A gynecological tumor rehabilitation intervention scheme dynamic generation method and system

Abstract

The application belongs to the technical field of gynecological tumor rehabilitation, and discloses a gynecological tumor rehabilitation intervention scheme dynamic generation method and system, which integrates electronic medical record clinical data, wearable physiological behavior data, patient self-report form data and gynecological tumor special evaluation data, completes time sequence alignment and standardized mapping with key treatment events such as operation day, first day of each cycle chemotherapy, radiotherapy start day, first day of afterloading treatment as time anchor points, divides multi-scale rehabilitation time windows and constructs full-cycle longitudinal data sets; combined with prior constraints in the field of gynecological tumor rehabilitation, a causal discovery algorithm is used to construct the causal directed acyclic graph corresponding to each time window, to mine the causal transmission relationship among symptoms, physiological indicators, treatment adverse reactions and behavior intervention, to identify the pre-causal markers and three-level transmission links of the latent period complications, and to output the core intervention target points and the latent period pre-intervention nodes.

Description

Technical Field

[0001] This invention belongs to the field of gynecological tumor rehabilitation technology, specifically a method and system for dynamically generating gynecological tumor rehabilitation intervention programs. Background Technology

[0002] The rehabilitation needs of gynecological cancer patients are characterized by multi-dimensionality, stages, and strong heterogeneity. Cervical cancer, as a common malignant tumor of the female reproductive system, is prone to complications such as lower extremity lymphedema, pelvic floor dysfunction, and radiation damage after treatment such as extensive hysterectomy, pelvic lymph node dissection, and concurrent chemoradiotherapy. The rehabilitation cycle is long and there are significant individual differences.

[0003] Currently, some digital rehabilitation intervention program generation methods have been applied in the field of cancer rehabilitation, but most of them adopt a static generation logic of "patient stratification-guideline matching," which has the following technical problems: Existing methods mostly match intervention plans with cross-sectional features of a single assessment, failing to explore the causal relationships of symptoms, physiological indicators, treatment events, and behavioral data in the longitudinal course of the patient's disease. They cannot distinguish the causal transmission and accompanying associations between symptoms, making it difficult to carry out pre-intervention based on antecedent factors, and are prone to misjudgment of intervention targets and misallocation of resources.

[0004] Existing methods only make simple stratified adjustments based on the group standard dose of clinical guidelines, without establishing an individual intervention dose-outcome response model. This makes it difficult to adapt to the heterogeneous differences in patients' surgical scope, treatment plan, and basic physical fitness, and it is difficult to achieve individualized titration of intervention dose. There is a risk of causing pelvic floor damage and aggravating lymphedema.

[0005] Gynecologic oncology rehabilitation covers multiple intervention domains, including symptom management, exercise rehabilitation, and nutritional support. Existing methods are mostly modular splicing of standardized programs in each domain. They lack cross-domain conflict detection and collaborative optimization mechanisms, making it impossible to identify explicit and implicit conflicts between intervention combinations or quantify synergistic gain effects. This can easily lead to decreased program effectiveness or even safety risks.

[0006] Existing methods are mostly based on matching recommendations with historical data of similar patients, which cannot perform counterfactual outcome extrapolation based on individual patient characteristics, making it difficult to quantify the differences in long-term rehabilitation outcomes corresponding to different intervention combinations, and the selection of treatment plans lacks sufficient evidence-based support. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for dynamically generating gynecological tumor rehabilitation intervention programs, so as to solve one or more problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamically generating gynecological tumor rehabilitation intervention programs, comprising the following specific steps: Furthermore, it integrates clinical data from electronic medical records, physiological and behavioral data collected by wearable devices, patient-reported symptom scales, psychological scales, dietary records, and gynecological tumor-specific assessment data such as lymphedema grading, pelvic floor function scores, pelvic pain scores, and reproductive function scores. Using key treatment events such as the day of surgery, the first day of each cycle of chemotherapy, the start day of radiotherapy, and the first day of brachytherapy as time anchors, the data is time-series aligned and standardized, multi-scale rehabilitation time windows are divided, and a longitudinal dataset covering the entire rehabilitation cycle is constructed. Combining prior constraints from the field of gynecologic oncology rehabilitation, a causal discovery algorithm is used to construct causal directed acyclic graphs corresponding to each time window, explore the causal transmission relationships between symptoms, physiological indicators, adverse treatment reactions, and behavioral interventions, identify the antecedent causal markers and three-level transmission links of latent complications, and output individual causal directed acyclic graphs, core intervention targets, indirect influence paths, and pre-intervention nodes in the latent period.

[0009] Furthermore, using a Bayesian hierarchical modeling framework, based on real-world data from gynecologic oncology rehabilitation, a priori model of group dose-effect under different rehabilitation scenarios, such as the postoperative acute phase, radiotherapy and chemotherapy phase, pelvic floor rehabilitation phase, follow-up phase, and endocrine therapy phase, was trained. Specific safety dose constraint rules for gynecologic oncology were embedded to quantify the nonlinear correlation between intervention dimensions such as exercise intensity, pelvic floor training frequency, nutritional intake level, psychological intervention frequency, and symptom management plan dosage and corresponding outcome indicators. For individual patients, based on the population dose-effect prior model, the individual posterior distribution is dynamically updated by combining the patient's baseline characteristics and longitudinal rehabilitation data. The corresponding bi-objective dose-effect surface for each intervention dimension is fitted and generated to determine the expected benefits, adverse reaction risks and individual tolerance boundaries for different intervention doses.

