates a comprehensive score by weighting five dimensions: knowledge mastery, operational standardization, clinical decision-making ability, communication and humanistic care ability, and emergency response ability. The counterfactual depth assessment unit performs counterfactual reasoning based on a structural causal model to obtain the actual cumulative rewards corresponding to the trainees' actual actions. Deducing the cumulative counterfactual reward corresponding to counterfactual actions Calculate the counterfactual therapeutic effect: The reward function is a weighted sum of cardiac function improvement, patient compliance, adverse event penalty, and guideline violation penalty; if the CTE is lower than the preset threshold, a counterfactual warning is triggered and a physiological evolution comparison chart is displayed.

7. The intelligent teaching system for cardiac rehabilitation nursing case database according to claim 1, characterized in that: The knowledge graph and case library self-iterative module includes a case quality assessment unit, a knowledge point association update unit, and a guide synchronization update unit; the case quality assessment unit quantifies the teaching value of cases based on teaching data, eliminates low-quality cases, and expands high-quality cases. The knowledge point association update unit optimizes the association mapping between cases and knowledge points based on the distribution of students' wrong questions, and improves the priority of pushing cases corresponding to weak knowledge points; the guideline synchronization update unit supports the import of clinical guidelines and automatically extracts the key points of updates, synchronizes them to the knowledge graph, and verifies the intervention plan of the updated case library; the module combines the periodic model iteration of federated learning to achieve a double closed-loop self-evolution of the case library and the knowledge system.

8. An intelligent teaching method for cardiac rehabilitation nursing case databases, characterized in that, The system implementation based on any one of claims 1-7 includes the following steps: S1. Multi-center federated feature extraction and full-cycle case library construction: Through a federated learning architecture, multi-center features are aggregated under the premise of privacy compliance to construct a five-dimensional full-cycle case library and simultaneously construct a causal-driven knowledge graph for cardiac rehabilitation nursing. S2. Initialization of multi-dimensional student profile: Collect basic student information, generate an initial ability profile through an introductory test, and identify the student's knowledge gaps and ability levels. S3. Intelligent Case Matching and Learning Path Planning: Based on the student's current profile and learning objectives, suitable cases are pushed through a multi-layer matching mechanism, and a phased personalized learning path is generated. S4. Immersive full-cycle scenario-based interactive teaching: Trainees participate in the nursing care of patients throughout the entire recovery cycle as responsible nurses, make decisions at key points, and the system simulates the condition of patients in real time through dual-engine virtual standardized patients and provides graded prompts. S5. Multi-dimensional teaching assessment and profile update: After the case study, quantitative scoring is carried out from multiple basic dimensions, and counterfactual reasoning is performed to generate an in-depth assessment report. The student profile is updated based on the assessment results. S6. Knowledge System and Case Library Self-Iteration: Regularly optimize case quality, update knowledge point associations, and synchronize clinical guidelines based on teaching data and multi-center clinical data to achieve continuous improvement of system capabilities.

9. The intelligent teaching method for cardiac rehabilitation nursing case database according to claim 8, characterized in that, In step S4, the deduction process of the dual-engine virtual standardized patient is as follows: the system encodes the trainee's structured intervention instructions into action vectors and inputs them into the temporal physiological evolution engine, while inputting the trainee's unstructured communication scripts into the behavioral psychological and emotional engine to generate compliance correction coefficients. The compliance correction coefficient acts on the exercise prescription execution weight of the physiological evolution engine, jointly driving the dynamic evolution of the patient's physiological indicators and psychological state. When the physiological indicators exceed the safety threshold, an adverse event scenario is triggered.

10. The intelligent teaching method for cardiac rehabilitation nursing case database according to claim 8, characterized in that, In step S5, the specific process of counterfactual reasoning is as follows: extract the student's actual decision and corresponding actual state trajectory for key state nodes, and calculate the actual cumulative reward; under the condition of keeping the background noise variable unchanged, replace the action with the optimal action recommended by the guide or common error-prone actions, deduce the counterfactual state trajectory and calculate the counterfactual cumulative reward; quantify the relative advantage and potential risk of the decision through the difference between the two, and generate a visual comparison result.

Images