Intelligent digital therapy recommendation system based on generative adversarial network
The digital therapy recommendation system, which integrates conditional generative adversarial networks and jellyfish search algorithms, solves the problems of personalization and dynamic adaptation in existing systems, realizes the accurate generation and dynamic optimization of personalized therapy content, and improves user experience and intervention effects.
Patent Information
- Application Number
- CN202510656923.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-09-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing digital therapy recommendation systems lack the ability to learn users' real-time status and individual responses, resulting in serious homogeneity of recommended content and significant individual differences in efficacy, which cannot meet the needs of personalization and dynamic adaptation.
By integrating the conditional generative adversarial network model and the jellyfish search optimization algorithm, we build user portraits by collecting multimodal health data, generate personalized therapy content, and optimize the model through a feedback update mechanism to achieve a personalized and dynamic recommendation system.
The accuracy and personalized adaptation capabilities of therapy content have been improved. The system has self-optimization and iteration capabilities and can respond to user changes in real time, thereby improving user stickiness and intervention effects.
Smart Images
Figure CN120596735A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence and medical information technology, and in particular to an intelligent digital therapy recommendation system based on a generative adversarial network. Background Art
[0002] In the modern healthcare service system, digital therapy, as an emerging non-drug intervention, has gradually been applied to multiple fields such as chronic disease management, mental health intervention, and behavioral regulation. Digital therapy (DTx) refers to the provision of evidence-based treatment plans through software programs. Its core lies in the use of digital methods to carry out continuous, personalized, and quantifiable health interventions. Currently, with the development of smart devices, wearable technology, and big data sensing methods, the scope of digital therapy is constantly expanding, and its system capabilities are becoming increasingly rich. However, before digital therapy can truly be implemented as a scalable, repeatable, and evaluable medical intervention path, a key challenge is how to achieve accurate and dynamic personalized therapy content generation and matching recommendations based on the user's multimodal health characteristics.
[0003] Existing digital therapy recommendation systems are generally based on traditional algorithms such as rule engines, knowledge base matching, or collaborative filtering. These systems typically use fixed labels or demographic parameters (such as age, gender, and disease type) as the classification basis, and promote universal intervention plans through manually set matching rules. This recommendation mechanism lacks the ability to learn from users' real-time status, behavioral preferences, and intervention feedback, and is unable to adapt to fluctuations in users' psychological states and individual responses. This leads to severe homogeneity in recommended content and significant individual differences in treatment efficacy, seriously affecting the effectiveness of digital therapy interventions and user stickiness.
[0004] In recent years, Generative Adversarial Networks (GANs) have been widely used in tasks such as image generation, speech synthesis, and text generation, demonstrating powerful sample generation capabilities. Conditional GANs (cGANs), as an improved form of GANs, further enhance the controllability and semantic relevance of generated samples by introducing conditional information. Studies have begun to explore the application of cGANs in medical content generation and health data enhancement. However, in digital therapy recommendation scenarios, cGAN models still suffer from problems such as unstable training, insufficient content diversity, and difficulty in adaptively adjusting structural parameters. Especially when it comes to complex user portraits and semantic intervention tasks, traditional fixed-structure cGAN models struggle to achieve a good balance between accuracy, diversity, and real-time performance, resulting in recommendation results that cannot effectively reflect individual health characteristics and changing trends.
[0005] On the other hand, most current deep learning models for generating content rely heavily on manual experience for structural parameter configuration, making it difficult to adaptively optimize based on task requirements and user feedback. This structural rigidity limits the model's generalization and personalized response capabilities. In medical intervention systems, even slightly inappropriate parameter settings can cause the generated content to deviate from the intervention goal or lack specificity, affecting user trust and efficacy evaluation. In digital therapies, in particular, user feedback information is lagging and multidimensional, making it difficult to meet dynamic adaptation requirements through static training alone.
[0006] To address these issues, some research has attempted to introduce evolutionary optimization algorithms (such as genetic algorithms and particle swarm optimization) to assist in neural network architecture search and parameter tuning. However, these algorithms suffer from local optimality traps, insufficient global search capabilities, and slow convergence efficiency, making them difficult to operate stably in high-dimensional nonlinear optimization problems. Furthermore, most existing optimization strategies are disconnected from deep generative models and lack a mechanism for synergizing the model's internal parameters with the system's external feedback data, making it difficult to form a truly closed-loop optimization recommendation system.