[0010] Furthermore, based on the patient's treatment cycle and rehabilitation assessment results, the core rehabilitation goals for the current stage are matched, and the overall goal is broken down into sub-goals in five gynecological oncology-specific intervention domains: symptom management, pelvic floor function rehabilitation, lymphedema prevention and control, nutritional support, and psychological and reproductive health protection. By combining the core intervention targets and latent pre-intervention nodes output by the individual causal directed acyclic graph, and calling the individual benefit-risk dual-objective dose-response surface, candidate intervention plans containing parameters such as intervention content, implementation frequency, intensity level, and execution method are generated for each intervention domain. Patient preference weight parameters are incorporated into the plan generation process. Based on individual causal directed acyclic graphs, causal mediation effects of candidate interventions are verified to confirm the effectiveness of interventions in causal transmission of rehabilitation outcomes through expected mediating variables. Candidates with spurious associations or target misalignments are eliminated, and a set of effective candidate interventions is output.

[0011] Furthermore, a three-level hierarchical gynecological rehabilitation conflict knowledge graph is pre-constructed, which covers three types of rules: explicit absolute conflict, implicit conditional conflict, and temporal phase conflict, and annotates the synergistic gain coefficients of different intervention combinations. After receiving the candidate solution set verified by causal mediation effect, conflict reasoning detection is performed based on a three-level hierarchical gynecological rehabilitation conflict knowledge graph to identify antagonistic terms and risk points in the solutions; A constrained multi-objective optimization algorithm is adopted, with the optimization objectives of maximizing rehabilitation outcomes, minimizing conflict risks, and minimizing patient executive burden. The objective weights are adjusted according to the patient's rehabilitation stage. The parameters of each domain intervention plan are coordinated and adjusted through parameter adjustment, time staggering, and execution unit splitting, and an intervention plan set is output.

[0012] Furthermore, a structural causal counterfactual reasoning framework was constructed based on individual causal directed acyclic graphs and individual benefit-risk dual-objective dose-response surfaces. Counterfactual inferences were performed on each candidate protocol in the protocol set at weekly, monthly, and quarterly time scales to quantify and predict the outcome indicators for the corresponding period after protocol implementation, including the core symptom relief rate, risk of lower extremity lymphedema, pelvic floor function recovery score, risk of adverse reactions, changes in physical condition, and degree of improvement in gynecological tumor-specific indicators. At the same time, the expected patient compliance level under different protocol parameters was predicted. We constructed a five-dimensional evaluation system that includes clinical benefit, safety, implementation difficulty, expected compliance, and patient preference matching. We determined the weight of each dimension using the entropy weight method, comprehensively ranked all candidate interventions, output the top-ranked intervention, and generated the corresponding outcome report.

[0013] Furthermore, real-time collection of patient behavioral data, combined with the planned execution standards of the intervention program, and the use of a hidden Markov model to assess the patient's true compliance status; when low compliance is identified, the core reasons for the decline in compliance are identified by combining concurrent symptom scores, intervention dose tolerance data and behavioral characteristics, and are classified into three typical scenarios: symptoms worsening leading to execution difficulties, the intensity of the program being too high and exceeding tolerance, and insufficient subjective willingness. For different reasons that trigger corresponding adaptation strategies, when the symptoms worsen and make execution difficult, the intensity of the corresponding intervention domain is reduced and symptom management intervention is added. When the intensity of the plan is too high and exceeds the tolerance, the dose-response curve based on the individual benefit-risk dual objective is reduced to the suboptimal tolerable dose and the execution unit is split. When subjective willingness is insufficient, positive incentive content is added and the intervention form is switched. The execution data is synchronized in real time to the individual causal directed acyclic graph model and the individual benefit-risk dual-objective dose-response surface model, supporting dynamic parameter updates.

[0014] Furthermore, a dual update path is set up for individual internal circulation and group external circulation. The individual internal circulation takes each treatment cycle or rehabilitation stage as a node to summarize the patient's protocol execution data and outcome changes at this stage. When the actual recovery outcome deviates from the predicted value by more than a threshold, the failure point in the intervention plan is traced back through the structural causal counterfactual reasoning framework to locate the link that caused the effect deviation, update the individual causal directed acyclic graph and the individual benefit-risk dual-objective dose-response surface, and optimize the accuracy of the next stage of plan generation. The external circulation of the population periodically aggregates and analyzes the rehabilitation data of desensitized gynecological tumor patients, incrementally updates the parameters of the population dose-effect prior model, the three-level hierarchical gynecological rehabilitation conflict knowledge graph and the structural causal counterfactual reasoning framework. The external circulation of the population has the ability to trace the source of rehabilitation outcomes, and can trace the implementation process and outcome differences of different intervention programs, and accumulate to form a best practice library for gynecological tumor rehabilitation.

