Artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients

Abstract

The application discloses a prostate cancer patient pelvic floor muscle recovery evaluation system based on artificial intelligence, which comprises an original data collection module, an original data optimization module, a time sequence double-end recovery evaluation model establishment module, a recovery evaluation model performance optimization module and a whole-cycle intelligent recovery evaluation module. The application relates to the technical field of medical data processing, in particular to a prostate cancer patient pelvic floor muscle recovery evaluation system based on artificial intelligence. The scheme innovatively proposes a time sequence double-end recovery evaluation model combining an in-hospital baseline end and a home continuous time sequence end, improves the real-time performance of patient pelvic floor muscle recovery evaluation, extracts short-term time sequence recovery features by using a bidirectional long short-term memory network, extracts long-term time sequence recovery features by combining an improved convolution residual network with a parallel activation function, significantly improves the accuracy of the recovery evaluation results, and introduces an improved optimization algorithm combined with a dynamic local search strategy, thereby significantly improving the stability and accuracy of the output results of the model.

Description

Artificial Intelligence-Based Pelvic Floor Muscle Recovery Assessment System for Prostate Cancer Patients Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to an artificial intelligence-based assessment system for pelvic floor muscle recovery in prostate cancer patients. Background Technology

[0002] The AI-based pelvic floor muscle recovery assessment system for prostate cancer patients is a system that uses artificial intelligence technology to collect, process, and comprehensively analyze multi-source data to assess the recovery of pelvic floor muscle function in prostate cancer patients. This system transforms the recovery process from traditional experience-based judgment into an objective, intelligent, and visualized evaluation model, providing decision support for clinicians and personalized rehabilitation guidance for patients.

[0003] However, traditional pelvic floor muscle recovery assessment systems for prostate cancer patients suffer from several technical problems. These include a single assessment dimension, lack of continuous monitoring, and inability to dynamically reflect the changing trends of pelvic floor muscle function over time, leading to inaccurate assessment results. Furthermore, existing recovery assessment models struggle to simultaneously capture both short-term fluctuations and long-term trends in dynamic, continuous data, resulting in inaccurate assessments. Finally, existing models for recovery assessment suffer from reliance on human experience during hyperparameter setting, low search efficiency, and a tendency to get trapped in local optima, leading to insufficient generalization ability and decreased assessment accuracy. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients. Addressing the technical problems of traditional prostate cancer patient pelvic floor muscle recovery assessment systems, such as limited assessment dimensions, lack of continuous monitoring, and inability to dynamically reflect the changing trends of pelvic floor muscle function over time, leading to inaccurate assessment results, this solution innovatively proposes a time-series dual-end recovery assessment model combining in-hospital baseline and home continuous time-series endpoints. This improves the completeness of the pelvic floor muscle recovery assessment, overcomes the shortcomings of traditional single-point detection methods that lead to delayed and biased conclusions, and enhances the real-time performance and accuracy of the pelvic floor muscle recovery assessment. It achieves continuous, intelligent, and full-cycle assessment of pelvic floor muscle function at follow-up points and in daily life. Furthermore, addressing the technical problem of existing recovery assessment models struggling to simultaneously capture short-term fluctuations and long-term trends in dynamic continuous data, resulting in inaccurate pelvic floor muscle recovery assessment results, this solution innovatively incorporates a time-series endpoint at home... This method employs a bidirectional long short-term memory network to extract short-term temporal recovery features and combines it with an improved convolutional residual network incorporating parallel activation functions to extract long-term temporal recovery features. This effectively enhances the ability to capture short-term dynamic fluctuations and improves the representation of the patient's long-term recovery process. It overcomes the limitations of traditional methods that rely solely on a single time scale, thus significantly improving the accuracy and reliability of recovery assessment results. This enables a dynamic, continuous, and precise intelligent assessment of the patient's pelvic floor muscle function recovery status. Addressing the technical problems of existing recovery assessment models that rely on human experience, have low search efficiency, and are prone to getting trapped in local optima during hyperparameter setting, leading to insufficient generalization ability and decreased assessment accuracy, this solution innovatively introduces an improved optimization algorithm incorporating a dynamic local search strategy. This achieves efficient global exploration and fine-tuned local optimization of the assessment model's hyperparameters, obtaining the optimal hyperparameter combination and significantly improving the stability and accuracy of the model's output results. This enables a precise and intelligent assessment of the patient's pelvic floor muscle recovery status.

[0005] The technical solution adopted by the present invention is as follows: The artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients provided by the present invention includes a raw data collection module, a raw data optimization module, a time-series dual-end recovery assessment model establishment module, a recovery assessment model performance optimization module, and a full-cycle intelligent recovery assessment module;

[0006] The raw data collection module is used to acquire multi-source data required for the patient's pelvic floor muscle recovery assessment, specifically by collecting data to obtain raw data for the pelvic floor muscle recovery assessment.

[0007] The original data optimization module is used to improve the reliability and modeling applicability of the original data. Specifically, it performs data cleaning, data standardization, data encoding, and feature screening on the original data to obtain optimized data for the evaluation of the patient's pelvic floor muscle recovery.

[0008] The module for establishing a time-series dual-end recovery assessment model is used to construct an intelligent assessment model that can simultaneously characterize the static baseline state and the dynamic recovery trend. Specifically, it extracts in-hospital baseline recovery features through the in-hospital baseline end, and extracts home-based time-series recovery features through the home-based continuous time-series end, which combines a bidirectional long short-term memory network and an improved convolutional residual network. The two types of features are then fused through a dual-end feature fusion layer to generate comprehensive recovery features, which are then input into the assessment result output layer to obtain the recovery status assessment result, thereby completing the construction of the time-series dual-end recovery assessment model.