[0007] Therefore, how to provide an intelligent digital therapy recommendation system based on generative adversarial networks is an urgent problem that those skilled in the art need to solve. Summary of the Invention
[0008] One purpose of the present invention is to propose an intelligent digital therapy recommendation system based on a generative adversarial network. The present invention integrates the conditional generative adversarial network model and the jellyfish search optimization algorithm, and describes in detail the entire process of generating and recommending personalized therapy content driven by the user's multimodal health data. It has the advantages of high accuracy in intervention content generation, strong personalized adaptation capabilities, and strong system self-optimization and iteration capabilities.
[0009] An intelligent digital therapy recommendation system based on a generative adversarial network according to an embodiment of the present invention includes:
[0010] The data processing module is used to collect the user's multimodal health data and perform cleaning, normalization and feature encoding to generate user feature labels;
[0011] The user profile module is used to build user profiles based on processed health data as conditional input to the generator;
[0012] Generator building module, used to build the generator in the conditional generative adversarial network and generate personalized digital therapy content based on user profiles;
[0013] The optimization control module is used to apply the jellyfish search algorithm to optimize the generator's structural parameters and training hyperparameters to improve the quality of content generation;
[0014] The recommendation matching module is used to match the generated content and output the recommendation solution based on the user status, behavior preferences and historical feedback;
[0015] The feedback update module is used to collect user feedback data and use it to update training samples and optimize algorithm strategies.
[0016] Optionally, modules can be connected using the following methods:
[0017] S1. Collect the user's multimodal health data, preprocess the multimodal health data, and construct an input feature set containing user feature labels;
[0018] S2. Construct a conditional generative adversarial network model, which includes a generator and a discriminator. The generator generates digital therapy content based on user feature labels, and the discriminator is used to distinguish between real content and generated content and provide a generative adversarial feedback signal.
[0019] S3. Optimize the generator parameters in the conditional generative adversarial network model through the jellyfish search algorithm. At the same time, introduce the simulated annealing perturbation mechanism in the iterative process to apply probabilistic perturbations to the current optimal solution to obtain the optimal generator parameter configuration;
[0020] S4. Utilize the optimized generator to combine multimodal health data to generate personalized digital therapy content, perform multi-dimensional matching based on user status, interaction preferences, and historical feedback scores, and output the optimal digital therapy recommendation plan;
[0021] S5. Collect users' actual usage feedback and effect evaluation data on recommended content to form a feedback dataset. The feedback data is used to update the training samples of the conditional generative adversarial network model and dynamically adjust the fitness function parameters and search strategy in the jellyfish search algorithm.
[0022] Optionally, the multimodal health data specifically includes the user's behavioral data, physiological monitoring data, and psychological assessment data, which are used to construct portrait information characterizing the user's health status as a conditional input for generating an adversarial network.
[0023] Optionally, the preprocessing of multimodal health data specifically includes data cleaning, missing value filling, normalization and feature vector encoding, which is used to improve the quality of multimodal health data and generate an input feature set in a unified format for processing by the conditional generative adversarial network model.
[0024] Optionally, the S2 specifically includes:
[0025] S21. Setting a user feature label vector c, wherein the user feature label vector is composed of a user profile generated after preprocessing multimodal health data, and is used as a conditional input to guide the generator to generate content;
[0026] S22. Construct a generator network G(z,c), where is a noise vector that follows a standard normal distribution. The generator input is a joint vector [z,c] and the output is a digital therapy content sample.
[0027] S23. Construct a discriminator network D(x,c), where x represents a real digital therapy sample or a generated sample. The discriminator determines the match between x and the user feature label vector c and outputs a true or false judgment value D(x,c);
[0028] S24. Define the adversarial loss function for:
[0029]
[0030] Among them, p data (x,c) represents the joint distribution of the real sample and its conditional label, p z (z) represents the noise distribution, p c (c) represents the prior distribution of conditional labels, Represents the joint distribution p from the real sample and the condition pair data The expected value obtained by sampling in (x,c), log2 is the logarithmic function, Represents the noise distribution p z (z) and the conditional distribution p c (c) is the expected value of the sample, G(z,c) is the generator function;
[0031] S25, introduce a multi-scale feature generation module into the generator, and set the multi-scale feature level number parameter N ms , extract local detail features and global structural information at the same time, and realize multi-scale information fusion through the pyramid convolution nested structure;
[0032] S26: embed a label-guided attention mechanism in the generator structure and set the attention embedding depth parameter D att , inject the user feature label vector c into the generator network D att Layer, which adjusts the response strength of different regions in the intermediate feature map through the conditional attention matrix;
[0033] S27. Set feature fusion ratio coefficient α fuse , used to control the local feature output F local and the global feature output F global The fusion ratio of , the final output result is:
[0034] F out =α fuse ·F local+(1-α fuse )·F global ;
[0035] S28, generator parameter θ G and the discriminator parameters θ D Alternating optimization training is performed to make the content generated by the generator under a given conditional user feature label vector more realistic and user-adaptive.