[0015] This invention also provides a dynamic generation system for gynecological tumor rehabilitation intervention programs, used to implement the above-mentioned method, comprising the following modules: The data processing module integrates rehabilitation data, uses key treatment events as time anchors to complete time-series alignment and standardization mapping, divides rehabilitation time windows and constructs a full-cycle longitudinal dataset, combines domain prior constraints and uses causal discovery algorithms to construct causal directed acyclic graphs corresponding to each time window, forms individual causal directed acyclic graphs, and outputs core intervention targets and latent period pre-intervention nodes. The dose fitting module uses a Bayesian hierarchical modeling framework to train a multi-scenario group dose effect prior model, embeds gynecological tumor-specific safety dose constraint rules, and combines individual baseline characteristics and longitudinal rehabilitation data to update the posterior distribution, and fits to generate an individual benefit-risk dual-objective dose effect surface. The scheme generation and verification module matches the current stage rehabilitation goals and decomposes them into intervention domain sub-goals. It calls the individual benefit-risk dual-objective dose-response surface to generate candidate intervention schemes. After verifying the causal mediation effect based on the individual causal directed acyclic graph, it outputs a set of valid candidate schemes. The optimization module has a built-in three-level hierarchical gynecological rehabilitation conflict knowledge graph, which detects antagonistic items and risk points in candidate solutions, and uses a constrained multi-objective optimization algorithm to collaboratively adjust the intervention parameters of each domain, and outputs a set of intervention solutions. The deduction and optimization module builds a structural causal counterfactual reasoning framework, performs counterfactual deduction on candidate solutions, completes the comprehensive ranking of solutions through a five-dimensional evaluation system, and outputs the optimal intervention solution. The perception and adaptation module collects patient behavior data, uses a hidden Markov model to assess compliance status and triggers corresponding adaptation strategies. The iterative update module sets up dual update paths for individuals and groups to complete incremental updates of model parameters and the three-level hierarchical gynecological rehabilitation conflict knowledge graph.

[0016] The beneficial effects of this invention are as follows: 1. This invention integrates electronic medical record clinical data, wearable physiological and behavioral data, patient self-reported scale data, and gynecological oncology-specific assessment data. Using key treatment events such as the surgery date, the first day of each chemotherapy cycle, the start date of radiotherapy, and the first day of brachytherapy as time anchors, it achieves temporal alignment and standardized mapping, divides multi-scale rehabilitation time windows, and constructs a full-cycle longitudinal dataset. Combining prior constraints from the field of gynecological oncology rehabilitation, it employs a causal discovery algorithm to construct causal directed acyclic graphs corresponding to each time window, mining causal transmission relationships between symptoms, physiological indicators, adverse treatment reactions, and behavioral interventions. It identifies pre-causal markers and three-level transmission chains for latent complications, outputting core intervention targets and pre-intervention nodes during the latent period. This accurately distinguishes causal transmission and accompanying associations between symptoms, avoiding misjudgment of intervention targets and misallocation of rehabilitation resources, enabling pre-intervention during the latent period of complications, improving the targeting and foresight of rehabilitation interventions, and reducing the risk of serious long-term complications.

[0017] 2. This invention employs a Bayesian hierarchical modeling framework based on real-world data from gynecological oncology rehabilitation. It trains prior models of dose-effects across multiple scenarios, including the postoperative acute phase and the radiotherapy / chemotherapy phase, embedding specific safety dose constraints for gynecological oncology to quantify the nonlinear correlation between each intervention dimension and its corresponding outcome indicators. By dynamically updating the posterior distribution using individual baseline characteristics and longitudinal rehabilitation data, it fits a dual-objective dose-effect surface representing individual benefit and risk, determining the expected benefit, adverse reaction risk, and individual tolerance boundaries corresponding to different intervention doses. Relying on a three-tiered hierarchical gynecological rehabilitation conflict knowledge graph to detect conflict items, it uses a constrained multi-objective optimization algorithm to collaboratively adjust intervention parameters across domains. This achieves individualized and precise titration of intervention doses, identifies and resolves cross-domain explicit and implicit conflicts, leverages the synergistic gain effect of intervention combinations, reduces rehabilitation safety risks, and improves the overall effectiveness and patient tolerance of the program.

[0018] 3. This invention constructs a structural causal counterfactual reasoning framework based on individual causal directed acyclic graphs and individual benefit-risk dual-objective dose-response surfaces. It performs weekly, monthly, and quarterly counterfactual deductions on candidate protocols at multiple time scales, quantifying and predicting multi-dimensional rehabilitation outcome indicators and patient expected compliance under different parameters. A five-dimensional evaluation system including clinical benefit and safety dimensions is constructed, and the weights of each dimension are determined using the entropy weight method to complete the comprehensive ranking of candidate protocols. A hidden Markov model is used to assess patient compliance and trigger corresponding adaptation strategies. A dual update path for individuals and groups is set to achieve incremental iteration of the model and knowledge graph. This can improve the evidence-based support for protocol selection, adapt to changes in patient compliance, continuously optimize the accuracy of individual protocol generation, and accumulate best practices for gynecological oncology rehabilitation. Attached Figure Description

[0019] Figure 1 This is a flowchart of the dynamic generation method for gynecological tumor rehabilitation intervention programs according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, this example uses a 48-year-old FIGO IB2 stage cervical squamous cell carcinoma patient as the subject. The patient underwent radical hysterectomy with pelvic lymph node dissection and received concurrent chemoradiotherapy with paclitaxel and cisplatin after surgery. The patient is currently 12 weeks postoperatively and has completed the 3rd cycle of chemotherapy.

[0022] The study integrated clinical data from the patient's electronic medical record, including surgical records, concurrent chemoradiotherapy regimen, lymph node dissection extent, blood routine and liver and kidney function test indicators. It also collected physiological data such as heart rate, sleep structure, daily steps, and lower limb activity duration recorded by a smart bracelet. Simultaneously, it collected the patient's self-reported pain numerical score, fatigue scale, anxiety self-rating scale, and daily dietary records. Additionally, it incorporated gynecological tumor-specific assessment data such as lymphedema grade 0, pelvic floor muscle strength grade 3, pelvic pain score of 2, and reproductive function score of 45.

[0023] Using the patient's surgery date as the initial time anchor, the first day of the third cycle of chemotherapy as the current primary anchor, and the radiotherapy start date as the secondary anchor, all data records are encoded using the relative number of days of the event to complete the time sequence alignment of multi-source data.

