A gynecological tumor rehabilitation period multi-source follow-up data fusion evaluation method and system

Abstract

The application belongs to the technical field of tumor rehabilitation evaluation, and discloses a gynecological tumor rehabilitation period multi-source follow-up data fusion evaluation method and system. Four core data domains of in-hospital clinical data, out-of-hospital time sequence sign data, patient self-reported outcome data and health behavior data are constructed, and each data domain is subdivided into a cervical cancer special disease subdomain to include multi-type follow-up data. Special disease semantic mapping rules are established to complete the coding unification and measurement caliber conversion of multi-source data. Special disease data quality checking rules are synchronously set, and a differential privacy algorithm is used to complete patient sensitive data desensitization processing. A follow-up time axis is constructed in units of natural weeks to realize time sequence data alignment, a stage partition type interpolation network is built, a recognized rehabilitation trajectory of each rehabilitation stage is embedded as an a priori constraint to complete missing data, and a confidence weight with missing time length attenuation and a stage sensitivity coefficient are set for the completed data.

Description

Technical Field

[0001] This invention belongs to the field of tumor rehabilitation assessment technology, specifically a method and system for multi-source follow-up data fusion assessment during the rehabilitation period of gynecological tumors. Background Technology

[0002] With the widespread adoption of cervical cancer screening systems and the standardized application of treatments such as radical surgery and concurrent chemoradiotherapy, the 5-year survival rate of cervical cancer patients has significantly improved. Follow-up during the recovery period, as a core component of the entire disease management process, plays a crucial role in early identification of signs of recurrence and metastasis, assessment of pelvic floor and physical function recovery, intervention for long-term complications related to treatment, and improvement of patients' quality of life. Currently, cervical cancer rehabilitation follow-up and assessment techniques face the following technical challenges:

[0003] Rehabilitation-related data is scattered across multiple independent channels, including in-hospital information systems, laboratory imaging systems, patient follow-up platforms, and wearable devices. It covers multimodal data such as structured numerical values, semi-structured scales, unstructured reports, and continuous time-series vital signs. Existing technologies lack a standardized data mapping framework for cervical cancer rehabilitation scenarios, resulting in low data interoperability. Clinical assessments rely on fragmented manual integration, which can easily lead to the omission of key rehabilitation information such as out-of-hospital symptoms and changes in physical fitness.

[0004] The existing assessment system relies heavily on objective indicators such as in-hospital tumor markers and pelvic imaging examinations. It does not adequately cover dimensions such as physical condition, pelvic floor dysfunction, symptom burden, and psychological state outside the hospital, and cannot reflect the true state of home rehabilitation. Moreover, it is mostly a static judgment at discrete time points, lacking continuous analysis of temporal trends, and it is difficult to capture weak abnormal signals in the early stage of relapse and the progressive decline in function.

[0005] Existing multi-source data fusion methods mostly employ fixed weighting, simple feature splicing, or general deep learning models. They do not incorporate the pathophysiological laws and stage characteristics of cervical cancer rehabilitation to set differentiated fusion logic, thus failing to dynamically adapt to the differences in assessment focus at different rehabilitation stages. Consequently, the accuracy of recurrence risk prediction and the differentiation of rehabilitation status are insufficient. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for the fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors, so as to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors, comprising the following specific steps: Furthermore, a cervical cancer rehabilitation follow-up data metadata set was constructed, dividing the follow-up data into four core data domains: in-hospital clinical data, out-of-hospital time-series vital signs data, patient self-reported outcome data, and health behavior data. Each data domain was further subdivided into cervical cancer-specific subdomains, and structured, semi-structured, and continuous time-series follow-up data of the corresponding dimensions were included respectively. Establish disease-specific semantic mapping rules for data, and match corresponding standard codes and unified measurement conversion rules for clinical laboratory indicators, disease diagnosis and adverse reactions, and scale data respectively, so as to unify the measurement caliber of data from different sources. Set up disease-specific data quality verification rules to identify and mark high-priority abnormal data, and use differential privacy algorithms to desensitize patient identity and sensitive medical history data.

[0008] Furthermore, a follow-up time axis is constructed based on the natural week as the basic time unit. Dimensionality reduction and aggregation are performed on high-frequency wearable time series data to generate statistical features. Low-frequency clinical test and imaging data are aligned to time nodes and mapped to the corresponding follow-up week observation nodes. Based on the patient's radical treatment method and postoperative duration, a standard rehabilitation stage is automatically divided. A stage-partitioned interpolation network is constructed. The recognized rehabilitation trajectory of each stage of cervical cancer is added to the loss function of the corresponding branch as a soft constraint in the form of a prior regularization term to guide the interpolation results to conform to the clinical rehabilitation pattern. When completing the data, the patient's rehabilitation stage is matched and the corresponding prior constraint is called to generate the interpolation results. To supplement the data, a confidence weight was assigned. The basic confidence weight of the supplemented data was lower than that of the measured data in the same period. Furthermore, the confidence of the supplemented data decreased with the duration of the missing data. A stage sensitivity coefficient was set to lower the confidence of the supplemented results at key follow-up nodes and trigger a clinical re-examination reminder.