[0009] The recovery evaluation model performance optimization module is used to improve the performance and stability of the time series two-end recovery evaluation model. Specifically, it first performs preliminary training on the model based on historical data, then obtains the optimal hyperparameter combination of the evaluation model by introducing an optimization algorithm improved by a dynamic local search strategy, and finally retrains the model in the final training stage by combining the optimal hyperparameter combination of the evaluation model to obtain the time series two-end recovery evaluation model with the best performance.

[0010] The full-cycle intelligent recovery assessment module specifically inputs real-time data into the optimal time-series dual-end recovery assessment model to obtain the patient's real-time pelvic floor muscle recovery status assessment results, thereby realizing intelligent assessment of the patient's pelvic floor muscle function.

[0011] Furthermore, the raw data collection module specifically collects data through the medical terminal management system and wearable devices to obtain raw data for pelvic floor muscle recovery assessment; the raw data for patient pelvic floor muscle recovery assessment includes historical pelvic floor muscle recovery assessment data and real-time pelvic floor muscle recovery assessment data; both the historical pelvic floor muscle recovery assessment data and the real-time pelvic floor muscle recovery assessment data include in-hospital phased test data and home continuous monitoring data; the historical pelvic floor muscle recovery assessment data also includes historical recovery assessment levels.

[0012] Furthermore, the raw data optimization module specifically includes the following steps:

[0013] Data cleaning and processing specifically includes outlier removal, missing value handling, data deduplication, and data validation.

[0014] Data standardization specifically involves using the min-max normalization method to map the numerical data of the original data to a uniform numerical range.

[0015] Data encoding specifically involves using label encoding methods to encode the category fields in the original data, transforming them into structured numerical values;

[0016] Feature selection involves measuring the importance of each feature based on correlation analysis, setting a threshold as a selection criterion, eliminating redundant features below the threshold, and selecting the feature set that is most discriminative for recovery assessment.

[0017] Furthermore, the module for establishing a time-series two-end recovery evaluation model specifically includes the following steps:

[0018] The in-hospital baseline is specifically analyzed and feature extracted from the in-hospital phased detection data using a gated bidirectional convolutional neural network to obtain the in-hospital baseline recovery features.

[0019] The home-based continuous time-series terminal specifically includes the following steps:

[0020] Short-term time-series recovery feature extraction involves inputting continuous home monitoring data into a bidirectional long short-term memory network to obtain short-term time-series recovery features.

[0021] Long-term temporal recovery feature extraction specifically involves inputting short-term temporal recovery features into an improved convolutional residual network to obtain long-term temporal recovery features.

[0022] The improved convolutional residual network specifically involves first inputting short-term temporal recovery features into the first convolutional unit. Preliminary temporal features are extracted by expanding causal convolution, combining weight normalization and random deactivation. Subsequently, the LeakyReLU activation function is used to enhance the response to negative inputs, generating the first convolutional feature. The first convolutional feature is then input into the second convolutional unit, where multi-scale dependencies are extracted by expanding causal convolution and weight normalization. A parallel activation structure of PReLU and hyperbolic tangent functions is set, and the output is weighted and fused and then randomly deactivated to generate the second convolutional feature. Finally, the short-term temporal recovery features are convolved with the second convolutional feature element-wise and added to it to output the long-term temporal recovery features.

[0023] Feature aggregation specifically involves aggregating short-term and long-term time-series recovery features using a feature weighted fusion method to obtain home-based time-series recovery features.

[0024] The dual-end feature fusion layer specifically integrates the in-hospital baseline recovery features and the home time-series recovery features using an attention-based mechanism to fuse the features from the two time-series ends, resulting in a comprehensive recovery feature.

[0025] The evaluation result output layer specifically involves inputting the comprehensive recovery features into a fully connected layer and mapping them to a Softmax function. This process calculates the probability distribution of each recovery level and determines the patient's current recovery level based on the maximum probability judgment strategy, thereby obtaining the recovery status evaluation result.

[0026] Furthermore, the performance optimization module for the recovery evaluation model specifically includes the following steps:

[0027] The initial training of the evaluation model involves using historical pelvic floor muscle recovery evaluation data, processed by the original data optimization module, as training data to perform initial training of the time-series dual-end recovery evaluation model, resulting in the initial trained time-series dual-end recovery evaluation model.

[0028] Adaptive hyperparameter search, specifically, involves obtaining the optimal hyperparameter combination for a time-series two-end recovery evaluation model through an improved optimization algorithm; it includes the following steps:

[0029] Initialize the search individuals by encoding the hyperparameters of the temporal two-end recovery evaluation model into search individual position vectors, and generate N search individual position vectors through a random initialization method. Each individual encoding represents a candidate combination of hyperparameters of the temporal two-end recovery evaluation model, thus obtaining the initial search population.

[0030] The fitness value of a search individual is calculated by calculating the fitness value of the search individual in the population; the performance of the temporal two-end recovery evaluation model based on the location of the search individual is used as the fitness value of the search individual.

[0031] The search process involves dividing individuals into groups, specifically selecting the individual with the best fitness as the leader and the remaining individuals as followers.