[0036] Optionally, the S3 specifically includes:
[0037] S31. Set the initial population size of the jellyfish search algorithm to P, and encode each jellyfish individual as a parameter vector v i =[N ms ,D att ,α fuse ], where N ms represents the number of multi-scale feature levels, D att represents the attention embedding depth parameter, α fuse Represents the feature fusion ratio coefficient;
[0038] S32, define the fitness function f(v i ) as individual performance evaluation indicators, where:
[0039]
[0040] in, is the adversarial loss function of the conditional generative adversarial network, Div(G(z,c)) is the diversity measure of the generated content, Match(G(z,c),c) is the semantic matching score between the generated content and the conditional label c, and λ1, λ2, and λ3 are weighting coefficients used to balance the three sub-goals;
[0041] S33. In each iteration, a global search is performed based on the "drifting with the flow" strategy of the jellyfish search algorithm.
[0042] Generate updated individual vector v′ i :
[0043] v′ i =v i +r·(gv i );
[0044] Among them, v i is the current jellyfish individual, g is the individual with the best fitness in the current population, and r is the drift ratio coefficient, which controls the update step size;
[0045] S34, execute the "active foraging" strategy of the jellyfish search algorithm, perform local perturbation search near the optimal individual, generate candidate solutions and pass the fitness function f(v′i ) to select and retain;
[0046] S35, introduce simulated annealing perturbation mechanism in each update process, set the temperature parameter T k With the annealing probability p k Perform disturbance acceptance judgment and accept the inferior solution when the following conditions are met:
[0047]
[0048] Among them, T k is the annealing temperature of the kth generation, exp() represents the exponential function, f(v′ i )-f(v i ) is the fitness change;
[0049] S36, after reaching the preset number of iterations or fitness convergence threshold, select the best individual v * As the final parameter configuration of the generator, it is used to generate personalized digital therapy content.
[0050] Optionally, the S4 specifically includes:
[0051] S41. Load the generator parameter configuration optimized by the jellyfish search algorithm, initialize the generator structure in the conditional generative adversarial network model, and prepare the input data set;
[0052] S42, calling the pre-processed multimodal health data, including the user's behavioral characteristics, physiological state information and psychological assessment labels, as a conditional input vector and inputting it into the generator;
[0053] S43. Based on the current user profile, use the generator to generate personalized digital therapy content, including cognitive intervention scripts, emotion regulation suggestions, and behavioral training tasks;
[0054] S44. Collect the user's current usage status information, including system access time, physiological state change trends and psychological response types, and extract behavioral preference characteristics during the user's interaction with the system;
[0055] S45. Combining the generated therapy content, user status, behavioral preferences, and historical feedback records, executing a multi-dimensional matching strategy, and evaluating content suitability through a similarity measurement mechanism;
[0056] S46. Filter and output the optimal digital therapy recommendation plan based on the evaluation results, and present personalized intervention content that is highly matched with the user's current health profile.
[0057] Optionally, the S5 specifically includes:
[0058] S51. After the user completes the use of the recommended digital therapy content, collect usage behavior data, including interactive feature information such as launch frequency, usage duration, operation sequence, and bounce rate;
[0059] S52. Synchronously collect a series of health status feedback data of the user after the intervention, including subjective assessment results and objective indicator changes;
[0060] S53. Structurally integrate the user's usage behavior data and efficacy feedback data to form a standardized feedback data set, and index and label it according to the user ID;
[0061] S54. Based on the newly formed feedback dataset, incrementally update the training sample set in the conditional generative adversarial network model to incorporate real interaction and efficacy results;
[0062] S55. Based on the matching degree changes, diversity responses, and user satisfaction performance recorded in the feedback data, dynamically adjust the fitness function parameter settings in the jellyfish search algorithm and update the evaluation weight of the optimization target;
[0063] S56. Adjust the search strategy in the jellyfish search algorithm based on the latest feedback trends and changes in user performance, including adjusting the search range, convergence step size, and perturbation probability.
