Intelligent follow-up visit system and method based on AI digital twinning of patient

Abstract

The invention discloses an intelligent follow-up visit system and method based on AI digital twinning of patients, and belongs to the technical field of intelligent follow-up visit of digital twinning. Comprising a patient data acquisition module, an AI twinborn model construction module, a personalized follow-up strategy generation module, a multi-modal risk assessment module, an intelligent interaction execution module, an effect assessment and feedback module and a model continuous optimization module. According to the method, the patient-specific AI digital twinborn model is constructed, the follow-up visit frequency, content and opportunity are dynamically optimized in combination with the reinforcement learning algorithm, meanwhile, the optimal follow-up visit time is determined by matching the living habits of the patient according to the K-nearest neighbor algorithm, the problem that traditional standardized follow-up visit conflicts with the living of the patient is avoided, and the multi-modal emotion recognition technology is matched. The AI can understand the emotion of the patient and generate a co-estrus response, the mechanical interaction experience is improved, the patient can accept follow-up visit more easily, finally the follow-up visit completion rate can be increased to 60% or above from traditional less than 10%, and medical management caused by follow-up visit deficiency is greatly reduced.

Description

An intelligent follow-up system and method based on patient AI digital twins Technical Field

[0001] This invention belongs to the field of intelligent follow-up technology of digital twins, specifically referring to an intelligent follow-up system and method based on patient AI digital twins. Background Technology

[0002] In the field of healthcare, follow-up management is a key link in ensuring postoperative recovery, chronic disease control, and health risk intervention for patients. In particular, for patients with chronic diseases (such as diabetes and hypertension), postoperative patients, and special disease populations, regular follow-up can effectively reduce the probability of disease deterioration and improve treatment outcomes. However, current mainstream medical follow-up systems and methods still have many technical bottlenecks and application defects, making it difficult to meet the personalized needs of patients and the requirements for high-quality development of medical services.

[0003] Traditional follow-up adopts a standardized "one-size-fits-all" approach, failing to consider individual patient differences in frequency, content, and timing. For example, patients of different ages and with different diseases have varying health needs and rehabilitation priorities, resulting in a follow-up completion rate generally below 10% and poor patient compliance. Secondly, the system lacks long-term memory capabilities; each follow-up interaction is independent, making it impossible to store related patient historical information, symptom trends, and preferences, leading to "data fragmentation." This makes it difficult to build a complete patient profile, increases communication costs, and makes follow-up decisions lack historical data support. Furthermore, risk management relies primarily on passive alerts, depending solely on preset rules to issue warnings when indicators exceed limits. It cannot combine historical patient data and real-time status (such as daily routines and emotions) to dynamically predict risks, and the reliance on single judgment rules leads to delayed warnings, low accuracy, and a tendency to miss the optimal intervention opportunity. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent follow-up system and method based on patient AI digital twins to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent follow-up system and method based on patient AI digital twins, comprising: a patient data acquisition module, an AI twin model construction module, a personalized follow-up strategy generation module, a multimodal risk assessment module, an intelligent interaction execution module, an effect evaluation and feedback module, and a model continuous optimization module;

[0006] The patient data acquisition module is used to collect multi-dimensional data such as patients' physiological indicators, behavioral data, dialogue history, symptom changes, medication response, disease knowledge, treatment experience, and behavioral patterns.

[0007] The AI ​​twin model construction module employs an LSTM neural network and a memory network to construct a three-layer memory system: short-term memory, medium-term memory, and long-term memory. Short-term memory stores immediate information such as recent conversation history, symptom changes, and medication responses. Medium-term memory integrates stage-specific data such as disease knowledge, treatment experience, and behavioral patterns. Long-term memory constructs a model of the patient's disease progression and personality traits based on long-term historical data. Through progressive information compression and priority retention mechanisms, it achieves unlimited contextual memory. The information retention priority is calculated as follows:

[0008] ,

[0009] in, For the first Priority of retaining each piece of information , , These are time weight, information importance weight, and information relevance weight, respectively. + + =1, This is a normalized value representing the information's present time. Score the importance of the information. A score is assigned to assess the relevance of the information to the patient's current disease status.

[0010] The personalized follow-up strategy generation module is based on the multidimensional features of the patient's AI twin and uses reinforcement learning algorithms to dynamically optimize the follow-up frequency, content and method to determine the best follow-up time.

[0011] The multimodal risk assessment module integrates multimodal information such as physiological indicators, behavioral data, and emotional states, uses a time-series prediction model to analyze the changing trends of health indicators, identifies abnormal signs based on the historical patterns of the patient's AI twin, and establishes a graded early warning mechanism.