[0024] During standardized mapping, clinical data were aligned and coded using a common terminology set for gynecologic oncology rehabilitation. Wearable physiological data were aggregated at the minute level and Z-score normalized. Scale data were uniformly converted to standard T-scores. Missing values ​​were filled using linear interpolation of adjacent observations within the same time window. Time periods with more than 30% consecutive missing values ​​were marked and filled using the mean of the same subtype of patients at the same stage combined with individual trends. These missing values ​​were not included in the effective sample set for core causal link calculations.

[0025] The rehabilitation time window is divided into multiple scales based on the patient's disease course. The postoperative acute phase window is subdivided into days within 14 days after surgery. The radiotherapy and chemotherapy phase window is subdivided into single chemotherapy cycles in sync with the 3-week chemotherapy cycle. The transition window from the end of the current radiotherapy and chemotherapy phase to the pelvic floor rehabilitation phase is subdivided into weeks. The subsequent pelvic floor rehabilitation phase window and follow-up period are subdivided into weeks and months, respectively. Finally, a longitudinal dataset of individual patients covering the entire rehabilitation cycle is constructed.

[0026] Combining a priori population causal network framework from the field of gynecologic oncology rehabilitation, this algorithm incorporates longitudinal observation data from various time windows of the patient for individualized edge weight calibration and local structure fine-tuning, generating an individualized causal directed acyclic graph adapted to the patient. The algorithm input is set with four types of nodes: exogenous nodes are treatment event variables such as radical hysterectomy, concurrent chemoradiotherapy with paclitaxel and cisplatin, and pelvic lymph node dissection; exposure nodes are intervention variables such as exercise intensity, frequency of pelvic floor training, and protein intake; mediating nodes are physiological symptom variables such as bilateral ankle circumference difference, pelvic floor muscle strength, fatigue score, and sleep duration; and outcome nodes are rehabilitation outcome variables such as lymphedema grade, pelvic floor function score, and quality of life score.

[0027] When enforcing domain prior constraints, treatment event nodes only serve as parent nodes, intervention nodes cannot act in reverse on treatment event nodes, and variables within the same time window do not have reverse causal edges. Using the population causal network skeleton as the initial structure, partial correlation tests are performed based on the longitudinal data of each time window to verify the significance of edges. Combining the V structure and domain prior rules, local edges are added, deleted, and oriented. After clinical logic verification, directed edges that do not conform to common sense in diagnosis and treatment are removed, and an individual causal directed acyclic graph with individual fitness scores is output. Edges with fitness scores below 0.7 are marked as paths to be verified.

[0028] To investigate the outcome of lower extremity lymphedema, a three-level conduction pathway was identified. The first-level pathway is the proximal pathway where increased lower extremity circumference difference directly affects the grade of lymphedema. The second-level pathway is the intermediate pathway where decreased daily activity level affects the increased lower extremity circumference difference. The third-level pathway is the distal initiation pathway where cancer-related fatigue aggravates decreased daily activity level.

[0029] The identified precursor causal marker for the latent period was a persistently elevated fatigue score, with the corresponding core intervention target being the duration of daily low-intensity walking. The precursor intervention point for the latent period was fatigue management intervention, which could be carried out 1 to 2 time windows before the onset of edema.

[0030] Using a Bayesian hierarchical modeling framework, a prior model of population dose effect under radiotherapy and chemotherapy scenarios was trained based on real-world data of gynecologic oncology rehabilitation. Specific safety dose constraints for gynecologic oncology were embedded to quantify the nonlinear correlation between intervention dimensions such as exercise intensity, pelvic floor training frequency, and protein intake and rehabilitation outcomes.

[0031] For this patient, specific safety constraints were embedded: high-intensity outdoor exercise was prohibited during the bone marrow suppression period of radiotherapy and chemotherapy; the duration of lower limb resistance training was limited to no more than 20 minutes per day during the high-risk state of lymphedema; the intensity of abdominal pressure exercises was limited to the low-intensity level during pelvic radiotherapy; and the corresponding intervention domain dose was automatically reduced by one level when grade II or above adverse reactions lasted for 3 days. All constraints were embedded into the reasoning process in a hard-truncation manner.

[0032] The population dose-response prior model is a two-layer Bayesian hierarchical structure. The individual layer fits the nonlinear response relationship between intervention dose and outcome for a single patient, while the population layer imposes distributional constraints on the individual layer parameters to share population rehabilitation patterns. The model input intervention dimensions are exercise intensity, frequency of pelvic floor training, and daily protein intake. Covariates include the patient's age of 48 years, extent of extensive hysterectomy, history of 3 cycles of chemotherapy, and baseline fitness score of 60. The model output includes rehabilitation benefit score, probability of adverse reactions, and individual tolerability score.

[0033] Bayesian hierarchical modeling formula for individual posterior distribution: This represents the individual-level model parameters for the i-th patient, corresponding to the response relationship parameters between the individual intervention dose and the outcome; This represents the baseline characteristics of the i-th patient and the longitudinal rehabilitation observation dataset; This represents the hyperparameters of the population layer, corresponding to the distribution parameters of the population dose-effect prior model, and carrying the prior knowledge of the rehabilitation patterns of all patients; The population prior distribution representing individual parameters is output by the population dose-effect prior model; The likelihood function represents the individual observation data and describes the degree of fit of the observation data under individual parameters. The posterior distribution of individual parameters is represented by the combination of population prior and individual data updates, and is used to fit the individual dose-response surface.