[0009] Furthermore, deep feature extraction is performed on the four core data domains respectively, and feature engineering is carried out for the characteristics of each data domain to extract numerical features, time series features, scale score features and behavioral features of the corresponding dimensions, and incorporate features related to cervical cancer-specific complications. Combining the core clinical dimensions of cervical cancer rehabilitation, and through unsupervised dimensionality reduction and clinical expert weight calibration, five rehabilitation characterization factors were constructed: tumor recurrence risk factor, physical function recovery factor, pelvic floor function rehabilitation factor, psychosocial adaptation factor, and health behavior compliance factor. Each of these rehabilitation characterization factors is nested with corresponding disease-specific complication subfactors. Clinical anchoring indicators are set for each of the rehabilitation characterization factors. The changing trend of the rehabilitation characterization factor scores is consistent with the clinical significance direction of the anchoring indicators. When the deviation exceeds the threshold, calibration is automatically triggered. At the same time, clinical logic verification rules between factors are set, and correlation constraints are set based on the pathophysiological laws of cervical cancer rehabilitation.

[0010] Furthermore, a three-stage attention fusion architecture is constructed. The first layer is set as an in-domain feature attention layer. For the feature set corresponding to each rehabilitation representation factor, the features are divided into groups according to the data domain. The attention network automatically learns the importance weights of different features in each data domain, strengthens the feature weights of core monitoring indicators, and weakens the interference of non-specific indicators. The second layer sets up a joint attention layer between temporal domains, which adjusts the global weights of the four core data domains according to the patient's recovery stage. At the same time, it amplifies the weights of indicators that show significant temporal change trends, automatically increases the weights of indicators that show a worsening trend in continuous follow-up, and sets up a recurrence signal-specific stepwise amplification mechanism to stepwise weight the continuous abnormal changes of core tumor markers and link them to increase the weights of their associated clinical signs and symptoms. The third layer sets up a rehabilitation stage gating layer as the top-level hard constraint. Based on the rehabilitation stage to which the patient belongs, it outputs the allowable range of weights for each data domain, limiting the weight adjustment range of the underlying attention network. At the same time, a cross-domain feature conflict resolution mechanism is set up in this layer to handle the trend deviation between objective indicators in the hospital and self-reported data from outside the hospital. Finally, a comprehensive rehabilitation feature vector that integrates information from all dimensions is output through the three-stage attention fusion architecture.

[0011] Furthermore, based on the comprehensive rehabilitation feature vector, rehabilitation status stratification and relapse risk classification and early warning are performed. The rehabilitation status stratification adopts density peak clustering algorithm combined with rehabilitation evaluation criteria in clinical diagnosis and treatment guidelines to divide the patient's rehabilitation status into four levels: high-quality rehabilitation, stable rehabilitation, limited rehabilitation, and high-risk status. The sub-item scores and dimensional distribution of each rehabilitation characteristic factor are output simultaneously to clarify the patient's rehabilitation shortcomings and dominant complication types. The recurrence risk grading and early warning system is based on a gradient boosting tree risk prediction model trained on a cervical cancer follow-up dataset. It takes a comprehensive rehabilitation feature vector as input and outputs the recurrence and metastasis probability of patients in different time windows, which are divided into three risk levels: low risk, medium risk, and high risk. The risk driving factors are identified by feature importance analysis, and the risk source categories are distinguished.

[0012] Furthermore, a factor-oriented intervention generation engine is constructed, a cervical cancer rehabilitation intervention knowledge base is established, and domestic and international cervical cancer rehabilitation guidelines, expert consensus and evidence-based medicine are integrated. The knowledge base includes multiple types of intervention measures, and each type of measure is set with a corresponding intervention intensity ladder. Each ladder is matched with clear rehabilitation characterization factor score thresholds and adaptation conditions. A hybrid matching strategy is used to generate intervention plans. The scores of each rehabilitation characteristic factor and disease-specific complication subfactor are used as inputs. Intervention measures are given for the corresponding steps of single-factor abnormal matching. For multi-factor abnormal matching, a combined intervention plan is automatically formed. The basic plan is optimized by combining guidelines and similar historical rehabilitation cases. The clinical pathway validation layer is set up to ensure that the plan meets the requirements of clinical standards, and finally an individualized intervention plan is generated, which includes intervention content, execution frequency, and review and adjustment suggestions.

[0013] Furthermore, after patients implement the intervention plan, the new data collected during subsequent follow-ups are automatically entered into the evaluation process. The changes in various rehabilitation characteristics before and after the intervention are compared, the intervention effect is quantified, and feedback is given to the clinical and patient ends to guide the adjustment of the plan. At the same time, the improvement of rehabilitation characteristic factors corresponding to individual intervention measures is automatically quantified, the effectiveness rate of the intervention population is calculated, and the improvement attribution results of factors corresponding to the intervention effect are fed back to the cervical cancer rehabilitation intervention knowledge base to adjust the matching priority and intensity parameters of intervention measures. A dual-track incremental update mechanism is set up to separate the group model iteration from the individual baseline calibration, and the newly added follow-up data and manual correction results are summarized. An incremental learning algorithm is used to incrementally train the three-stage attention fusion architecture and gradient boosting tree risk prediction model. Patient-specific rehabilitation baselines are updated synchronously, and subsequent follow-up assessments are prioritized for comparison with these baselines.