[0012] The intelligent interactive execution module combines technologies such as voice emotion analysis and text emotion recognition to understand the patient's emotional state and provide personalized care and communication.

[0013] The effect evaluation and feedback module is used to evaluate the follow-up effect and feed the evaluation results back to other modules; the model continuous optimization module continuously optimizes the AI ​​twin model and follow-up strategy based on the feedback results.

[0014] In the personalized follow-up strategy generation module, when dynamically optimizing the follow-up frequency using a reinforcement learning algorithm, patient follow-up compliance, degree of health improvement, and satisfaction are used as key parameters of the reward function. The reward function formula is as follows:

[0015] ,

[0016] in, To enhance the immediate reward value of learning, , , These are, respectively, the weights for compliance, health improvement, and satisfaction. + + =1, To ensure patient follow-up compliance, To the extent of improvement in the patient's health Rate patient satisfaction.

[0017] The multimodal risk assessment module uses a time-series forecasting model to analyze the changing trends of health indicators and employs the ARIMA model to predict short-term health indicators. The prediction formula is as follows:

[0018] ,

[0019] in, For the first Predicted values ​​of health indicators at any given time. , , , These are the autoregressive coefficients. , , , Historical health indicator values, , , , The moving average coefficient is... , , , This is the historical error term. This represents the error term at the current time.

[0020] In the intelligent interaction execution module, the voice emotion recognition extracts features such as tone, speech rate, and pauses, and uses a support vector machine model for emotion classification. The classification decision function is as follows:

[0021] ,

[0022] in, For the emotion classification results, The speech feature vector to be classified. For Lagrange multipliers, For the emotion labels of the training samples, For kernel function, For classification threshold, This represents the number of training samples.

[0023] This includes a privacy protection module that uses end-to-end encryption technology to encrypt and store patient data. It employs federated learning technology to optimize the model while protecting privacy, and ensures patient data security through blockchain-based access control and audit trails. The blockchain access verification formula is as follows:

[0024] ,

[0025] in, This is the access permission verification value. For user identification, Preset permission levels for users. This is the current access timestamp. For hash functions, For XOR operation, Preset a key for the system, when When the value matches the verification value stored in the blockchain, the user gains the corresponding access privileges.

[0026] Specifically, in the AI ​​twin model construction module, when constructing the patient's disease progression trajectory using long-term memory, a temporal convolutional network is employed to process the long-term history. Feature extraction is performed on historical data, and the feature extraction formula is as follows:

[0027] ,

[0028] in, This represents the feature vector of the disease progression trajectory in long-term memory. , , , For patients Historical data for each time period, for The model's weight parameters, This is the feature extraction function for temporal convolutional networks.

[0029] In the personalized follow-up strategy generation module, the personalized time window prediction uses the K-nearest neighbor algorithm to determine the optimal follow-up time based on the patient's lifestyle and response patterns, and the similarity calculation uses the Euclidean distance formula.

[0030] ,

[0031] in, The feature vector of the time window to be predicted With historical time window feature vector Euclidean distance, For feature dimension, , They are respectively , The value at the k-th feature dimension is selected based on the minimum distance. Using the follow-up timing corresponding to a historical time window as a reference, the current optimal follow-up timing is determined.

[0032] The tiered early warning mechanism of the multimodal risk assessment module classifies mild alerts, moderate concerns, and severe interventions based on risk values, which are calculated as follows:

[0033] ,

[0034] in, For the patient's health risk value, For the first Risk weights for each health indicator For the first The degree to which a health indicator deviates from the normal range, The number of health indicators used in risk assessment; when A mild alert is triggered when the value is less than 0.3; a warning is triggered when the value is less than or equal to 0.3. When <0.7, moderate attention is triggered; when A value ≥0.7 triggers a severe intervention.

[0035] In the effect evaluation and feedback module, the follow-up effect evaluation adopts a comprehensive scoring system, and the comprehensive scoring formula is as follows:

[0036] ,

[0037] in, To evaluate the overall follow-up effectiveness, , , These are the completion rate weight, the effect weight, and the feedback weight, respectively. + + =1, To achieve the follow-up completion rate, Rate the health improvement effect. Rate patient feedback.

[0038] This includes the following steps:

[0039] Step 1: Collect multi-dimensional patient data through the patient data collection module, including physiological indicators, behavioral data, dialogue history, symptom changes, medication response, disease knowledge, treatment experience, behavioral patterns, etc.