[0034] Based on the population prior model, the individual posterior distribution was updated using the patient's 12-week longitudinal rehabilitation data. A third-order B-spline function was used to fit and generate a bi-objective dose-response surface for individual benefit and risk. The calculated tolerance boundary for low-intensity walking was 45 minutes per day, corresponding to a 15% probability threshold for adverse reactions. The tolerance boundary for pelvic floor muscle training was 5 times per week; exceeding this significantly increased the risk of pelvic floor pain. The tolerance boundary for daily protein intake was 1.2 g / kg body weight; exceeding this increased the risk of gastrointestinal burden.

[0035] Matching the patient's rehabilitation stage at the end of radiotherapy and chemotherapy, the core rehabilitation goals were determined to be controlling cancer-related fatigue, improving pelvic floor muscle strength, preventing and controlling lower extremity lymphedema, and maintaining nutritional status. Considering the patient's current treatment stage at the end of radiotherapy and chemotherapy and age characteristics, the focus was on breaking down sub-goals into four intervention domains: symptom management, pelvic floor function rehabilitation, lymphedema prevention and control, and nutritional support. No mandatory intervention goals were set for the psychological and reproductive health protection domain. The symptom management domain aimed to reduce the fatigue score to below 3 points, the pelvic floor rehabilitation domain aimed to improve pelvic floor muscle strength to grade 4, the lymphedema prevention and control domain aimed to maintain lymphedema grade 0, and the nutritional support domain aimed to maintain serum albumin at no less than 35g / L.

[0036] By combining the core intervention targets and pre-intervention nodes output by the individual causal directed acyclic graph, candidate solutions are generated by calling the individual benefit-risk dual-objective dose-response surface, and patient preference weights are incorporated. This patient prefers home-based non-device intervention, can accept a total daily intervention duration of 40 minutes, and tends to perform exercise-related interventions in the morning.

[0037] The plan is generated by a three-step screening logic, matching intervention content such as low-intensity walking, home pelvic floor muscle training, and high-protein diet guidance. Within the safe tolerance boundary, three groups of gradient dose parameters (low, medium, and high) are selected. The intervention form and time period are adjusted according to the patient's preferences. Three candidate plans are generated for each intervention domain, and cross-domain combinations form a complete set of 81 candidate plans.

[0038] Causal mediation effect verification was performed based on individual causal directed acyclic graphs. Core mediating variables were preset for each domain: fatigue score and sleep duration for symptom management, pelvic floor muscle contraction force for pelvic floor rehabilitation, lower limb circumference difference for lymphedema prevention and control, and serum albumin level for nutritional support. The total effect of the intervention on the outcome and the indirect effect of the mediating pathway were calculated in sequence to obtain the proportion of the mediating effect.

[0039] Decomposition and proportion formula of causal mediation effect: This represents the overall effect of the intervention on the rehabilitation outcome, i.e., the total change in the outcome variable after the intervention was implemented; This represents the direct effect of the intervention, that is, the effect size of the intervention acting directly on the rehabilitation outcome without the intervention being mediated by pre-set variables; This refers to the indirect effect of the intervention, also known as the mediating effect, which is the effect size of the intervention indirectly affecting the rehabilitation outcome through the transmission path of a pre-defined mediating variable. This indicates the proportion of the mediating effect, that is, the proportion of the indirect effect in the total effect, and is used to determine the transmission effectiveness of the intervention target.

[0040] When the mediating effect ratio is less than 30% and the direct effect size does not reach the minimum effective threshold in clinical practice, it is judged as target misalignment and eliminated. After verification, 9 groups of protocols with insufficient mediating effect were eliminated, and a set of 72 effective candidate protocols was output.

[0041] The built-in three-level hierarchical gynecological rehabilitation conflict knowledge graph was invoked to perform conflict detection on 72 effective candidate solutions. Through graph pattern matching identification, 24 solutions had temporal phase conflicts, that is, lower limb walking training and pelvic floor muscle contraction training were performed at the same time every morning, and the superimposed abdominal pressure easily aggravated pelvic floor soreness.

[0042] Six treatment plans were found to be at a high-risk borderline state, meaning that in the patient's high-risk state of lymphedema, the daily walking time of some plans was close to the upper limit of tolerance, the safety margin for edema aggravation was insufficient, and there were no obvious absolute conflicting plans. At the same time, the synergistic gain coefficient of low-intensity walking combined with high-protein nutritional intervention was 1.12, indicating a synergistic effect.

[0043] A non-dominated sorting genetic algorithm with an elitist strategy was used to perform multi-objective optimization. The target weights were set according to the patient's late stage of radiotherapy and chemotherapy: rehabilitation benefit weight 0.4, safety conflict weight 0.4, and execution burden weight 0.2. The hard constraints were that the intervention dose should not exceed the tolerance boundary and the total daily intervention time should not exceed 45 minutes.

[0044] By resolving timing conflicts through staggered scheduling, pelvic floor muscle contraction training was adjusted to be performed in the afternoon, while walking training was retained in the morning. High-risk borderline states were mitigated by adjusting parameters, and the daily walking duration of the corresponding program was reduced to within a safe tolerance range, supplemented with lower limb elevation guidance.

[0045] After 100 iterations of optimization, eight sets of Pareto optimal solutions were output. All solutions had no explicit absolute conflict, implicit conditional conflict, or temporal phase conflict, and high-risk edge states were eliminated. The total daily intervention time was controlled within the range of 30 to 40 minutes, and the comprehensive rehabilitation benefit score covered different gradients to meet the needs of differentiated selection.