[0064] The beneficial effects of the present invention are:
[0065] The intelligent digital therapy recommendation system based on a generative adversarial network proposed in this invention can effectively overcome technical problems existing in existing digital therapy recommendation technologies, such as content homogeneity, lack of dynamic adaptability, and insufficient response to individual differences. The system constructs personalized portraits by collecting multimodal health data from users and introduces a conditional generative adversarial network model. The generator integrates a multi-scale feature extraction structure and a label-guided attention mechanism, so that the generated therapy content has both rich semantic expression capabilities and can accurately fit the user's current health status and intervention needs. At the same time, the system uses a jellyfish search algorithm to globally optimize the generator's structural parameters and hyperparameters, improving the stability, diversity, and consistency of content generation, and avoiding the performance limitations caused by the fixed structure of traditional generative models.
[0066] During the recommendation execution phase, the present invention not only combines user behavioral preferences, historical efficacy feedback, and state change information to perform multi-dimensional matching of generated content, but also establishes a feedback-driven model update mechanism. By collecting users' actual usage data of recommended content and intervention effect evaluation results, the system can update training samples in real time and dynamically adjust the fitness function and search strategy of the optimization algorithm, thereby achieving self-evolution and continuous optimization of the recommendation system. This closed-loop process of "generation-recommendation-feedback-re-optimization" enables the present invention to not only generate highly personalized intervention content, but also have the intelligent decision-making ability to dynamically respond to user changes.
[0067] In summary, the present invention has achieved innovative breakthroughs in model structure, optimization mechanism and system feedback strategy, which can significantly improve the comprehensive performance of the digital therapy system in terms of personalization, effectiveness and intelligent adaptation, and has a wide range of application values in clinical auxiliary intervention, mental health management and intelligent rehabilitation recommendation. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0069] Figure 1 This is a flowchart of an intelligent digital therapy recommendation system based on generative adversarial networks proposed in the present invention;
[0070] Figure 2 This is a structural diagram of the conditional generative adversarial network model of the intelligent digital therapy recommendation system based on generative adversarial networks proposed in the present invention. DETAILED DESCRIPTION
[0071] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.
[0072] refer to Figure 1 and Figure 2 , an intelligent digital therapy recommendation system based on generative adversarial networks, including:
[0073] The data processing module is used to collect the user's multimodal health data and perform cleaning, normalization and feature encoding to generate user feature labels;
[0074] The user profile module is used to build user profiles based on processed health data as conditional input to the generator;
[0075] Generator building module, used to build the generator in the conditional generative adversarial network and generate personalized digital therapy content based on user profiles;
[0076] The optimization control module is used to apply the jellyfish search algorithm to optimize the generator's structural parameters and training hyperparameters to improve the quality of content generation;
[0077] The recommendation matching module is used to match the generated content and output the recommendation solution based on the user status, behavior preferences and historical feedback;
[0078] The feedback update module is used to collect user feedback data and use it to update training samples and optimize algorithm strategies.
[0079] In this embodiment, the modules are connected through the following methods:
[0080] S1. Collect the user's multimodal health data, preprocess the multimodal health data, and construct an input feature set containing user feature labels;
[0081] S2. Construct a conditional generative adversarial network model, which includes a generator and a discriminator. The generator generates digital therapy content based on user feature labels, and the discriminator is used to distinguish between real content and generated content and provide a generative adversarial feedback signal.
[0082] S3. Optimize the generator parameters in the conditional generative adversarial network model through the jellyfish search algorithm. At the same time, introduce the simulated annealing perturbation mechanism in the iterative process to apply probabilistic perturbations to the current optimal solution to obtain the optimal generator parameter configuration;
[0083] S4. Utilize the optimized generator to combine multimodal health data to generate personalized digital therapy content, perform multi-dimensional matching based on user status, interaction preferences, and historical feedback scores, and output the optimal digital therapy recommendation plan;
[0084] S5. Collect users' actual usage feedback and effect evaluation data on recommended content to form a feedback dataset. The feedback data is used to update the training samples of the conditional generative adversarial network model and dynamically adjust the fitness function parameters and search strategy in the jellyfish search algorithm.
[0085] In this embodiment, the multimodal health data specifically includes the user's behavioral data, physiological monitoring data, and psychological assessment data, which are used to construct portrait information representing the user's health status as a conditional input for generating an adversarial network.
[0086] In this embodiment, the preprocessing of multimodal health data specifically includes data cleaning, missing value filling, normalization and feature vector encoding, which is used to improve the quality of multimodal health data and generate an input feature set in a unified format for conditional generative adversarial network model processing.
[0087] In this embodiment, S2 specifically includes:
[0088] S21. Setting a user feature label vector c, wherein the user feature label vector is composed of a user profile generated after preprocessing multimodal health data, and is used as a conditional input to guide the generator to generate content;
[0089] S22. Construct a generator network G(z,c), where is a noise vector that follows a standard normal distribution. The generator input is a joint vector [z,c] and the output is a digital therapy content sample.