[0032] Individual position iterative updates specifically involve updating the leader's and follower's individual positions; this includes the following steps:

[0033] The leader's individual position is updated; specifically, the leader's individual position is updated using the following formula:

[0034] ;

[0035] In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. This represents the current optimal individual position in the j-th dimension. They represent Random numbers within a range and These represent the upper and lower bounds of the global search space, respectively. Indicates search control parameters. express Random numbers within a certain range;

[0036] Follower individual position update, specifically, updating the position of each follower individual to the average position of itself and the previous search individual;

[0037] Dynamic local search involves recalculating the fitness value of all search individuals after iteratively updating the individual's position, and selecting the individual with the best fitness value as the current iteration's best individual. If the fitness value of the current iteration's best individual is not better than that of the previous iteration's best individual, the number of unimproved individuals is incremented by one; otherwise, the number of unimproved individuals is reset to zero, and a local search judgment is performed. When the cumulative number of unimproved individuals exceeds four, the local search process is triggered.

[0038] The local search process specifically involves first selecting k search individuals within a preset radius based on Euclidean distance, centered on each search individual, to form its neighborhood set. Within this neighborhood set, the individual with the best fitness value is determined as the optimal neighboring individual. This is then combined with the globally optimal individual to update the local search position, generating local search candidate individuals. If the candidate individual's fitness value is better than the current iteration's optimal individual, it replaces the current iteration's optimal individual, and the unimproved count is reset to zero. Otherwise, the current iteration's optimal individual remains unchanged, and the unimproved count remains constant. The formula used is as follows:

[0039] ;

[0040] In the formula, Indicates the location of candidate individuals in a local search. Indicates the first The current iteration's optimal individual position is obtained after iterative updates. This represents the location of the best individual in the neighborhood. This represents the globally optimal individual position. and They represent A random number within a given range;

[0041] The global optimal position update for a search individual involves evaluating the fitness value of the search individual in the current iteration and comparing it with the global optimal position of the current search individual based on the fitness value of the current best individual. If the fitness value of the current best individual is better, then the global optimal position of the search individual is updated.

[0042] The iterative search terminates when the global optimal position of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached. The global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the evaluation model.

[0043] The final training of the evaluation model involves adjusting the hyperparameters of the initially trained temporal two-end recovery evaluation model based on the optimal hyperparameter combination of the evaluation model, and using historical pelvic floor muscle recovery evaluation data processed by the original data optimization module as training data to retrain the initially trained temporal two-end recovery evaluation model to obtain the optimal performance temporal two-end recovery evaluation model.

[0044] Furthermore, the full-cycle intelligent recovery assessment module specifically inputs real-time pelvic floor muscle recovery assessment data into the optimal performance time-series dual-end recovery assessment model to obtain the patient's real-time pelvic floor muscle recovery status assessment result, thereby realizing a full-cycle continuous intelligent assessment of the patient's pelvic floor muscle function.

[0045] The beneficial effects achieved by the present invention using the above solution are as follows:

[0046] (1) In response to the technical problems of the traditional pelvic floor muscle recovery assessment system for prostate cancer patients, such as the single assessment dimension, lack of continuous monitoring, and inability to dynamically reflect the trend of pelvic floor muscle function changes over time, resulting in inaccurate pelvic floor muscle recovery assessment results, this solution innovatively proposes a time-series dual-end recovery assessment model that combines the baseline end in the hospital and the continuous time end at home. This improves the completeness of the pelvic floor muscle recovery assessment, overcomes the shortcomings of the traditional single-point detection that leads to delayed and biased conclusions, enhances the real-time and accuracy of the pelvic floor muscle recovery assessment, and realizes continuous, intelligent and full-cycle assessment of the patient's pelvic floor muscle function at follow-up points and in daily life.

[0047] (2) In view of the technical problem that existing recovery assessment models have difficulty capturing both short-term fluctuations and long-term trends in time series when processing dynamic continuous data, resulting in inaccurate assessment results of patients' pelvic floor muscle recovery, this solution innovatively uses a bidirectional long short-term memory network to extract short-term recovery features in the home continuous time series, and combines it with an improved convolutional residual network that introduces parallel activation functions to extract long-term recovery features. This effectively improves the ability to capture short-term dynamic fluctuations, improves the expression effect of patients' long-term recovery process, overcomes the limitation of traditional methods that rely on only a single time scale, and thus significantly improves the accuracy and reliability of recovery assessment results, realizing a dynamic, continuous and precise intelligent assessment of patients' pelvic floor muscle function recovery status.

[0048] (3) In view of the technical problems of existing recovery assessment models, which rely on human experience, have low search efficiency and are prone to getting trapped in local optima during the hyperparameter setting process, resulting in insufficient generalization ability and decreased assessment accuracy, this solution innovatively introduces an improved optimization algorithm that combines dynamic local search strategy. This realizes efficient global exploration and fine local optimization of the hyperparameters of the assessment model, obtains the optimal hyperparameter combination of the assessment model, significantly improves the stability and accuracy of the model's output results, and realizes precise and intelligent assessment of the patient's pelvic floor muscle recovery status. Attached Figure Description

[0049] Figure 1 is a schematic diagram of the modules of the artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients provided by the present invention;

[0050] Figure 2 is a flowchart illustrating the process of establishing a time-series two-end recovery assessment model module;

[0051] Figure 3 is a flowchart illustrating the performance optimization module of the recovery evaluation model;

[0052] Figure 4 is a flowchart illustrating the home-based continuous time series endpoint in the module for establishing a time-series dual-end recovery assessment model.

[0053] Figure 5 is a flowchart illustrating the adaptive hyperparameter search process in the performance optimization module of the recovery evaluation model.