[0014] This invention also provides a multi-source follow-up data fusion and evaluation system for the recovery period of gynecological tumors, used to implement the above-mentioned method, comprising the following modules: The data collection and standardization module constructs a data set of cervical cancer rehabilitation follow-up data, divides it into four core data domains and sets up disease-specific subdomains, establishes disease-specific semantic mapping rules to unify data coding and measurement standards, and simultaneously performs disease-specific data quality verification and differential privacy desensitization processing of sensitive patient data. The time alignment and data completion module constructs a follow-up timeline in natural weeks to complete the alignment of data time nodes, builds a stage-partitioned interpolation network, embeds prior constraints of rehabilitation trajectories in each rehabilitation stage to perform adaptive completion of missing data, and configures confidence weights for the completed data. The feature extraction and characterization construction module performs feature engineering on the four core data domains, extracts features and incorporates disease-specific complication features, constructs five major rehabilitation characterization factors and corresponding disease-specific sub-factors, and ensures the clinical consistency of factor output through anchoring calibration and logical verification. The three-stage attention fusion module builds a three-stage attention fusion architecture, which sequentially completes intra-domain feature weighting, cross-domain dynamic weight adjustment and stage gating constraints, and outputs a comprehensive rehabilitation feature vector. The rehabilitation assessment and risk warning module, based on the comprehensive rehabilitation feature vector, completes the stratification of rehabilitation status and the classification of relapse risk, analyzes risk driving factors and distinguishes the risk source categories; The individualized intervention generation module combines a rehabilitation intervention knowledge base with historical similar cases and outputs individualized intervention plans after clinical pathway verification. The incremental iteration module enables feedback on intervention effects, incremental model training, and updates to patient-specific rehabilitation baselines.

[0015] The beneficial effects of this invention are as follows: 1. This invention constructs four core data domains: in-hospital clinical data, out-of-hospital time-series vital sign data, patient-reported outcome data, and health behavior data. Each data domain is further subdivided into cervical cancer-specific subdomains to incorporate multiple types of follow-up data. Specific semantic mapping rules are established to unify the encoding and convert measurement standards for multi-source data. Specific data quality verification rules are simultaneously set, and a differential privacy algorithm is used to anonymize sensitive patient data. A follow-up timeline is constructed using natural weeks to align time-series data. A stage-partitioned interpolation network is built, embedding recognized recovery trajectories for each recovery stage as prior constraints to complete missing data. A confidence weight and stage sensitivity coefficient are set for the completed data, which decay with the duration of missing data. This improves data standardization and overall usability, ensuring the continuity and accuracy of assessments in data-missing scenarios, while also considering patient privacy and providing clinical risk warnings at key follow-up points.

[0016] 2. This invention performs deep feature engineering on four core data domains, extracting numerical, temporal, scale, and behavioral features and incorporating them into the cervical cancer-specific complication feature dimension. Through unsupervised dimensionality reduction combined with clinical expert weight calibration, it constructs five rehabilitation characterization factors and nested sub-factors of specific complications, along with a clinical anchoring calibration mechanism and clinical logic verification rules between factors. A three-stage attention fusion architecture is built, which sequentially learns intra-domain feature weights through an intra-domain feature attention layer, performs cross-domain dynamic weighting through a temporal domain joint attention layer, and applies top-level constraints and resolves cross-domain conflicts through a rehabilitation stage gating layer, ultimately generating a comprehensive rehabilitation feature vector. This approach aligns with the pathophysiological laws and stage characteristics of cervical cancer rehabilitation, accurately capturing weak abnormal signals in the early stages of recurrence and the progressive decline in function, thus improving the differentiation of rehabilitation status and the accuracy of recurrence risk prediction.

[0017] 3. This invention uses density peak clustering combined with clinical standards to stratify rehabilitation status based on comprehensive rehabilitation feature vectors. A gradient boosting tree risk prediction model outputs relapse risk levels across multiple time windows, analyzes risk drivers, and distinguishes risk source categories. A factor-oriented intervention generation engine is constructed, integrating evidence-based medicine to build a rehabilitation intervention knowledge base. Intervention measures are matched with factor scores and historically similar cases, and individualized plans are generated after clinical pathway validation. An intervention effect feedback mechanism is set up, establishing a dual-track incremental update mechanism that separates group model iteration from individual baseline calibration, simultaneously updating model parameters and patient-specific rehabilitation baselines. This improves the individualized adaptability and clinical compliance of intervention plans, and gradually eliminates assessment bias caused by individual differences through continuous iterative optimization of model performance. Attached Figure Description

[0018] Figure 1 This is a flowchart of the multi-source follow-up data fusion and evaluation method for the gynecological tumor rehabilitation period of the present invention. Detailed Implementation

[0019] The following is a complete description of this protocol with reference to the accompanying drawings and specific clinical scenarios. This embodiment uses a 38-year-old patient with stage IB2 cervical squamous cell carcinoma as the subject of the procedure. The patient underwent laparoscopic radical hysterectomy combined with bilateral pelvic lymph node dissection. The patient was in the late acute phase of the postoperative period at the 8th week after surgery. During home rehabilitation, the patient wore a smart bracelet, filled out a patient self-report questionnaire weekly, and returned to the hospital monthly to complete tumor marker and gynecological examination follow-up.

[0020] like Figure 1 As shown, the multi-source follow-up data fusion and evaluation process for gynecological tumor rehabilitation in this embodiment proceeds sequentially according to the following steps.

[0021] First, a cervical cancer rehabilitation follow-up data metadata set was constructed, dividing the patient's follow-up data into four core data domains: in-hospital clinical data, out-of-hospital time-series vital sign data, patient self-reported outcome data, and health behavior data. Each data domain was further divided into cervical cancer-specific subdomains, which respectively contained the corresponding types of structured test data, semi-structured scale data, and continuous time-series vital sign data.

[0022] The in-hospital clinical data is divided into three subdomains: tumor monitoring, surgical treatment, and complication diagnosis and treatment. These subdomains include the patient's tumor marker results such as SCC and CA125, radical surgery records and postoperative pathology reports, and postoperative vaginal stump healing status. The out-of-hospital time-series vital signs data is divided into two subdomains: physical activity and vital signs. These subdomains include daily steps, resting heart rate, and sleep duration collected by wristbands. The patient self-reported outcome data is divided into three subdomains: symptom burden, pelvic floor function self-assessment, and psychological state. These subdomains include the scores of the QLQ-C30 scale, pelvic floor dysfunction scale, and anxiety self-assessment scale completed by the patient. The health behavior data is divided into two subdomains: follow-up compliance and lifestyle. These subdomains include follow-up attendance rate and daily diet and rest records.