[0090] S23. Construct a discriminator network D(x,c), where x represents a real digital therapy sample or a generated sample. The discriminator determines the match between x and the user feature label vector c and outputs a true or false judgment value D(x,c);
[0091] S24. Define the adversarial loss function for:
[0092]
[0093] Among them, p data (x,c) represents the joint distribution of the real sample and its conditional label, p z (z) represents the noise distribution, p c (c) represents the prior distribution of conditional labels, Represents the joint distribution p from the real sample and the condition pair data The expected value obtained by sampling in (x,c), log2 is the logarithmic function, Represents the noise distribution p z (z) and the conditional distribution p c (c) is the expected value of the sample, G(z,c) is the generator function;
[0094] S25, introduce a multi-scale feature generation module into the generator, and set the multi-scale feature level number parameter N ms , extract local detail features and global structural information at the same time, and realize multi-scale information fusion through the pyramid convolution nested structure;
[0095] S26: embed a label-guided attention mechanism in the generator structure and set the attention embedding depth parameter D att , inject the user feature label vector c into the generator network D att Layer, which adjusts the response strength of different regions in the intermediate feature map through the conditional attention matrix;
[0096] S27. Set feature fusion ratio coefficient α fuse , used to control the local feature output F localand the global feature output F global The fusion ratio of , the final output result is:
[0097] F out =α fuse ·F local +(1-α fuse )·F global ;
[0098] S28, generator parameter θ G and the discriminator parameters θ D Alternating optimization training is performed to make the content generated by the generator under a given conditional user feature label vector more realistic and user-adaptive.
[0099] In this embodiment, S3 specifically includes:
[0100] S31. Set the initial population size of the jellyfish search algorithm to P, and encode each jellyfish individual as a parameter vector v i =[N ms ,D att ,α fuse ], where N ms represents the number of multi-scale feature levels, D att represents the attention embedding depth parameter, α fuse Represents the feature fusion ratio coefficient;
[0101] S32, define the fitness function f(v i ) as individual performance evaluation indicators, where:
[0102]
[0103] in, is the adversarial loss function of the conditional generative adversarial network, Div(G(z,c)) is the diversity measure of the generated content, Match(G(z,c),c) is the semantic matching score between the generated content and the conditional label c, and λ1, λ2, and λ3 are weighting coefficients used to balance the three sub-goals;
[0104] S33. In each iteration, a global search is performed according to the “drifting with the flow” strategy of the jellyfish search algorithm to generate an updated individual vector v′ i :
[0105] v′ i =v i +r·(gv i );
[0106] Among them, v i is the current jellyfish individual, g is the individual with the best fitness in the current population, and r is the drift ratio coefficient, which controls the update step size;
[0107] S34, execute the "active foraging" strategy of the jellyfish search algorithm, perform local perturbation search near the optimal individual, generate candidate solutions and pass the fitness function f(v′ i ) to select and retain;
[0108] S35, introduce simulated annealing perturbation mechanism in each update process, set the temperature parameter T k With the annealing probability p k Perform disturbance acceptance judgment and accept the inferior solution when the following conditions are met:
[0109]
[0110] Among them, T k is the annealing temperature of the kth generation, exp() represents the exponential function, f(v′ i )-f(v i ) is the fitness change;
[0111] S36, after reaching the preset number of iterations or fitness convergence threshold, select the best individual v * As the final parameter configuration of the generator, it is used to generate personalized digital therapy content.
[0112] In this embodiment, the S4 specifically includes:
[0113] S41. Load the generator parameter configuration optimized by the jellyfish search algorithm, initialize the generator structure in the conditional generative adversarial network model, and prepare the input data set;
[0114] S42, calling the pre-processed multimodal health data, including the user's behavioral characteristics, physiological state information and psychological assessment labels, as a conditional input vector and inputting it into the generator;
[0115] S43. Based on the current user profile, use the generator to generate personalized digital therapy content, including cognitive intervention scripts, emotion regulation suggestions, and behavioral training tasks;
[0116] S44. Collect the user's current usage status information, including system access time, physiological state change trends and psychological response types, and extract behavioral preference characteristics during the user's interaction with the system;
[0117] S45. Combining the generated therapy content, user status, behavioral preferences, and historical feedback records, executing a multi-dimensional matching strategy, and evaluating content suitability through a similarity measurement mechanism;
[0118] S46. Filter and output the optimal digital therapy recommendation plan based on the evaluation results, and present personalized intervention content that is highly matched with the user's current health profile.