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0057] Example 1, referring to Figure 1, the artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients provided by the present invention includes a raw data collection module, a raw data optimization module, a time-series dual-end recovery assessment model establishment module, a recovery assessment model performance optimization module, and a full-cycle intelligent recovery assessment module;

[0058] The raw data collection module is used to acquire multi-source data required for the patient's pelvic floor muscle recovery assessment. Specifically, it collects data to obtain raw data for pelvic floor muscle recovery assessment and sends the data to the raw data optimization module.

[0059] The raw data optimization module receives data sent by the raw data collection module to improve the reliability and modeling applicability of the raw data. Specifically, it performs data cleaning, data standardization, data encoding, and feature screening on the raw data to obtain optimized data for the patient's pelvic floor muscle recovery assessment. The data is then sent to the recovery assessment model performance optimization module and the full-cycle intelligent recovery assessment module.

[0060] The module for establishing a time-series dual-end recovery assessment model is used to construct an intelligent assessment model that can simultaneously characterize the static baseline state and the dynamic recovery trend. Specifically, it extracts in-hospital baseline recovery features through the in-hospital baseline end, and extracts home-based time-series recovery features through the home-based continuous time-series end, which combines a bidirectional long short-term memory network and an improved convolutional residual network. The two types of features are then fused through a dual-end feature fusion layer to generate comprehensive recovery features, which are then input into the assessment result output layer to obtain the recovery status assessment result. This completes the construction of the time-series dual-end recovery assessment model, and the data is sent to the recovery assessment model performance optimization module.

[0061] The recovery assessment model performance optimization module receives data sent by the time-series two-end recovery assessment model module to improve the performance and stability of the time-series two-end recovery assessment model. Specifically, it first performs preliminary training on the model based on historical data, then obtains the optimal hyperparameter combination of the assessment model by introducing an optimization algorithm improved by a dynamic local search strategy, and retrains the model in the final training stage by combining the optimal hyperparameter combination of the assessment model to obtain the time-series two-end recovery assessment model with the best performance, and sends the data to the full-cycle intelligent recovery assessment module.

[0062] The full-cycle intelligent recovery assessment module receives data from the original data optimization module and the recovery assessment model performance optimization module. Specifically, it inputs real-time data into the optimal performance time-series dual-end recovery assessment model to obtain the patient's real-time pelvic floor muscle recovery status assessment result, thereby realizing intelligent assessment of the patient's pelvic floor muscle function.

[0063] Example 2, referring to Figure 1, is based on the above example. Specifically, the raw data collection module collects data through the medical terminal management system and wearable devices to obtain raw data for pelvic floor muscle recovery assessment. This raw data includes historical pelvic floor muscle recovery assessment data and real-time pelvic floor muscle recovery assessment data. Both historical and real-time data include in-hospital interim monitoring data and home continuous monitoring data. The historical data also includes historical recovery assessment levels, which include complete recovery, partial recovery, slow recovery, and poor recovery. The in-hospital interim monitoring data is data obtained during routine follow-up examinations at medical institutions, including imaging parameters. The imaging parameters include pelvic floor muscle thickness, contractility, displacement, and symmetry; the pelvic floor electromyography parameters include muscle contraction and relaxation amplitude, frequency, and duration; the urodynamic parameters include maximum urine flow rate, urethral pressure, and residual urine volume; the home continuous monitoring data includes physiological signal data, behavioral performance data, and patient symptom self-assessment data; the physiological signal data includes the average electromyography amplitude, peak amplitude, frequency, and duration during pelvic floor muscle contraction and relaxation; the behavioral data includes the number of recovery training sessions, recovery training duration, recovery training completion rate, water intake, urination frequency, and bowel movement regularity; the patient symptom self-assessment data includes urinary frequency, urgency, frequency of urinary incontinence, symptom severity, and quality of life score.

[0064] Example 3, referring to Figure 1, is based on the above examples. The original data optimization module specifically performs data cleaning, data standardization, data encoding, and feature filtering on the patient's pelvic floor muscle recovery assessment to obtain optimized data for the patient's pelvic floor muscle recovery assessment; including the following steps:

[0065] Data cleaning is used to process noisy data and missing values ​​in the original data, specifically including outlier removal, missing value handling, data deduplication, and data validation.

[0066] The removal of outliers specifically involves performing statistical distribution analysis on the original data using the interquartile range method to detect outlier data points that deviate from the normal range, and then removing these outlier data points.

[0067] The missing value processing specifically involves filling in missing values ​​in the original data using the nearest neighbor filling method, so that missing records can be repaired while maintaining the reasonable distribution of data.

[0068] The data deduplication process specifically involves identifying and deleting identical records through timestamp matching and data hash verification.

[0069] The data verification is used to verify the rationality of data from different sources. Specifically, it involves using medical logic verification rules to remove data that does not conform to medical principles.

[0070] Data standardization is used to process data of different dimensions and sources in a unified manner. Specifically, it uses the min-max normalization method to map the numerical data of the original data to a unified numerical range.

[0071] Data encoding is used to represent unstructured or categorical data numerically. Specifically, it involves using label encoding methods to encode the category fields in the original data and convert them into structured numerical values.

[0072] Feature screening is used to select the core indicators most relevant to pelvic floor muscle recovery from multi-dimensional features. Specifically, it measures the importance of each feature based on correlation analysis, sets a threshold as a screening criterion, removes redundant features below the threshold, and selects the feature set with the most discriminative power for recovery assessment.