[0023] Specific semantic mapping rules were established, and standardized coding was performed on the above three types of data. Clinical laboratory indicators were matched with LOINC standard coding, with SCC units standardized as ng / mL and CA125 units as U / mL, and classified into the tumor monitoring subdomain. Surgery and diagnosis were matched with ICD-10 and ICD-9-CM-3 codes, mapped to labels such as cervical squamous cell carcinoma stage IB2 and radical hysterectomy. Scale data were matched with the PRO standard coding system, and mapped to subdomains such as physical function and pelvic floor function according to dimensions. All mapping results corresponded one-to-one with the characteristic dimensions of subsequent rehabilitation characterization factors.

[0024] Three types of disease-specific data quality verification rules were set up to identify abnormal data. The numerical range verification marked that the patient's SCC value of 12.5 ng / mL in week 6 exceeded the normal range after surgery. The temporal change verification identified that the patient's resting heart rate suddenly increased to 110 beats / min on a certain day, which deviated from the mean of the three adjacent days by 3 times the standard deviation. The disease-specific logic verification found that the patient reported no lower limb swelling symptoms, but the physical examination record showed grade I lymphedema in the left lower limb. The two were clinically contradictory, so they were marked as high-priority abnormalities and manual review was triggered.

[0025] Abnormal data is processed in a tiered manner. Low-priority heart rate mutation data is included in the missing data completion process. Mid-priority abnormal data with numerical ranges are marked as pending verification and their input weight is reduced when participating in the completion of adjacent node data. High-priority clinical contradictory data have patient self-reported entries with low credibility removed, while in-hospital objective physical examination records are retained. The corresponding self-reported node is marked as invalid data and is not included in the feature calculation of the corresponding dimension. At the same time, a special follow-up is triggered to supplement the collection of valid self-reported data for that node.

[0026] Differential privacy algorithms are used to desensitize patient sensitive data in a graded manner. Identification fields such as patient name and ID number are processed by adding noise perturbation to the numerical values; fields such as reproductive history and pathological stage are processed by interval generalization, such as "stage IB2" being generalized to "stage I". A graded privacy budget is set in the desensitization process, with different budget thresholds corresponding to single-field queries and multi-field joint queries, so as to preserve the clinical statistical value of the data while ensuring privacy.

[0027] A follow-up timeline was constructed for the patient based on the natural week. The daily high-frequency time-series data collected by the wristband was dimensionality-reduced and aggregated to generate three statistical features: weekly average steps, weekly resting heart rate fluctuation, and weekly average sleep duration. The monthly low-frequency clinical test data such as SCC and CA125 were aligned to the follow-up observation node of the current week and mapped to the test feature values ​​of the corresponding follow-up week.

[0028] Based on the radical surgical treatment method and the duration of 8 weeks after surgery, patients are automatically determined to be in the acute postoperative phase. The recovery phase of radical surgical patients is uniformly divided into three stages: 0-3 months after surgery is the acute postoperative phase, 3-12 months after surgery is the functional recovery phase, and more than 12 months after surgery is the long-term follow-up stable phase.

[0029] During the construction phase, a partitioned interpolation network is used to complete missing data. The network is set up with independent interpolation branches according to three recovery stages. Each branch contains two hidden layers, and the number of neurons in each layer is twice the feature dimension of the corresponding data domain. The activation function is ReLU.

[0030] The recognized rehabilitation trajectory at each stage is added to the loss function of the corresponding branch in the form of a priori regularization term as a soft constraint to complete the interpolation results and guide them to conform to the clinical rehabilitation pattern. Among them, the acute postoperative branch is embedded with the trajectory constraint of rapid decline of tumor markers and stepwise recovery of physical function.

[0031] For the missing SCC test data in week 7 of the patient, the postoperative acute phase branch network was called. The measured tumor marker sequences of week 5 and week 8 and the abnormal SCC sequence marked as to be reviewed in week 6 were input. The abnormal data in week 6 were reduced in input weight and then used in interpolation calculation. The initial interpolation result was output. Then, the measured data of the adjacent two weeks were combined for weighted correction to generate the complete value of SCC in week 7.

[0032] To complete the data, a confidence weight was assigned. The base confidence weight of the completed data was 60% of the measured data in the same period. For each additional follow-up week of missing data, the confidence weight was reduced by 10%, with a minimum of 20%. In this example, the patient only had 1 week of missing SCC data, so the confidence weight of the completed data was 60% of the measured data.

[0033] Formula for calculating the confidence weight of the completed data: To complete the confidence weight of the data, the value is taken as a weight ratio relative to the measured data of the same period; The baseline confidence weight coefficient, used to complete the data, is fixed at 0.6. The confidence decay coefficient for single-week missing values ​​is fixed at 0.1. For consecutive follow-up weeks where data is missing; The minimum confidence threshold for completing the data is fixed at 0.2.

[0034] A stage sensitivity coefficient is set to further lower the confidence level of the supplementary results at key follow-up points such as 3 months, 6 months, and 12 months post-surgery, and simultaneously trigger clinical follow-up reminders; if data is missing at the 3-month post-surgery follow-up point, the supplementary confidence level is multiplied by a sensitivity coefficient of 0.7 on the original basis, and a follow-up reminder is automatically pushed to the attending physician.