[0040] Step 2: The AI ​​twin model construction module uses LSTM neural network and memory network, combined with the information retention priority calculation method, to construct a three-layer memory system consisting of short-term memory, medium-term memory and long-term memory, forming a patient AI twin model;

[0041] Step 3: The personalized follow-up strategy generation module uses the multi-dimensional features of the patient's AI twin, employs reinforcement learning algorithms, and combines the aforementioned reward function to dynamically optimize the follow-up frequency, content, and method. It then combines the aforementioned Euclidean distance similarity calculation to determine the optimal follow-up timing and generate a personalized follow-up plan.

[0042] Step 4: The multimodal risk assessment module integrates multimodal information, uses the ARIMA model to predict the trend of health indicator changes, identifies abnormal signs, and then combines the risk value calculation to establish a graded early warning mechanism to provide early warning of potential health risks.

[0043] Step 5: The intelligent interaction execution module interacts with the patient according to the personalized follow-up plan, and performs voice emotion recognition and text sentiment analysis in combination with the SVM model to understand the patient's emotional state and provide warm and personalized care.

[0044] Step 6: The effect evaluation and feedback module evaluates the follow-up effect according to the comprehensive scoring formula and feeds the evaluation results back to the model continuous optimization module;

[0045] Step 7: The model continuous optimization module adjusts and optimizes the AI ​​twin model and follow-up strategy based on the feedback results. At the same time, the privacy protection module ensures the security of patient data in accordance with the blockchain access permission verification method, and realizes the iterative optimization of the follow-up system.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. This invention constructs a patient-specific AI digital twin model, dynamically optimizes the frequency, content, and timing of follow-up visits using reinforcement learning algorithms, and determines the optimal follow-up time based on the K-nearest neighbor algorithm to match the patient's lifestyle habits. This avoids the problem of conflicts between traditional standardized follow-up visits and the patient's life. Combined with multimodal emotion recognition technology, the AI ​​can understand the patient's emotions and generate empathetic responses, improving the mechanical interaction experience and making follow-up visits more acceptable to patients. Ultimately, the follow-up completion rate can be increased from less than 10% to over 60%, while patient satisfaction can be increased to over 95%. This significantly reduces medical management loopholes caused by missing follow-ups and improves the overall efficiency of medical services.

[0048] 2. This invention integrates a patient's physiological indicators, behavioral data, and emotional state into a multimodal risk assessment module. Through a time-series prediction model, it predicts the trend of health indicator changes 24-72 hours in advance. Combined with a risk value grading mechanism: risk value <0.3 indicates mild warning, 0.3 ≤ risk value <0.7 indicates moderate attention, and risk value ≥0.7 indicates severe intervention, it achieves proactive identification and differentiated intervention for potential health risks. This mechanism allows medical staff to intervene before the risk worsens, effectively reducing the situation of worsening condition due to delayed risk discovery.

[0049] 3. This invention prevents data storage leakage by employing end-to-end encryption technology, achieves "data remains still while the model moves" through federated learning, and combines blockchain technology to enable controllable access permissions and traceable behavior. This not only avoids the "data silo" problem of traditional systems but also solves the privacy leakage risk in data sharing. The system can fully explore the value of patient data, optimize AI twin models and follow-up strategies, while eliminating patients' concerns about data privacy, laying a foundation of trust for the long-term promotion and application of the system. Attached Figure Description

[0050] Figure 1 is a flowchart of the operation of an intelligent follow-up system and method based on patient AI digital twin according to the present invention;

[0051] Figure 2 is a flowchart of the operation of an intelligent follow-up system and method based on patient AI digital twin according to the present invention.

[0052] Figure 3 is a flowchart of the operation of an intelligent follow-up system and method based on patient AI digital twin according to the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example

[0055] Please refer to Figures 1-3. The present invention provides a technical solution: an intelligent follow-up system based on patient AI digital twins, characterized in that it includes: a patient data acquisition module, an AI twin model construction module, a personalized follow-up strategy generation module, a multimodal risk assessment module, an intelligent interaction execution module, an effect evaluation and feedback module, and a model continuous optimization module.

[0056] The patient data acquisition module is used to collect multi-dimensional data such as patients' physiological indicators, behavioral data, dialogue history, symptom changes, medication response, disease knowledge, treatment experience, and behavioral patterns.