[0073] Example 4, referring to Figures 1, 2, and 4, is based on the above examples. Further, the module for establishing the time-series two-end recovery evaluation model specifically includes the following steps:

[0074] The in-hospital baseline is used to process the in-hospital interim test data obtained during regular follow-up examinations of patients in medical institutions, in order to extract high-precision baseline recovery features that can comprehensively represent the patient's pelvic floor muscle function status; specifically, the in-hospital interim test data is analyzed and features are extracted through a gated bidirectional convolutional neural network to obtain the in-hospital baseline recovery features;

[0075] The bidirectional convolutional units in the gated bidirectional convolutional neural network perform convolution calculations on the input data in both forward and reverse sequences, extract the forward and reverse dependencies between different parameters, and combine the gating mechanism to adaptively suppress noise features and enhance key clinical features, thereby obtaining a baseline representation of in-hospital baseline recovery features that can fully preserve the semantic correlation between different parameters while ensuring the authority of medical testing, and forming a high signal-to-noise ratio and robust baseline representation.

[0076] The home-based continuous time-series endpoint is used to process continuous home monitoring data acquired by patients in their home environment to extract home-based time-series recovery characteristics that dynamically reflect the changing trend of pelvic floor muscle function over time; specifically, it includes the following steps:

[0077] Short-term temporal recovery feature extraction is used to extract recovery features that reflect short-term temporal functional fluctuations of the pelvic floor muscles. Specifically, continuous home monitoring data is input into a bidirectional long short-term memory network. The input data is recursively modeled through forward and backward sequence pathways to capture forward and backward dependencies between adjacent moments in the time series. Short-term information is stored and updated through memory units to obtain short-term temporal recovery features that can dynamically reflect the patient's short-term monitoring period.

[0078] Long-term temporal recovery feature extraction is used to extract long-term recovery features that can reflect the trend of pelvic floor muscle function over time. Specifically, short-term temporal recovery features are input into an improved convolutional residual network to obtain long-term temporal recovery features that can characterize the full-cycle change pattern of the pelvic floor muscle recovery process.

[0079] The improved convolutional residual network specifically involves first inputting short-term temporal recovery features into the first convolutional unit. Preliminary temporal features are extracted through dilated causal convolution, combined with weight normalization and random deactivation. Then, the LeakyReLU activation function enhances the response to negative inputs, generating the first convolutional feature. This first convolutional feature is input into the second convolutional unit, where multi-scale dependencies are extracted through dilated causal convolution and weight normalization. A parallel activation structure using PReLU and hyperbolic tangent functions is implemented. The output is then weighted, fused, and randomly deactivated to generate the second convolutional feature. Finally, the short-term temporal recovery feature is convolved with the second convolutional feature element-wise and added to it to output the long-term temporal recovery feature. The formula used is as follows:

[0080] The kernel size in the dilated causal convolution is set to 7, and the dilation factor is set to 5.

[0081] ;

[0082] ;

[0083] ;

[0084] In the formula, This represents the first convolutional feature. This indicates a random deactivation operation. This represents the LeakyReLU activation function. Indicates short-term time series recovery characteristics. This represents a one-dimensional temporal convolution operation. and This represents the random deactivation probability of two convolutional units, with a set value range. between, This represents the first convolutional kernel after weight normalization. and These represent the weights of the second-layer convolutional kernel, corresponding to the parallel PReLU and Tanh branches, respectively. express Activation function This represents the hyperbolic tangent activation function. This indicates the fusion weights, and the range of values ​​can be set. between, This represents the second convolutional feature. This represents a 1×1 convolution operation. This indicates an element-wise addition operation. Indicates long-term time series recovery characteristics;

[0085] Feature aggregation is used to obtain the final temporal recovery features that comprehensively reflect the patient's pelvic floor muscle function in both short-term fluctuations and long-term trends. Specifically, short-term and long-term temporal recovery features are aggregated using a feature weighting fusion method to obtain the home-based temporal recovery features. The formula used is as follows:

[0086] ;

[0087] In the formula, Indicates the characteristics of the recovery of home-based time sequence, This represents the feature aggregation weight coefficient, and its value range is set. between;

[0088] A dual-end feature fusion layer is used to obtain comprehensive recovery features that simultaneously reflect the static baseline state and dynamic temporal recovery trend of the patient's pelvic floor muscle function. Specifically, the in-hospital baseline recovery features and home temporal recovery features are fused using an attention-based method to obtain comprehensive recovery features. The formula used is as follows:

[0089] ;

[0090] In the formula, Indicates comprehensive recovery characteristics. Indicates baseline recovery characteristics within the hospital. , and These represent the query projection matrix, key projection matrix, and value projection matrix, respectively. Indicates the scaling factor; This represents the transpose operation of the key projection matrix. Indicates feature concatenation operation;

[0091] The assessment result output layer is used to intelligently grade and assess the recovery status of the patient's pelvic floor muscle function. Specifically, the comprehensive recovery features are input into the fully connected layer and mapped to the Softmax function to calculate the probability distribution of each recovery level. Based on the maximum probability judgment strategy, the patient's current recovery level is determined, thereby obtaining the recovery status assessment result.

[0092] ;

[0093] ;

[0094] In the formula, This represents the probability distribution of recovery levels. This represents the weight matrix of the fully connected layer. This represents the bias term parameters of the fully connected layer. This indicates the results of the recovery status assessment.