[0035] Feature engineering was carried out on the four core data domains. Numerical features such as weekly change rate of SCC and absolute value of CA125 were extracted from in-hospital clinical data; time-series features such as week-on-week change rate of steps and heart rate variability were extracted from out-of-hospital time-series data; scale features such as physical function score and pelvic floor dysfunction score were extracted from self-reported data; and behavioral features such as follow-up compliance rate and daily exercise duration were extracted from health behavior data.

[0036] Simultaneously incorporating the characteristics of cervical cancer-specific complications, for the mild lymphedema of the left lower extremity in this patient after surgery, the difference in lower extremity circumference and edema grading characteristics were extracted from the specialist physical examination record, and the lower extremity swelling and pain score was extracted from the symptom scale. Together, they constituted the characteristic dimension of lymphedema complication, which was classified as a subfactor of disease-specific complications under the pelvic floor function rehabilitation factor.

[0037] Combining the core clinical dimensions of cervical cancer rehabilitation, we first used principal component analysis to reduce the dimensionality of all initial features and extracted five principal components with a cumulative variance contribution rate of over 85%. Then, we incorporated the weighted scores of three gynecological oncology rehabilitation experts using analytic hierarchy process (AHP), and weighted the expert weights with the principal component loading coefficients to finally construct five rehabilitation characterization factors: tumor recurrence risk factor, physical function recovery factor, pelvic floor function rehabilitation factor, psychosocial adaptation factor, and health behavior compliance factor. Among them, a higher score for the tumor recurrence risk factor indicates a higher risk of recurrence, while higher scores for the other four factors indicate a better rehabilitation status in the corresponding dimension.

[0038] Among them, the characteristics related to complications such as lymphedema are aggregated separately into disease-specific complication subfactors, which are nested under the corresponding rehabilitation characterization factors. The current scores of the five factors in this patient are as follows: tumor recurrence risk factor 72 points, physical function recovery factor 65 points, pelvic floor function rehabilitation factor 58 points, psychosocial adaptation factor 70 points, and health behavior compliance factor 80 points.

[0039] Clinical anchoring indicators were set for the five rehabilitation characterization factors: tumor recurrence risk factor anchored to SCC level, physical function recovery factor anchored to physical performance status score, pelvic floor function rehabilitation factor anchored to pelvic floor muscle strength grade, psychosocial adaptation factor anchored to anxiety and depression scale score, and health behavior compliance factor anchored to follow-up compliance rate. The trend of factor score change must be consistent with the clinical significance of the anchoring indicator, and automatic calibration is triggered when the deviation exceeds the threshold.

[0040] In this embodiment, the patient's SCC decreased for two consecutive weeks, but the tumor recurrence risk factor score increased slightly. The deviation between the two exceeded the threshold, and the system automatically triggered feature verification. First, it checked whether there were any abnormalities in the input data of non-anchored features. If there were no abnormalities, the factor score was retained and the reason for the deviation was marked. If it was confirmed that there was data noise in the low correlation feature, the weight of the corresponding feature was reduced until the deviation between the factor score trend and the anchor index fell back to within the threshold. After calibration, the patient's tumor recurrence risk factor score was adjusted to 68 points, which was consistent with the downward trend of SCC.

[0041] Simultaneously, clinical logic verification between factors is performed. The verification found that the patient's pelvic floor muscle strength grade was 3 and the pelvic floor function score was low, but the overall physical function score was high, indicating a clinical logic deviation. This automatically triggered feature verification. First, it checked whether there were any abnormalities in the input data corresponding to the physical function factor. If there were no abnormalities, the factor score was retained and the reason for the deviation was marked. If it was confirmed that there was data noise or weight allocation deviation, the weight calibration process for the physical function factor was executed.

[0042] A three-stage attention fusion architecture is constructed to perform multi-domain feature fusion. The three-layer network is connected in series, and the core parameters include multiple configurations. The intra-domain feature attention layer adopts a 4-head self-attention structure and calculates weights using scaled dot product attention. The query vector is generated based on the correlation between features and core indicators of cervical cancer rehabilitation. The inter-temporal joint attention layer adopts a gated attention unit structure, and the recurrence signal step amplification mechanism is implemented through a piecewise weight function. The rehabilitation stage gate layer uses a sigmoid gate function to generate weight intervals. Finally, a 64-dimensional comprehensive rehabilitation feature vector is output, which corresponds one-to-one with the input dimensions of subsequent clustering and risk prediction models.

[0043] The first layer is the domain feature attention layer, which is a feature set corresponding to the tumor recurrence risk factors. The features are divided into groups according to data domains such as in-hospital clinical and out-of-hospital signs. The importance weights of each feature in the domain are learned through a multi-head self-attention network.

[0044] In this embodiment, the weights of core tumor markers such as SCC and CA125 in the in-hospital clinical data domain are strengthened, while the weights of non-specific indicators such as weight fluctuations are weakened, and the final output is a weighted aggregated feature vector within the domain.

[0045] The second layer is the joint attention layer between time domains. Based on the stage attributes of the patient's acute postoperative period, the global weight of the in-hospital clinical data domain is increased, the weight of the health behavior data domain is decreased, and the weight of indicators with significant time-series changes is amplified, while the weight of indicators with continuous deterioration is automatically increased.

[0046] A recurrence signal amplification mechanism is set up. If a patient's SCC exceeds the normal range for two consecutive follow-up weeks, the corresponding feature weight is amplified by 1.5 times, and the weight of related symptoms such as lower limb swelling and pain and pelvic distension is increased in a synchronized manner; if it is abnormal for three consecutive weeks or more, the weight is amplified by 2.5 times.