[0046] Based on the patient's individual causal directed acyclic graph and individual benefit-risk dual-objective dose-response surface, a structural causal counterfactual reasoning framework was constructed. The individual causal directed acyclic graph connecting the transition window at the end of the radiotherapy and chemotherapy period and the pelvic floor rehabilitation period window was used. Based on the time window progressive transmission rule, weekly and monthly multi-timescale extrapolation was performed on 8 groups of Pareto optimal plans. First, the values ​​of exogenous variables in the causal graph were corrected based on the patient's 12-week observation data, and the distribution of individual unobserved factors was fitted.

[0047] The intervention parameters of each group are then applied to the corresponding nodes as intervention parameters to cut off the influence of their parent nodes. Finally, the outcome variable values ​​at each time point are calculated in a forward manner. The extrapolation process combines the patient's expected compliance level output by the Hidden Markov Model with a differentiated compliance decay coefficient, with a preset baseline compliance of 85%.

[0048] Structural counterfactual intervention outcome expectation formula: It represents the expected counterfactual outcome, that is, the expected value of the recovery outcome variable Y when the intervention variable X is forcibly set to the value x; This represents the intervention quantifier in the structural causal model, which means that active intervention fixes variable X to x, while severing the causal connection between X and its original parent node to eliminate the influence of confounding factors. This represents the intervention variable, corresponding to the intervention parameters in the candidate program, such as exercise duration, training frequency, etc. Variables representing recovery outcomes, such as lymphedema grading and pelvic floor function scores; This represents the set of exogenous unobserved variables in a causal graph, corresponding to confounding factors that are not observed in the individual. Represents the space of all possible values ​​for the exogenous variable U; This represents the prior probability distribution of the exogenous, unobserved variable; This represents the conditional probability distribution of the outcome variable Y given the values ​​of the intervention and the exogenous variable.

[0049] The core results of the simulation output show that the top-ranked schemes can reduce the fatigue score from 5 to 3.8 points in weekly simulations, with no significant decrease in pelvic floor muscle strength and lymphedema maintained at grade 0. Monthly simulations show that pelvic floor muscle strength can be improved to grade 4-, the risk of lymphedema is controlled within 8%, serum albumin is maintained above 36g / L, and the probability of mild adverse reactions is about 12%.

[0050] A five-dimensional evaluation system was constructed to rank the protocols, with the following scores: clinical benefit 78 points, safety 85 points, execution difficulty 82 points, expected compliance 79 points, and patient preference matching 88 points.

[0051] Objective weights were calculated using the entropy weighting method, and subjective weights were set by combining them with the clinical priority at the end of the radiotherapy and chemotherapy period. The combined weights were: clinical benefit 0.28, safety 0.28, execution difficulty 0.16, expected compliance 0.14, and preference matching degree 0.14. After weighted calculation, the highest comprehensive score was determined to be 30 minutes of low-intensity walking every morning, 10 minutes of pelvic floor muscle training 5 times a week, and a daily protein intake of 1.0 g / kg body weight. This was identified as the optimal intervention and a corresponding outcome report was generated.

[0052] During the implementation of the program, wearable exercise data, pelvic floor training APP execution records, diet check-in data and symptom reporting records of the patient were collected in real time. Hidden Markov model was used to assess compliance status. The model was set with three hidden states: high compliance, medium compliance and low compliance. The observed variables were daily intervention completion rate, number of consecutive missed execution days, training duration target rate and check-in interval.

[0053] Starting from day 8 of the treatment regimen, the patient's daily intervention completion rate was less than 50% for three consecutive days, indicating a low compliance status. It was determined that the patient's fatigue score increased from 3.8 to 5.2 during the same period, an increase of more than 30%. The intervention dose was within the tolerance limit and the patient actively reported significant fatigue. Excluding factors such as intensity exceeding tolerance and subjective willingness, it was determined that the difficulty in implementation was due to the aggravation of symptoms after chemotherapy.

[0054] The patient's fatigue level was assessed as meeting the criteria for a Grade II adverse reaction and had persisted for 3 days. This triggered the corresponding adaptation strategy and safety constraint rules, reducing the daily walking time from 30 minutes to 20 minutes and the frequency of pelvic floor training from 5 times per week to 3 times per week. Symptom management interventions were supplemented with 15 minutes of relaxation training at noon every day. After the adaptation and adjustment, a 7-day observation and review cycle was set, and implementation data and symptom changes were continuously collected.

[0055] On the 7th day after the adjustment, the patient's fatigue score dropped to 4.1 points, and compliance improved to high compliance. After assessing the current tolerance status using the individual benefit-risk dual-objective dose-response surface model, the intervention intensity was gradually increased. Initially, the patient was restored to 25 minutes of walking and 4 times a week of pelvic floor training. After a 3-day adaptation period, the patient gradually returned to the original optimal dose. All adjustment data were synchronized in real time to the individual causal directed acyclic graph and dose-response surface model to support dynamic parameter updates.

[0056] A dual update path is set up for individual internal circulation and group external circulation. The individual internal circulation uses a single chemotherapy cycle as the update node, summarizing the patient's regimen execution data and outcome changes for this cycle. After the end of this cycle, the actual outcome is compared with the counterfactual predicted value. If the patient's pelvic floor muscle strength at the end of the month is actually grade 3+, while the predicted value is grade 4-, the deviation exceeds the 15% threshold, triggering the failure node backtracking.