[0095] By performing the above operations, this solution addresses the technical problems of traditional pelvic floor muscle recovery assessment systems for prostate cancer patients, which suffer from a single assessment dimension, lack of continuous monitoring, and inability to dynamically reflect the changing trend of pelvic floor muscle function over time, leading to inaccurate assessment results. This solution innovatively proposes a time-series dual-end recovery assessment model combining in-hospital baseline and home continuous time-series endpoints. This improves the completeness of pelvic floor muscle recovery assessment, overcomes the shortcomings of traditional single-point detection methods that lead to delayed and biased conclusions, and enhances the real-time and accuracy of pelvic floor muscle recovery assessment. It achieves continuous, intelligent, and full-cycle assessment of pelvic floor muscle function at follow-up points and in daily life. Furthermore, it addresses the limitations of existing recovery assessment models in processing... There is a technical challenge in capturing both short-term fluctuations and long-term trends in dynamic continuous data, leading to inaccurate assessments of pelvic floor muscle recovery. This innovative approach employs a bidirectional long short-term memory network to extract short-term recovery features in a home-based continuous time-series dataset, and combines this with an improved convolutional residual network incorporating parallel activation functions to extract long-term recovery features. This effectively enhances the ability to capture short-term dynamic fluctuations, improves the representation of the patient's long-term recovery process, and overcomes the limitations of traditional methods that rely solely on a single time scale. Consequently, it significantly improves the accuracy and reliability of recovery assessment results, achieving a dynamic, continuous, and precise intelligent assessment of the patient's pelvic floor muscle function recovery status.

[0096] Example 5, referring to Figures 1, 3, and 5, is based on the above examples. The performance optimization module of the recovery evaluation model specifically includes the following steps:

[0097] The initial training of the evaluation model is used to learn parameters and achieve performance convergence of the time-series dual-end recovery evaluation model based on historical data. Specifically, the historical pelvic floor muscle recovery evaluation data processed by the original data optimization module is used as training data to perform initial training of the time-series dual-end recovery evaluation model, resulting in the initial trained time-series dual-end recovery evaluation model.

[0098] The initial training of the time-series two-end recovery evaluation model specifically uses the cross-entropy loss function as the objective function to measure the difference between the predicted probability distribution and the actual recovery level distribution. The weight matrix and bias parameters of the evaluation model are iteratively updated using the backpropagation algorithm and gradient descent optimization method. During the training process, the model parameters are continuously optimized through multiple rounds of iteration. When the preset maximum number of training times is reached or the loss function converges to a set threshold, the iterative training is stopped.

[0099] Adaptive hyperparameter search, specifically, involves obtaining the optimal hyperparameter combination for a time-series two-end recovery evaluation model through an improved optimization algorithm; it includes the following steps:

[0100] Initialize the search individuals by encoding the hyperparameters of the temporal two-end recovery evaluation model into search individual position vectors, and generate N search individual position vectors through a random initialization method. Each individual encoding represents a candidate combination of hyperparameters of the temporal two-end recovery evaluation model, thus obtaining the initial search population.

[0101] The hyperparameter encoding of the temporal dual-end recovery evaluation model includes learning rate, Dropout probability, fusion weight, and feature aggregation weight coefficient;

[0102] The fitness value of a search individual is calculated by calculating the fitness value of the search individual in the population; the performance of the temporal two-end recovery evaluation model based on the location of the search individual is used as the fitness value of the search individual.

[0103] The search process involves dividing individuals into groups, specifically selecting the individual with the best fitness as the leader and the remaining individuals as followers.

[0104] Individual position iterative updates specifically involve updating the leader's and follower's individual positions; this includes the following steps:

[0105] The leader's individual position is updated; specifically, the leader's individual position is updated using the following formula:

[0106] ;

[0107] ;

[0108] In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. This represents the current optimal individual position in the j-th dimension. Indicates the current iteration number. Indicates the maximum number of iterations. They represent Random numbers within a range and These represent the upper and lower bounds of the global search space, respectively. Represents the natural constant. Indicates search control parameters. express Random numbers within a certain range;

[0109] The follower's position is updated by setting the position of each follower to the average of its own position and the position of the previous searched individual; the formula used is as follows:

[0110] ;

[0111] In the formula, Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first During the nth iteration, the 1st The position of a follower individual in the j-th dimension, where i represents the index of the follower individual;

[0112] Dynamic local search is used to prevent the algorithm from getting stuck in local optima during the population evolution process. Specifically, after the individual position is updated iteratively, the fitness value of all search individuals is recalculated, and the individual with the best fitness value is selected as the current iteration's best individual. If the fitness value of the current iteration's best individual is not better than the fitness value of the previous iteration's best individual, the number of unimproved individuals is incremented by one; otherwise, the number of unimproved individuals is reset to zero, and a local search judgment is performed. When the cumulative number of unimproved individuals exceeds four, the local search process is triggered.

[0113] The local search process specifically involves first selecting k search individuals within a preset radius based on Euclidean distance, centered on each search individual, to form its neighborhood set. Within this neighborhood set, the individual with the best fitness value is determined as the optimal neighboring individual. This is then combined with the globally optimal individual to update the local search position, generating local search candidate individuals. If the candidate individual's fitness value is better than the current iteration's optimal individual, it replaces the current iteration's optimal individual, and the unimproved count is reset to zero. Otherwise, the current iteration's optimal individual remains unchanged, and the unimproved count remains constant. The formula used is as follows:

[0114] ;

[0115] In the formula, Indicates the location of candidate individuals in a local search. Indicates the first The current iteration's optimal individual position is obtained after iterative updates. This represents the location of the best individual in the neighborhood. This represents the globally optimal individual position. and They represent Random numbers within a certain range;

[0116] The global optimal position update for a search individual involves evaluating the fitness value of the search individual in the current iteration and comparing it with the global optimal position of the current search individual based on the fitness value of the current best individual. If the fitness value of the current best individual is better, then the global optimal position of the search individual is updated.