[0047] Formula for weighting coefficient of recurrence signal step amplification: The weighting amplification factor for the core tumor marker features; The number of follow-up weeks for core tumor markers that consistently exceed the normal reference range.

[0048] Associated features are amplified according to the level of clinical evidence. High-association features are amplified proportionally to tumor markers, while medium- and low-association features are amplified proportionally. In this patient, the current SCC is showing a downward trend, and the stepwise amplification mechanism has not been triggered.

[0049] The third layer is the rehabilitation stage gating layer, which serves as the top-level hard constraint to output the allowable weight range of each data domain and limit the weight adjustment range of the bottom-level attention. The allowable weight range for in-hospital clinical data in the postoperative acute phase is 0.35-0.45, and the allowable weight range for out-of-hospital self-reported data is 0.15-0.25. The upper and lower limits of the range are preset according to the clinical assessment priority of this stage.

[0050] This layer establishes a cross-domain feature conflict resolution mechanism. If the in-hospital SCC index is stable, but the patient's self-reported lower abdominal distension continues to worsen, indicating a divergence, and the cosine similarity of the trends between the two is less than 0.3, the in-hospital objective test index is used as the core basis for recurrence risk assessment, and its weight remains unchanged. Only the weight of the self-reported symptom in the tumor recurrence risk factor is reduced to 40% of its original weight, while its weight in the corresponding factors such as pelvic floor function and physical function is retained. At the same time, a symptom-specific follow-up is triggered to investigate the symptom causes. If the in-hospital index is abnormal but the patient has no corresponding symptoms, the in-hospital index is used as the core, and the priority of imaging review is increased. Finally, the architecture outputs a 64-dimensional comprehensive rehabilitation feature vector.

[0051] Based on a 64-dimensional comprehensive rehabilitation feature vector, a density peak clustering algorithm combined with cervical cancer rehabilitation guidelines was used to complete the stratification of rehabilitation status. The algorithm calculates the sample spacing using Euclidean distance, determines the number of clusters by the inflection point of the product of local density and distance, and then maps and matches each cluster to four clinical levels: high-quality rehabilitation, stable rehabilitation, limited rehabilitation, and high-risk status based on the clinical features of each cluster.

[0052] The clustering results of this patient indicate a stable rehabilitation level. The scores of five factors were output simultaneously, clearly showing that pelvic floor function rehabilitation factor is the main rehabilitation weakness, and left lower extremity lymphedema is the dominant complication type.

[0053] The LightGBM gradient boosting tree model trained on a 5000-case annotated cervical cancer follow-up dataset was used to predict recurrence risk. The core parameters of the model include 200 trees, a maximum depth of 8, a learning rate of 0.05, and a minimum number of samples per leaf node of 20. The model was trained using binary cross-entropy as the loss function and was completed through five-fold cross-validation.

[0054] The model takes a comprehensive rehabilitation feature vector as input and outputs the recurrence and metastasis probabilities for three time windows: 6 months, 12 months, and 24 months. The corresponding grading thresholds are set according to the time window: 6 months: low risk <5%, intermediate risk 5%-20%, high risk ≥20%; 12 months: low risk <10%, intermediate risk 10%-30%, high risk ≥30%; 24 months: low risk <15%, intermediate risk 15%-40%, high risk ≥40%.

[0055] The patient's relapse rate at 6 months was 3.2%, classifying it as low-risk. Analyzing the significance of features using SHAP values ​​identified the core risk drivers as lower extremity lymphedema and decreased physical performance, which fall under the category of functional rehabilitation influencing factors. It was also noted that clinical attention should be paid to the progression of lymphedema and guidance on improving physical performance.

[0056] We construct a factor-driven intervention generation engine and establish a three-tiered cervical cancer rehabilitation intervention knowledge base. The basic intervention level includes general healthy lifestyle guidance, the specialist intervention level corresponds to specific rehabilitation programs for five rehabilitation characteristic factors, and the symptomatic intervention level corresponds to the treatment measures for various specific disease complications. Each type of measure has three levels of intervention intensity: low, medium, and high. Each level is matched with clear factor score thresholds and adaptation conditions.

[0057] A hybrid matching strategy was used to generate a treatment plan. First, abnormal scores of pelvic floor function rehabilitation factors and accompanying abnormalities of lymphedema subfactors were identified in patients. Pelvic floor muscle training specialist intervention and lower extremity lymphedema symptomatic intervention were then matched. Next, corresponding intervention intensity steps were matched according to the degree of deviation of each factor and subfactor. The pelvic floor function rehabilitation factor was matched with a secondary intervention intensity, and the lymphedema subfactor was matched with a corresponding secondary intervention intensity. Finally, basic dietary and lifestyle guidance was added to form an initial treatment plan.

[0058] The matching process verifies the suitability of the measures. After confirming that there are no contraindications to implementation, the historical similar case library is called up. The similarity is calculated using weighted distance based on the scores of the five factors. The intervention improvement effects of the top 5 similar cases are selected, and the intervention frequency and intensity are optimized and adjusted to form a basic intervention plan.

[0059] The clinical pathway validation layer is set up to verify the compliance of the protocol from three dimensions: intervention contraindication validation to screen patients for contraindications to compression therapy such as skin ulceration and cardiovascular disease; intensity matching validation to confirm that the intensity of the secondary intervention matches the physical condition of patients in the postoperative acute phase; and follow-up cycle validation to match the guidelines for tumor markers combined with pelvic ultrasound 3 months after surgery.