[0057] The AI ​​twin model construction module employs an LSTM neural network and a memory network to construct a three-layer memory system: short-term memory, medium-term memory, and long-term memory. Short-term memory stores immediate information such as recent conversation history, symptom changes, and medication responses. Medium-term memory integrates stage-specific data such as disease knowledge, treatment experience, and behavioral patterns. Long-term memory constructs a model of the patient's disease progression and personality traits based on long-term historical data. Through progressive information compression and priority retention mechanisms, it achieves unlimited contextual memory. The information retention priority is calculated as follows:

[0058] ,

[0059] in, For the first Priority of retaining each piece of information , , These are time weight, information importance weight, and information relevance weight, respectively. + + =1, This is a normalized value representing the information's present time. Score the importance of the information. A score is assigned to assess the relevance of the information to the patient's current disease status.

[0060] The personalized follow-up strategy generation module is based on the multidimensional features of the patient's AI twin and uses reinforcement learning algorithms to dynamically optimize the follow-up frequency, content and method to determine the best follow-up time.

[0061] The multimodal risk assessment module integrates multimodal information such as physiological indicators, behavioral data, and emotional states, uses a time-series prediction model to analyze the changing trends of health indicators, identifies abnormal signs based on the historical patterns of the patient's AI twin, and establishes a graded early warning mechanism.

[0062] The intelligent interactive execution module combines technologies such as voice emotion analysis and text emotion recognition to understand the patient's emotional state and provide personalized care and communication.

[0063] The effect evaluation and feedback module is used to evaluate the follow-up effect and feed the evaluation results back to other modules; the model continuous optimization module continuously optimizes the AI ​​twin model and follow-up strategy based on the feedback results.

[0064] It should be noted that the patient data acquisition module collects multi-dimensional patient data and transmits it to the AI ​​twin model construction module. This module uses LSTM and memory networks to build a three-layer memory system, and filters key information through an information priority retention mechanism to form a patient AI twin model. Subsequent modules (personalized follow-up, multimodal risk assessment, etc.) carry out corresponding work based on the model output. The effect evaluation module feeds the results back to the model continuous optimization module to achieve system iteration, thereby realizing the comprehensive integration of patient data and the construction of a personalized AI twin model. This provides data and model support for subsequent functions such as accurate follow-up and risk warning, breaking the limitations of traditional follow-up systems with scattered data and no personalized models.

[0065] In the personalized follow-up strategy generation module, when dynamically optimizing the follow-up frequency using a reinforcement learning algorithm, patient follow-up compliance, degree of health improvement, and satisfaction are used as key parameters of the reward function. The reward function formula is as follows:

[0066] ,

[0067] in, To enhance the immediate reward value of learning, , , These are, respectively, the weights for compliance, health improvement, and satisfaction. + + =1, To ensure patient follow-up compliance, To the extent of improvement in the patient's health Rate patient satisfaction.

[0068] It should also be understood that, with patient follow-up compliance, degree of health improvement, and satisfaction as core parameters, the system calculates immediate reward values ​​for reinforcement learning through preset weights. The system dynamically adjusts the follow-up frequency, content, and method based on the reward values, optimizing the follow-up strategy towards increasing reward values. This makes the follow-up frequency more aligned with the patient's actual situation, improves patient follow-up compliance, promotes patient health improvement, enhances patient satisfaction with follow-up services, and solves the problem of poor adaptability of traditional standardized follow-up frequencies.

[0069] The multimodal risk assessment module uses a time-series forecasting model to analyze the changing trends of health indicators and employs the ARIMA model to predict short-term health indicators. The prediction formula is as follows:

[0070] ,

[0071] in, For the first Predicted values ​​of health indicators at any given time. , , , These are the autoregressive coefficients. , , , Historical health indicator values, , , , The moving average coefficient is... , , , For historical error terms, This represents the error term at the current time.

[0072] In addition, by inputting patients' historical health indicator data, the ARIMA model uses autoregressive coefficients to mine the correlation of historical indicators and combines moving average coefficients to correct errors, predicting health indicators at a future point in time, capturing the trend of indicator changes, and predicting changes in patients' health indicators in advance. This helps medical staff to discover potential health risks as early as possible, prevent the risks from worsening, and make up for the shortcomings of traditional follow-up that rely solely on real-time data and cannot make predictions in advance.

[0073] In the intelligent interaction execution module, the voice emotion recognition extracts features such as tone, speech rate, and pauses, and uses a support vector machine model for emotion classification. The classification decision function is as follows:

[0074] ,

[0075] in, For the emotion classification results, The speech feature vector to be classified. For Lagrange multipliers, For the emotion labels of the training samples, For kernel function, For classification threshold, This represents the number of training samples.