[0117] The iterative search terminates when the global optimal position of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached. The global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the evaluation model.

[0118] The final training of the evaluation model involves adjusting the hyperparameters of the initially trained temporal two-end recovery evaluation model based on the optimal hyperparameter combination of the evaluation model, and using historical pelvic floor muscle recovery evaluation data processed by the original data optimization module as training data to retrain the initially trained temporal two-end recovery evaluation model to obtain the optimal performance temporal two-end recovery evaluation model.

[0119] By performing the above operations, this solution addresses the technical problems of existing recovery assessment models, which rely on human experience, have low search efficiency, and are prone to getting trapped in local optima during hyperparameter setting, resulting in insufficient generalization ability and decreased assessment accuracy. It innovatively introduces an improved optimization algorithm that combines a dynamic local search strategy, achieving efficient global exploration and refined local optimization of the assessment model's hyperparameters. This yields the optimal hyperparameter combination for the assessment model, significantly improving the stability and accuracy of the model's output results, and enabling precise and intelligent assessment of the patient's pelvic floor muscle recovery status.

[0120] Example 6, referring to Figure 1, is based on the above examples. The full-cycle intelligent recovery assessment module specifically inputs real-time pelvic floor muscle recovery assessment data into the optimal performance time-series dual-end recovery assessment model to obtain the patient's real-time pelvic floor muscle recovery status assessment result, thereby realizing the full-cycle continuous intelligent assessment of the patient's pelvic floor muscle function.

[0121] The full-cycle continuous intelligent assessment specifically refers to the complete rehabilitation process of the patient from the start of surgery or rehabilitation intervention to the long-term follow-up stage. The follow-up period at different stages corresponds to the time point of data collection for in-hospital phased testing, thereby realizing continuous dynamic assessment of the patient's pelvic floor muscle function at different stages such as early recovery, mid-term consolidation and long-term maintenance.

[0122] The specific follow-up period refers to multiple stages of 2 weeks, 4 weeks, 8 weeks, 12 weeks, 24 weeks, and 48 weeks. Within the same stage, the real-time pelvic floor muscle recovery assessment data consists of both the unchanged in-hospital stage test data and the continuously updated home monitoring data.

[0123] It should be noted that, in this document, 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.

[0124] 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.

[0125] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An AI-based assessment system for pelvic floor muscle recovery in prostate cancer patients, characterized by: The system includes a raw data collection module, a raw data optimization module, a module for establishing a time-series dual-end recovery assessment model, a recovery assessment model performance optimization module, and a full-cycle intelligent recovery assessment module. The raw data collection module specifically collects data to obtain raw data for pelvic floor muscle recovery assessment. The raw data optimization module performs data cleaning, standardization, encoding, and feature filtering on the raw data to obtain optimized data for patient pelvic floor muscle recovery assessment. The module for establishing a time-series dual-end recovery assessment model is used to construct an intelligent assessment model that can simultaneously represent static baseline status and dynamic recovery trends. Specifically, it extracts in-hospital baseline recovery features from the in-hospital baseline and combines bidirectional long short-term memory... The network and an improved convolutional residual network with parallel activation functions are used to extract home-based temporal recovery features at the continuous temporal end. These two types of features are then fused through a dual-end feature fusion layer to generate comprehensive recovery features, which are then input into the evaluation result output layer to obtain the recovery status evaluation result, thus completing the construction of the temporal dual-end recovery evaluation model. Specifically, the steps include: In-hospital baseline end, used to extract high-precision baseline recovery features that comprehensively represent the patient's pelvic floor muscle function status; specifically, using a gated bidirectional convolutional neural network to analyze and extract features from in-hospital phased detection data to obtain in-hospital baseline recovery features; and home-based continuous temporal end, used to extract home-based temporal recovery features that dynamically reflect the changing trend of the patient's pelvic floor muscle function over time. Specifically, the process includes the following steps: short-term temporal recovery feature extraction, specifically by inputting continuous home monitoring data into a bidirectional long short-term memory network to obtain short-term temporal recovery features; long-term temporal recovery feature extraction, specifically by inputting the short-term temporal recovery features into an improved convolutional residual network to obtain long-term temporal recovery features; the improved convolutional residual network specifically involves first inputting the short-term temporal recovery features into the first convolutional unit, extracting preliminary temporal features through dilated causal convolution, combined with weight normalization and random deactivation operations, and then enhancing the response capability to negative inputs through the LeakyReLU activation function to generate the first convolutional feature, which is then input into the second convolutional unit through dilated causal convolution... Multi-scale dependencies are extracted through product and weight normalization. A parallel activation structure using PReLU and hyperbolic tangent functions is set up. The output is weighted and fused and then randomly deactivated to generate a second convolutional feature. Finally, the short-term temporal recovery feature is convolved with the second convolutional feature element-wise to output the long-term temporal recovery feature. Feature aggregation is performed by combining the short-term and long-term temporal recovery features using a feature weighted fusion method to obtain the home temporal recovery feature. The dual-end feature fusion layer is performed by fusing the in-hospital baseline recovery feature and the home temporal recovery feature using an attention-based method to obtain the comprehensive recovery feature.The evaluation result output layer specifically involves inputting comprehensive recovery features into a fully connected layer and mapping them to a Softmax function. It calculates the probability distribution of each recovery level and determines the patient's current recovery level based on a maximum probability judgment strategy, thus obtaining the recovery status evaluation result. The recovery evaluation model performance optimization module first performs preliminary training on the model based on historical data. Then, it obtains the optimal hyperparameter combination of the evaluation model through an improved optimization algorithm incorporating a dynamic local search strategy. Finally, in the final training phase, it retrains the model using the optimal hyperparameter combination to obtain the best-performing temporal dual-end recovery evaluation model. Specifically, the dynamic local search strategy involves recalculating the fitness value of all search individuals after iterative updates of individual positions and selecting the individual with the best fitness value as the current iteration's optimal individual. If the fitness value of the current iteration's optimal individual is not better than... If the fitness value of the best individual in the previous iteration is used, the number of unimproved individuals is incremented by one; otherwise, the number of unimproved individuals is reset to zero, and a local search is performed. When the cumulative number of unimproved individuals exceeds four, a local search process is triggered. Specifically, the local search process involves first selecting k search individuals within a preset radius based on Euclidean distance, centered on each search individual, to form its neighborhood set. Within this neighborhood set, the individual with the best fitness value is determined as the neighborhood best individual of that search individual. The local search position is then updated in conjunction with the globally best individual, generating a local search candidate individual. If the candidate individual's fitness value is better than that of the current iteration's best individual, it replaces the current iteration's best individual, and the number of unimproved individuals is reset to zero. Otherwise, the current iteration's best individual remains unchanged, and the number of unimproved individuals remains the same. The formula used is as follows: In the formula, Indicates the location of candidate individuals in a local search. Indicates the first The current iteration's optimal individual position is obtained after iterative updates. This represents the location of the best individual in the neighborhood. This represents the globally optimal individual position. and They represent The random number within the range; the full-cycle intelligent recovery assessment module specifically inputs real-time pelvic floor muscle recovery assessment data into the optimal performance time-series dual-end recovery assessment model to obtain the patient's real-time pelvic floor muscle recovery status assessment result, realizing the full-cycle continuous intelligent assessment of the patient's pelvic floor muscle function.