[0057] Tracing back along the individual causal directed acyclic graph, the deviation of the pelvic floor muscle strength outcome node was located. Tracing upward, it was found that the actual execution frequency of pelvic floor muscle training was 20% lower than the preset value and lower than the projected baseline compliance of 85%. This was determined to be an effect deviation caused by intervention execution compliance deviation. Based on this, differentiated updates were performed: for outcome deviation caused by compliance execution deviation, only the emission probability matrix of the compliance assessment model was updated, without adjusting the causal graph structure and dose-response surface parameters, to optimize the compliance prediction accuracy of the next stage plan; if the deviation was caused by misjudgment of the causal link or the fitting deviation of the dose-response relationship, the individual causal directed acyclic graph and the individual benefit-risk dual-objective dose-response surface were updated simultaneously.

[0058] The group external circulation aggregates all desensitization and rehabilitation data monthly, incrementally updates the group dose-effect prior model, the three-level hierarchical gynecological rehabilitation conflict knowledge graph and the structural causal counterfactual reasoning framework parameters, and conducts outcome source analysis to select programs with core rehabilitation outcomes better than the average level of the same subtype, adverse reaction rates lower than the average level, and compliance rates of over 80% for inclusion in the best practice library.

[0059] The patient's adapted treatment plan met the inclusion criteria and was included in the best practice library for the subtype of concurrent chemoradiotherapy after cervical squamous cell carcinoma surgery. After the samples of the same subtype have accumulated to a batch size, they will be included in the incremental training dataset of the population model to support the improvement of the accuracy of treatment plan generation for similar patients.

[0060] The dynamic generation system for gynecological tumor rehabilitation intervention programs corresponding to this embodiment operates according to the following process based on the above-mentioned cervical cancer patient case: The data processing module receives the patient's multi-source rehabilitation data, completes time-series alignment and standardization with the surgery day and the first day of chemotherapy as anchor points, divides multi-scale time windows and constructs a longitudinal dataset, combines the prior population causal network skeleton in the field of gynecologic oncology rehabilitation, integrates the patient's longitudinal observation data to complete individualized edge weight calibration and local structure fine-tuning, generates an individual causal directed acyclic graph, and outputs fatigue score pre-markers and walking intervention core targets.

[0061] The dose fitting module calls the prior model of the group dose effect during radiotherapy and chemotherapy, embeds the specific safety constraints corresponding to the patient, updates the posterior distribution by combining individual baseline and longitudinal data, fits and generates a dual-objective dose effect surface of individual benefit and risk, and outputs the tolerance boundary of each intervention domain.

[0062] The solution generation and verification module matches the patient's current stage of rehabilitation goals and decomposes them into four sub-domain sub-goals, generates a gradient candidate solution set, and outputs a set of effective solutions after completing the causal mediation effect verification.

[0063] The optimization module calls a three-level hierarchical gynecological rehabilitation conflict knowledge graph to complete conflict detection, uses a multi-objective optimization algorithm to adjust the intervention timing and parameters, and outputs a Pareto optimal solution set.

[0064] The deduction and optimization module builds a counterfactual reasoning framework to complete the deduction of outcomes at multiple time scales, and outputs the optimal intervention plan after being ranked by a five-dimensional evaluation system.

[0065] The perception and adaptation module collects patient behavior data in real time, assesses compliance status through a hidden Markov model, and triggers dynamic adaptation strategies.

[0066] The iterative update module initiates an internal loop update at the end of the cycle, simultaneously connecting to the external loop of the group to complete the incremental iteration of the model and knowledge graph.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamically generating gynecological tumor rehabilitation intervention programs, characterized in that, Includes the following steps: Collect rehabilitation data, perform time-series alignment and standardization mapping with key treatment events as time anchors, divide rehabilitation time windows and construct longitudinal datasets; combine domain priors, construct individual causal directed acyclic graphs through causal discovery algorithms, and output core intervention targets and latent period pre-intervention nodes. A Bayesian hierarchical modeling framework was used to train a multi-scenario population dose-effect prior model, which was embedded with gynecologic oncology-specific safety constraints. The posterior distribution was updated by combining individual baseline characteristics and longitudinal rehabilitation data, and a bi-objective dose-effect surface of individual benefit and risk was fitted. Match the patient’s current rehabilitation goals, break them down into sub-goals of each intervention domain, generate candidate intervention plans, and verify them through causal mediation effects to output a set of effective candidate plans. The conflict items of the intervention scheme were detected by calling the three-level hierarchical gynecological rehabilitation conflict knowledge graph. The intervention parameters of each domain were adjusted by a constrained multi-objective optimization algorithm, and the adjusted intervention scheme set was output. Based on the individual causal directed acyclic graph and the individual benefit-risk bi-objective dose-response surface, a structural causal counterfactual reasoning framework is constructed. The candidate solutions are counterfactually deduced, and the optimal intervention solution is output after being ranked by the evaluation system. Collect patient behavior data, assess compliance status using a hidden Markov model, and adjust the intervention plan accordingly based on the compliance status assessment results. The system establishes a dual-update path consisting of an individual internal circulation and a group external circulation. The individual internal circulation updates the individual causal directed acyclic graph and the individual benefit-risk dual-objective dose-response surface by retrospectively analyzing intervention failure nodes based on stage rehabilitation data. The group external circulation aggregates desensitized group rehabilitation data and updates the group dose-response prior model and the three-level hierarchical gynecological rehabilitation conflict knowledge graph.

2. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The rehabilitation data includes electronic medical record clinical data, physiological and behavioral data collected by wearable devices, patient-reported symptom scales, psychological scales and dietary records, as well as gynecological tumor-specific assessment data. The time alignment uses key treatment events, including the day of surgery, the first day of chemotherapy, the start date of radiotherapy, and the first day of brachytherapy, as time anchors to divide the multi-scale rehabilitation time windows and construct the longitudinal dataset of the entire rehabilitation cycle.

3. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 2, characterized in that, The individual causal directed acyclic graphs are constructed according to each of the aforementioned recovery time windows to explore the causal transmission relationship between symptoms, physiological indicators, adverse treatment reactions, and behavioral interventions, identify the antecedent causal markers and tertiary transmission links of latent complications, and output the indirect influence path.

4. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The population dose-effect prior model covers five rehabilitation scenarios: postoperative acute phase, radiotherapy and chemotherapy phase, pelvic floor rehabilitation phase, follow-up phase, and endocrine therapy phase. It quantifies the nonlinear correlation between five intervention dimensions—exercise intensity, pelvic floor training frequency, nutritional intake level, psychological intervention frequency, and symptom management plan dosage—and corresponding rehabilitation outcome indicators. The specific safety constraints for gynecologic oncology are embedded in the entire reasoning process of the population dose-effect prior model.

5. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 4, characterized in that, The individual benefit-risk bi-objective dose-response surface is generated by updating the posterior distribution based on the population dose-response prior model, combined with the patient's individual baseline characteristics and the longitudinal rehabilitation data, and outputs three types of results for different intervention doses: expected rehabilitation benefit, risk of adverse reactions, and individual tolerance boundary.

6. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The rehabilitation goals are broken down into sub-goals in five specific intervention domains: symptom management, pelvic floor function rehabilitation, lymphedema prevention and control, nutritional support, and psychological and reproductive health protection. The candidate intervention plan includes four types of parameters: intervention content, implementation frequency, intensity level, and execution method. The candidate plan generation process incorporates patient preference weight parameters. The causal mediation effect verification is used to verify the effectiveness of the causal transmission path of the intervention measures and to eliminate candidate solutions with false associations or misaligned targets.

7. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The three-level hierarchical gynecological rehabilitation conflict knowledge graph covers three types of rules: explicit absolute conflict, implicit conditional conflict, and temporal phase conflict, and marks the synergistic gain coefficients of different intervention combinations. The constrained multi-objective optimization algorithm aims to maximize rehabilitation outcomes, minimize conflict risks, and minimize patient execution burden. It adjusts the target weights according to the rehabilitation stage and completes the coordinated adjustment of intervention parameters in each domain through methods including parameter adjustment, time-series staggering, and execution unit splitting.

8. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The compliance assessment, combined with the intervention program's implementation standards, identifies the reasons for decreased compliance and categorizes them into three typical scenarios: worsening symptoms leading to difficulty in implementation, exceeding tolerance for program intensity, and insufficient subjective willingness. The corresponding adaptation strategies are to reduce the intervention intensity and supplement symptom management, reduce to the suboptimal tolerable dose and split the execution unit, and increase positive incentives and switch the intervention form; the execution data is synchronized to the corresponding model in real time to support the dynamic updating of model parameters.

9. The method for dynamically generating gynecological tumor rehabilitation intervention programs according to claim 1, characterized in that, The individual internal circulation uses each treatment cycle or rehabilitation stage as an update node, summarizing the stage execution data and outcome changes. When the deviation between the actual outcome and the predicted value exceeds the threshold, the intervention failure node is traced back and the link that caused the effect deviation is located. The individual causal directed acyclic graph and the individual benefit-risk bi-objective dose-response surface are updated. The external circulation of the population periodically aggregates desensitized rehabilitation data, incrementally updates the parameters of the population dose-effect prior model, the three-level hierarchical gynecological rehabilitation conflict knowledge graph, and the structural causal counterfactual reasoning framework, and precipitates a best practice library for gynecological tumor rehabilitation.

10. A dynamic generation system for gynecological tumor rehabilitation intervention programs, used to implement the method described in any one of claims 1-9, characterized in that, include: The data processing module is used to integrate multi-source rehabilitation data, complete time alignment, standardization mapping and rehabilitation time window division with key treatment events as time anchors, construct a full-cycle longitudinal dataset, combine domain prior constraints, construct individual causal directed acyclic graphs through causal discovery algorithms, and output core intervention targets and latent period pre-intervention nodes. The dose fitting module is used to train a multi-scenario group dose effect prior model using a Bayesian hierarchical modeling framework, embed gynecologic oncology-specific safety constraints, update the posterior distribution by combining individual baseline characteristics and longitudinal rehabilitation data, and fit the individual benefit-risk bi-objective dose effect surface. The protocol generation and verification module is used to match the patient's current stage of rehabilitation goals and break them down into sub-goals of each intervention domain, generate candidate intervention protocols, complete causal mediation effect verification, and output a set of valid candidate protocols. The optimization module has a built-in three-level hierarchical gynecological rehabilitation conflict knowledge graph to detect conflict items in candidate solutions. It uses a constrained multi-objective optimization algorithm to adjust the intervention parameters of each domain and outputs a set of intervention solutions. The deduction and optimization module is used to build a structural causal counterfactual reasoning framework, perform counterfactual deductions on candidate solutions at multiple time scales, and complete the sorting of solutions and output the optimal intervention solution through a multi-dimensional evaluation system. The perception and adaptation module is used to collect patient behavior data, use a hidden Markov model to assess compliance status, trigger corresponding adaptation strategies, and complete the adjustment of the intervention plan. The iterative update module is used to set up dual update paths for individual internal circulation and group external circulation, and to complete the incremental updates of the group dose-effect prior model parameters, the three-level hierarchical gynecological rehabilitation conflict knowledge graph, and the structural causal counterfactual reasoning framework.

Images