[0119] In this embodiment, the S5 specifically includes:
[0120] S51. After the user completes the use of the recommended digital therapy content, collect usage behavior data, including interactive feature information such as launch frequency, usage duration, operation sequence, and bounce rate;
[0121] S52. Synchronously collect a series of health status feedback data of the user after the intervention, including subjective assessment results and objective indicator changes;
[0122] S53. Structurally integrate the user's usage behavior data and efficacy feedback data to form a standardized feedback data set, and index and label it according to the user ID;
[0123] S54. Based on the newly formed feedback dataset, incrementally update the training sample set in the conditional generative adversarial network model to incorporate real interaction and efficacy results;
[0124] S55. Based on the matching degree changes, diversity responses, and user satisfaction performance recorded in the feedback data, dynamically adjust the fitness function parameter settings in the jellyfish search algorithm and update the evaluation weight of the optimization target;
[0125] S56. Adjust the search strategy in the jellyfish search algorithm based on the latest feedback trends and changes in user performance, including adjusting the search range, convergence step size, and perturbation probability.
[0126] Example 1:
[0127] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the psychological counseling project of a community health service center. At the beginning of 2024, a push system based on a fixed content library was adopted. The platform provided users with unified templates for therapy content, such as "universal meditation audio" and "basic cognitive training text". Despite the wide coverage, the system cannot perceive the actual status of the user, nor can it adjust the content based on individual data. According to the statistics of the project in the first quarter of 2024, there were 168 registered users on the platform, with an average of only 48 active users per day. The average completion rate of therapy content use was 45%, and the satisfaction score was below 3.2 points (out of 5 points). Users' willingness to continue participating was insufficient.
[0128] To address this issue, we introduced our proposed "Intelligent Digital Therapy Recommendation System Based on Generative Adversarial Networks" at the community center for a three-month pilot program starting in April 2024. The system is easy to access: after logging into the mini-program, users submit multimodal health data by answering questions and uploading wearable device data. The system automatically collects information such as sleep scores, heart rate variability, activity frequency, and self-rated mood.
[0129] The system first uses a preprocessing module to standardize this data and construct user profiles, which serve as conditional inputs to guide content generation. Subsequently, a deployed conditional generative adversarial network model, combined with an optimized generator structure, generates personalized intervention content. The generator structure was optimized using the jellyfish search algorithm during initial deployment to ensure optimal generation capabilities. Consequently, the intervention content users receive is no longer a fixed template; instead, it is dynamically generated by the model with diverse content, such as "Three-Stage Pre-Sleep Meditation Guide" and "Anxiety Identification and Shifting Thinking Exercises."
[0130] Each time a user completes an intervention, the system automatically collects behavioral feedback (such as usage time and bounce rate) and self-assessment of effectiveness (such as the soothing index and concentration improvement score). Based on this feedback data, the system updates the model's training samples and optimization strategies. This feedback-driven mechanism ensures that the system continuously learns and responds to user changes, achieving closed-loop optimization.
[0131] During the pilot period from April to June 2024, a total of 156 users continuously participated in the system, and the system operated stably, interacted naturally, and provided rich content. We conducted a comparative evaluation of key indicators before and after the system's operation, measuring the application value of the system from multiple dimensions such as usage activity, completion rate, satisfaction, and intervention effectiveness.
[0132] Table 1 Comparison of key indicators before and after the pilot of the intelligent digital therapy recommendation system
[0133]
[0134] According to the data in Table 1, it can be clearly seen that the system proposed in the present invention has achieved significant improvements in intervention effect, user experience, and usage activity in actual applications, verifying the scientificity and practicality of the system design.
[0135] First, in terms of user activity, the average daily active users was 48 before the system was introduced, but this number increased to 73 after the introduction of the system, a 52.1% increase. This change demonstrates that the optimized recommendation mechanism and personalized therapy generation model have significantly increased user login frequency and willingness to participate, resulting in more frequent user interaction with the system and stronger stickiness.
[0136] Secondly, regarding the execution of therapeutic interventions, the average completion rate of intervention content jumped from 45.2% to 79.6%, a 76.2% increase. This demonstrates that personalized therapeutic content not only better aligns with the user's current status and needs, but also significantly increases user engagement and sustained completion, avoiding the user fatigue and interruptions associated with traditional, fixed content.
[0137] In terms of subjective user experience, satisfaction scores increased from 3.2 to 4.5, a 40.6% increase. This indicates a significant improvement in user acceptance of content format, push cadence, and intervention experience, demonstrating the system's advantages in contextual adaptability and user-friendly human-computer interaction.