2. The artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients according to claim 1, characterized in that: The performance optimization module for the recovery assessment model specifically includes the following steps: preliminary training of the assessment model, specifically using historical pelvic floor muscle recovery assessment data processed by the original data optimization module as training data to perform preliminary training of the time-series dual-end recovery assessment model, resulting in a preliminary trained time-series dual-end recovery assessment model; adaptive hyperparameter search, specifically obtaining the optimal hyperparameter combination of the time-series dual-end recovery assessment model by introducing an optimization algorithm improved by a dynamic local search strategy; and final training of the assessment model, specifically adjusting the hyperparameters of the preliminary trained time-series dual-end recovery assessment model based on the optimal hyperparameter combination of the assessment model, and retraining the preliminary trained time-series dual-end recovery assessment model using historical pelvic floor muscle recovery assessment data processed by the original data optimization module as training data, to obtain the time-series dual-end recovery assessment model with optimal performance.

3. The artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients according to claim 2, characterized in that: The adaptive hyperparameter search specifically includes the following steps: initializing search individuals, specifically encoding the hyperparameters of the temporal two-ends recovery evaluation model into search individual position vectors, and generating N search individual position vectors through a random initialization method, where each individual represents a candidate combination of hyperparameters for the temporal two-ends recovery evaluation model, thus obtaining an initial search population; calculating the fitness value of search individuals, specifically calculating the fitness value of search individuals in the population; using the performance of the temporal two-ends recovery evaluation model established based on the search individual positions as the fitness value of the search individuals; partitioning search individuals, specifically selecting the individual with the best fitness value as the leader individual, and the remaining individuals as followers individuals; iteratively updating individual positions, specifically updating the leader individual position and the follower individual position; including the following step: leader individual position update, specifically updating the leader individual position; the formula used is as follows: In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. This indicates the position of the current best individual in the j-th dimension. They represent Random numbers within a range and These represent the upper and lower bounds of the global search space, respectively. Indicates search control parameters. express The search involves several steps: random numbers within a given range; follower individual position updates, specifically updating the position of each follower individual to the average of its own position and the position of the previous search individual; dynamic local search; updating the global optimal position of the search individual, specifically evaluating the fitness value of the search individual in the current iteration and comparing it with the global optimal position of the current search individual based on the fitness value of the current iteration's best individual. If the fitness value of the current iteration's best individual is better, then the global optimal position of the search individual is updated; iterative search termination, specifically terminating the search and obtaining the global optimal position of the search individual when the global optimal position of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached. The global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the evaluation model.

4. The artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients according to claim 1, characterized in that: The raw data collection module specifically collects data through the medical terminal management system and wearable devices to obtain raw data for pelvic floor muscle recovery assessment. The raw data for patient pelvic floor muscle recovery assessment includes historical pelvic floor muscle recovery assessment data and real-time pelvic floor muscle recovery assessment data. Both the historical and real-time pelvic floor muscle recovery assessment data include in-hospital periodic test data and home continuous monitoring data. The historical pelvic floor muscle recovery assessment data also includes historical recovery assessment levels.

5. The artificial intelligence-based pelvic floor muscle recovery assessment system for prostate cancer patients according to claim 1, characterized in that: The original data optimization module specifically includes the following steps: data cleaning, specifically outlier removal, missing value handling, data deduplication, and data validation; data standardization, specifically using the min-max normalization method to map the numerical data of the original data to a unified numerical range; data encoding, specifically using a label encoding method to encode the category fields in the original data, converting them into structured numerical values; and feature selection, specifically measuring the importance of each feature based on correlation analysis, setting a threshold as a selection criterion, removing redundant features below the threshold, and selecting the feature set with the most discriminative power for recovery assessment.

Images