[0060] Once all verifications pass, a final individualized plan will be generated, including pelvic floor muscle training twice a day for 15 minutes each time, wearing lower limb compression stockings and completing one ankle pump exercise daily, moderate-intensity walking three times a week, and a follow-up examination at the hospital three months post-surgery. If the verification fails, the system will automatically return to the engine to adjust the parameters and re-verify.

[0061] Four weeks after the patient implemented the intervention plan, the newly collected follow-up data was automatically entered into the evaluation process. The changes in factor scores before and after the intervention were compared. The pelvic floor function rehabilitation factor increased from 58 to 67 points, and the lower extremity lymphedema score decreased. The quantitative results of the intervention effect were simultaneously pushed to the clinical end and the patient end.

[0062] By combining the intervention effect weights of the same population and removing the cumulative effect of other interventions during the same period, the independent improvement of pelvic floor muscle training on the patient's pelvic floor function factor was estimated to be 7 points. Among the same batch of intervention population, the proportion of patients whose factor improvement reached the clinical minimum important difference value was 68%. The attribution result of this improvement was fed back into the intervention knowledge base, and the matching priority of the corresponding intervention was increased.

[0063] A dual-track incremental update mechanism is set up to separate the population model iteration from the individual baseline calibration.

[0064] At the population model level, streaming incremental learning is adopted. Incremental training is triggered every 100 newly added valid follow-up data, freezing 70% of the parameters at the bottom layer of the model and only fine-tuning the top-level attention parameters and classifier parameters. The learning rate is 0.2 times that of the initial training to avoid catastrophic forgetting. The training weight of manually corrected samples from senior physicians is higher than that of ordinary samples.

[0065] At the individual baseline level, after a patient has accumulated three or more stable follow-up data, the average of the first three stable follow-up data is taken as the initial specific rehabilitation baseline; if there are fewer than three stable follow-up data, the baseline data of the population with the same stage and the same treatment method is used as the temporary baseline.

[0066] The baseline was updated every two new, verified, and stable follow-up data points without abnormal fluctuations. A weighted moving average was used for calculation, with the historical baseline accounting for 60% and the mean of the two new follow-up data points accounting for 40%. After the patient completed the 4-week intervention, their specific rehabilitation baseline was updated simultaneously. Subsequent assessments were compared with this baseline rather than the general standard to eliminate individual differences and biases.

[0067] Individualized Rehabilitation Baseline Update Formula: For the updated patient-specific recovery baseline; The baseline for historical rehabilitation before the update; The mean values ​​of the corresponding features for the two newly added stable follow-up data; The weighting coefficient for the historical baseline is fixed at 0.6. The weighting coefficient for the newly added follow-up data is fixed at 0.4.

[0068] This embodiment also provides a multi-source follow-up data fusion and evaluation system for the rehabilitation period of gynecological tumors, which is used to implement the above method and includes 7 modules that operate in a coordinated manner.

[0069] The data acquisition and standardization module completes the access, coding unification, and desensitization verification of multi-source patient data; the time sequence alignment and data completion module constructs a weekly follow-up timeline, performs missing data completion, and assigns confidence scores; the feature extraction and representation construction module extracts multi-dimensional features and generates five rehabilitation representation factors; the three-stage attention fusion module performs three-level attention fusion and outputs a comprehensive feature vector; the rehabilitation assessment and risk warning module outputs rehabilitation stratification and relapse risk levels; the individualized intervention generation module matches and verifies the output intervention plan; and the incremental iteration module completes effect feedback and dual-track updates.

[0070] For the rehabilitation assessment of this patient, each module sequentially performs the data processing, feature fusion, assessment and early warning, and intervention generation process according to the above procedure.

[0071] This embodiment, through the above-described full-process processing, achieves standardized fusion of multi-source follow-up data from a single cervical cancer patient after surgery, enabling precise rehabilitation assessment and individualized intervention generation. This verifies the feasibility and clinical suitability of this method in home-based rehabilitation follow-up scenarios. Those skilled in the art can make adaptive adjustments based on the principles of this solution for patients with different stages and treatment methods; all related modifications fall within the scope of protection of this invention.

[0072] 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.

[0073] 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 fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors, characterized in that, Includes the following steps: Collect follow-up data in the context of cervical cancer rehabilitation, and complete data standardization coding, unified measurement standards, disease-specific data quality verification, and desensitization of sensitive patient data; Construct a follow-up timeline to align time-series data, and combine prior constraints of the rehabilitation phase to perform missing data completion and confidence weighting; Feature engineering was performed on multi-domain follow-up data to extract multi-dimensional features and incorporate disease-specific complication feature dimensions. A rehabilitation characterization factor system was constructed and anchoring calibration and inter-factor logical verification were performed. A three-stage attention fusion architecture is used to fuse multi-domain features and generate a comprehensive rehabilitation feature vector. Based on the comprehensive rehabilitation feature vector, rehabilitation status is stratified and relapse risk is classified, risk driving factors are analyzed and risk source categories are distinguished. By combining a rehabilitation intervention knowledge base with similar historical cases, and after verification through clinical pathways, individualized rehabilitation intervention plans are generated. An intervention effect feedback mechanism is set up to quantify the intervention improvement effect and provide feedback to adjust the intervention knowledge base parameters. A dual-track incremental update mechanism is used to complete the incremental training of the model and the update of the patient's rehabilitation baseline.

2. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, The follow-up data is divided into four core data domains: in-hospital clinical data, out-of-hospital time-series vital signs data, patient self-reported outcome data, and health behavior data. Each data domain is further subdivided into cervical cancer subdomains and the corresponding types of follow-up data are included. Establish disease-specific semantic mapping rules, match standard codes to clinical laboratory indicators, disease diagnosis and adverse reactions, and scale data respectively, and unify the measurement caliber. Set the data quality verification rules for the specific disease, identify and mark abnormal data, and use a differential privacy algorithm to perform the desensitization processing on the sensitive patient data.

3. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, The follow-up time axis is constructed based on the natural week as the basic time unit. The high-frequency time series data is reduced and aggregated, and the low-frequency clinical test and imaging data are aligned with time nodes. The rehabilitation stages are divided based on the patient's treatment method and postoperative duration. A stage-partitioned interpolation network with embedded prior constraints of the rehabilitation trajectory is constructed to complete the missing data. To supplement the data, the confidence weight of the configuration is set, the stage sensitivity coefficient is adjusted to lower the confidence of the key node supplementation results and trigger clinical review reminders.

4. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, Numerical, temporal, scale, and behavioral features are extracted from each of the data domains and incorporated into the disease-specific complication feature dimensions related to cervical cancer-specific complications. Combining the core clinical dimensions of cervical cancer rehabilitation, five rehabilitation characterization factors and corresponding sub-factors of specific disease complications were constructed through unsupervised dimensionality reduction and clinical expert weight calibration, forming the rehabilitation characterization factor system. Clinical anchoring indicators are set for each of the rehabilitation characterization factors, and logical verification rules are established between the factors. When the deviation exceeds the threshold, the anchoring calibration is automatically triggered.

5. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, The first layer of the three-stage attention fusion architecture is the intra-domain feature attention layer. In the intra-domain feature attention layer, features are grouped according to the data domain. The importance weights of intra-domain features are learned through the attention network to strengthen the weights of core monitoring indicators and weaken the interference of non-specific indicators. The second layer is a joint attention layer between time domains. It adjusts the global weights of each data domain according to the recovery stage, amplifies the weights of indicators with significant time-series changes and continuous deterioration, sets up a relapse signal stepwise amplification mechanism, performs stepwise weighting on continuous abnormal changes of core tumor markers, and links up the weights of associated clinical signs and symptoms.

6. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, The third layer of the three-stage attention fusion architecture is the rehabilitation stage gating layer, which constrains the weight adjustment range of the underlying attention network based on the allowable weight range of each data domain output by the patient to which the rehabilitation stage belongs. The rehabilitation stage gate layer also sets up a cross-domain feature conflict resolution mechanism to handle the trend deviation between objective indicators within the hospital and self-reported data outside the hospital, and finally outputs the comprehensive rehabilitation feature vector that integrates information from all dimensions.

7. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, Based on the comprehensive rehabilitation feature vector, the density peak clustering algorithm combined with the clinical rehabilitation evaluation criteria is used to divide the patient's rehabilitation status into four levels, complete the stratification of the rehabilitation status, and simultaneously output the sub-scores of each rehabilitation characterization factor to clarify the rehabilitation shortcomings and the dominant complication types. A gradient boosting tree risk prediction model is trained based on a cervical cancer follow-up dataset. The model outputs the recurrence and metastasis probabilities for different time windows and classifies the risk levels, thus completing the recurrence risk classification. The risk driving factors are identified and the risk source categories are distinguished through feature importance analysis.

8. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, Establish a knowledge base for rehabilitation interventions for cervical cancer that integrates evidence-based medicine, including multiple types of intervention measures and setting corresponding intervention intensity levels, with each level matched with the score thresholds and fitting conditions of the rehabilitation characterization factors; A hybrid matching strategy is adopted, with intervention measures corresponding to single-factor abnormal matches and combined intervention measures for multi-factor abnormal matches. The basic plan is optimized by combining clinical guideline requirements and the historical similar cases, and the individualized rehabilitation intervention plan is generated after being verified by the clinical pathway.

9. The method for fusion and evaluation of multi-source follow-up data during the recovery period of gynecological tumors according to claim 1, characterized in that, By comparing the changes in each rehabilitation characteristic factor before and after the intervention, the improvement of the corresponding factor by a single intervention measure is quantified, the effectiveness rate of the intervention population is calculated, and the matching priority and intensity parameters of the rehabilitation intervention knowledge base are adjusted accordingly. The dual-track incremental update mechanism separates the group model iteration from the individual baseline calibration. Based on the newly added follow-up data and the results of manual correction, the model is incrementally trained, and the patient's recovery baseline is updated simultaneously as a reference for subsequent evaluation.

10. A multi-source follow-up data fusion and evaluation system for the recovery period of gynecological tumors, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition and standardization module divides the data into four core data domains and sets up disease-specific subdomains, establishes disease-specific semantic mapping rules, unifies data coding and measurement standards, and performs disease-specific data quality verification and desensitization of sensitive patient data. The time-series alignment and data completion module constructs a follow-up timeline to complete data alignment, and completes missing data completion and confidence weighting through an interpolation network with embedded prior constraints. The feature extraction and characterization construction module carries out multi-domain feature engineering and incorporates the characteristic dimensions of disease-specific complications, constructs a rehabilitation characterization factor system, and performs calibration and logical verification. The three-stage attention fusion module builds a three-stage attention fusion architecture, completes multi-domain feature fusion, and outputs a comprehensive rehabilitation feature vector. The rehabilitation assessment and risk warning module completes the stratification of rehabilitation status, the classification of relapse risk, and the analysis of risk driving factors. The individualized intervention generation module combines a rehabilitation intervention knowledge base with historical similar cases and outputs individualized rehabilitation intervention plans after clinical pathway verification. The incremental iteration module enables feedback on intervention effects, incremental model training, and updates to the patient's rehabilitation baseline.

Images