[0138] More importantly, in terms of intervention effectiveness, users' anxiety scores (GAD-7) decreased by an average of 106.2%, from 1.6 to 3.3 points; their sleep scores (PSQI) also improved by 133.3%, from 0.9 to 2.1 points. These data demonstrate that the system's recommended therapies are not only superior in form but also deliver effective interventions in mental health, further demonstrating the generator's accuracy in content personalization.
[0139] Finally, in terms of user stickiness, the usage rate for three weeks or more doubled from 28.6% to 61.5% to 114.7%. This demonstrates that users not only have a positive initial experience but are also willing to participate in the intervention process over the long term, forming stable behavioral habits and laying the foundation for long-term health management.
[0140] In summary, the pilot results clearly show that the system of the present invention is superior to the traditional recommendation model in many key indicators, especially in improving personalized experience, enhancing intervention effectiveness and system intelligent adaptive capabilities, and has shown significant advantages, and has extremely strong promotion value and practicality.
[0141] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.
Claims
1. An intelligent digital therapy recommendation system based on generative adversarial networks, characterized by: include: The data processing module is used to collect the user's multimodal health data and perform cleaning, normalization and feature encoding to generate user feature labels; The user profile module is used to build user profiles based on processed health data as conditional input to the generator; Generator building module, used to build the generator in the conditional generative adversarial network and generate personalized digital therapy content based on user profiles; The optimization control module is used to apply the jellyfish search algorithm to optimize the generator's structural parameters and training hyperparameters to improve the quality of content generation; The recommendation matching module is used to match the generated content and output the recommendation solution based on the user status, behavior preferences and historical feedback; The feedback update module is used to collect user feedback data and use it to update training samples and optimize algorithm strategies.
2. A method for recommending intelligent digital therapy based on a generative adversarial network, applied to the intelligent digital therapy recommendation system based on a generative adversarial network according to claim 1, characterized in that: The steps include: S1. Collect the user's multimodal health data, preprocess the multimodal health data, and construct an input feature set containing user feature labels; S2. Construct a conditional generative adversarial network model, which includes a generator and a discriminator. The generator generates digital therapy content based on user feature labels, and the discriminator is used to distinguish between real content and generated content and provide a generative adversarial feedback signal. S3. Optimize the generator parameters in the conditional generative adversarial network model through the jellyfish search algorithm. At the same time, introduce the simulated annealing perturbation mechanism in the iterative process to apply probabilistic perturbations to the current optimal solution to obtain the optimal generator parameter configuration; S4. Utilize the optimized generator to combine multimodal health data to generate personalized digital therapy content, perform multi-dimensional matching based on user status, interaction preferences, and historical feedback scores, and output the optimal digital therapy recommendation plan; S5. Collect users' actual usage feedback and effect evaluation data on recommended content to form a feedback dataset. The feedback data is used to update the training samples of the conditional generative adversarial network model and dynamically adjust the fitness function parameters and search strategy in the jellyfish search algorithm.
3. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 2, characterized in that: The multimodal health data specifically includes the user's behavioral data, physiological monitoring data and psychological assessment data, which are used to construct portrait information representing the user's health status as conditional input for the generative adversarial network.
4. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 2, characterized in that: The preprocessing of multimodal health data specifically includes data cleaning, missing value filling, normalization and feature vector encoding, which is used to improve the quality of multimodal health data and generate an input feature set in a unified format for conditional generative adversarial network model processing.
5. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 2, characterized in that: The S2 specifically includes: S21. Setting a user feature label vector c, wherein the user feature label vector is composed of a user profile generated after preprocessing multimodal health data, and is used as a conditional input to guide the generator to generate content; S22. Construct a generator network G(z,c), where is a noise vector that follows a standard normal distribution. The generator input is a joint vector [z,c] and the output is a digital therapy content sample. S23. Construct a discriminator network D(x,c), where x represents a real digital therapy sample or a generated sample. The discriminator determines the match between x and the user feature label vector c and outputs a true or false judgment value D(x,c); S24. Define the adversarial loss function for: Among them, p data (x,c) represents the joint distribution of the real sample and its conditional label, p z (z) represents the noise distribution, p c (c) represents the prior distribution of conditional labels, Represents the joint distribution p from the real sample and the condition pair data The expected value obtained by sampling in (x,c), log2 is the logarithmic function, Represents the noise distribution p z (z) and the conditional distribution p c (c) is the expected value of the sample, G(z,c) is the generator function; S25, introduce a multi-scale feature generation module into the generator, and set the multi-scale feature level number parameter N ms , extract local detail features and global structural information at the same time, and realize multi-scale information fusion through the pyramid convolution nested structure; S26: embed a label-guided attention mechanism in the generator structure and set the attention embedding depth parameter D att , inject the user feature label vector c into the generator network D att Layer, which adjusts the response strength of different regions in the intermediate feature map through the conditional attention matrix; S27. Set feature fusion ratio coefficient α fuse , used to control the local feature output F local and the global feature output F global The fusion ratio of , the final output result is: F out =a fuse ·F local +(1-a fuse )·F global ; S28, generator parameter θ G and the discriminator parameters θ D Alternating optimization training is performed to make the content generated by the generator under a given conditional user feature label vector more realistic and user-adaptive.
6. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 5, characterized in that: The S3 specifically includes: S31. Set the initial population size of the jellyfish search algorithm to P, and encode each jellyfish individual as a parameter vector v i =[N ms ,D att ,α fuse ], where N ms represents the number of multi-scale feature levels, D att represents the attention embedding depth parameter, α fuse Represents the feature fusion ratio coefficient; S32, define the fitness function f(v i ) as individual performance evaluation indicators, where: in, is the adversarial loss function of the conditional generative adversarial network, Div(G(z,c)) is the diversity measure of the generated content, Match(G(z,c),c) is the semantic matching score between the generated content and the conditional label c, and λ1, λ2, and λ3 are weighting coefficients used to balance the three sub-goals; S33. In each iteration, a global search is performed according to the "drift with the flow" strategy of the jellyfish search algorithm to generate an updated individual vector v′ i : v′ i =v i +r·(gv i ); Among them, v i is the current jellyfish individual, g is the individual with the best fitness in the current population, and r is the drift ratio coefficient, which controls the update step size; S34, execute the "active foraging" strategy of the jellyfish search algorithm, perform local perturbation search near the optimal individual, generate candidate solutions and pass the fitness function f(v′ i ) to select and retain; S35, introduce simulated annealing perturbation mechanism in each update process, set the temperature parameter T k With the annealing probability p k Perform disturbance acceptance judgment and accept the inferior solution when the following conditions are met: Among them, T k is the annealing temperature of the kth generation, exp() represents the exponential function, f(v′ i )-f(v i ) is the fitness change; S36, after reaching the preset number of iterations or fitness convergence threshold, select the best individual v * As the final parameter configuration of the generator, it is used to generate personalized digital therapy content.
7. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 6, characterized in that: The S4 specifically includes: S41. Load the generator parameter configuration optimized by the jellyfish search algorithm, initialize the generator structure in the conditional generative adversarial network model, and prepare the input data set; S42, calling the pre-processed multimodal health data, including the user's behavioral characteristics, physiological state information and psychological assessment labels, as a conditional input vector and inputting it into the generator; S43. Based on the current user profile, use the generator to generate personalized digital therapy content, including cognitive intervention scripts, emotion regulation suggestions, and behavioral training tasks; S44. Collect the user's current usage status information, including system access time, physiological state change trends and psychological response types, and extract behavioral preference characteristics during the user's interaction with the system; S45. Combining the generated therapy content, user status, behavioral preferences, and historical feedback records, executing a multi-dimensional matching strategy, and evaluating content suitability through a similarity measurement mechanism; S46. Filter and output the optimal digital therapy recommendation plan based on the evaluation results, and present personalized intervention content that is highly matched with the user's current health profile.
8. The method for recommending intelligent digital therapy based on generative adversarial networks according to claim 7, characterized in that: The S5 specifically includes: S51. After the user completes the use of the recommended digital therapy content, collect usage behavior data, including interactive feature information such as launch frequency, usage duration, operation sequence, and bounce rate; S52. Synchronously collect a series of health status feedback data of the user after the intervention, including subjective assessment results and objective indicator changes; S53. Structurally integrate the user's usage behavior data and efficacy feedback data to form a standardized feedback data set, and index and label it according to the user ID; S54. Based on the newly formed feedback dataset, incrementally update the training sample set in the conditional generative adversarial network model to incorporate real interaction and efficacy results; S55. Based on the matching degree changes, diversity responses, and user satisfaction performance recorded in the feedback data, dynamically adjust the fitness function parameter settings in the jellyfish search algorithm and update the evaluation weight of the optimization target; S56. Adjust the search strategy in the jellyfish search algorithm based on the latest feedback trends and changes in user performance, including adjusting the search range, convergence step size, and perturbation probability.