[0076] It should also be noted that by extracting features such as tone, speech rate, and pauses from the patient's speech, a feature vector is formed and input into the SVM model. Based on the emotion labels and Lagrange multipliers of the training samples, the model determines the patient's emotion category through kernel function calculation and classification threshold judgment, accurately identifies the patient's emotional state, and enables the system to adjust the interaction mode according to the emotion, thereby providing communication and care that matches the patient's psychological state, improving the mechanical interaction experience of traditional follow-up, and increasing the patient's acceptance of follow-up.

[0077] This includes a privacy protection module that uses end-to-end encryption technology to encrypt and store patient data. It employs federated learning technology to optimize the model while protecting privacy, and ensures patient data security through blockchain-based access control and audit trails. The blockchain access verification formula is as follows:

[0078] ,

[0079] in, This is the access permission verification value. For user identification, Preset permission levels for users. This is the current access timestamp. For hash functions, For XOR operation, Preset a key for the system, when When the value matches the verification value stored in the blockchain, the user gains the corresponding access privileges.

[0080] It should also be understood that patient data is stored using end-to-end encryption; during the model optimization phase, federated learning is used, with training conducted locally on each data node, transmitting only model parameters; when a user accesses the system, an access permission verification value is generated using a hash function and an XOR operation, which is then compared with the value stored on the blockchain. Access is granted only after successful verification, comprehensively ensuring patient data security, preventing data leakage and unauthorized access, eliminating patients' concerns about data privacy, and simultaneously achieving model optimization while protecting privacy, thus balancing data security and system performance.

[0081] Specifically, in the AI ​​twin model construction module, when constructing the patient's disease progression trajectory using long-term memory, a temporal convolutional network is employed to process the long-term history. Feature extraction is performed on historical data, and the feature extraction formula is as follows:

[0082] ,

[0083] in, This represents the feature vector of the disease progression trajectory in long-term memory. , , , For patients Historical data for each time period, for The model's weight parameters, This is the feature extraction function for temporal convolutional networks.

[0084] It should be noted that by inputting long-term historical data of patients from multiple time periods into the TCN model, the model performs convolution operations on the data through weight parameters, extracts the temporal features in the data, constructs feature vectors of the patient's disease development trajectory and personality characteristics, forms long-term memory, accurately mines key information in the patient's long-term historical data, clearly presents the disease development pattern and the patient's personality characteristics, and makes the AI ​​twin model more closely match the patient's real situation, providing in-depth data support for subsequent personalized services.

[0085] In the personalized follow-up strategy generation module, the personalized time window prediction uses the K-nearest neighbor algorithm to determine the optimal follow-up time based on the patient's lifestyle and response patterns, and the similarity calculation uses the Euclidean distance formula.

[0086] ,

[0087] in, The feature vector of the time window to be predicted With historical time window feature vector Euclidean distance, For feature dimension, , They are respectively , The value at the k-th feature dimension is selected based on the minimum distance. Using the follow-up timing corresponding to a historical time window as a reference, the current optimal follow-up timing is determined.

[0088] In addition, feature vectors of the time window to be predicted and historical time windows are constructed, the Euclidean distance between the two is calculated, and the follow-up timings corresponding to the K historical windows with the smallest distances are selected. The optimal follow-up timing is determined comprehensively, so that the follow-up timing is highly matched with the patient's lifestyle and response pattern, reducing the interference of follow-up on the patient's daily life, increasing the probability of the patient participating in follow-up on time, and solving the problem that the traditional follow-up timing is fixed and easily conflicts with the patient's life.

[0089] The tiered early warning mechanism of the multimodal risk assessment module classifies mild alerts, moderate concerns, and severe interventions based on risk values, which are calculated as follows:

[0090] ,

[0091] in, For the patient's health risk value, For the first Risk weights of each health indicator For the first The degree to which a health indicator deviates from the normal range, The number of health indicators used in risk assessment; when A mild alert is triggered when the value is less than 0.3; a warning is triggered when the value is less than or equal to 0.3. When <0.7, moderate attention is triggered; when A value ≥0.7 triggers a severe intervention.

[0092] It should also be understood that risk weights are set according to the importance of health indicators, the degree to which each indicator deviates from the normal range is calculated, and the risk value is obtained by weighted summation. Based on the range of the risk value (<0.3, 0.3≤risk value<0.7, ≥0.7), corresponding mild reminders, moderate attention, and severe intervention warnings are triggered. Differentiated treatment is carried out for different levels of health risks. Mild risks reduce unnecessary medical resource input, while severe risks ensure timely intervention, thereby improving the efficiency and accuracy of risk management and avoiding overreaction or underreaction of traditional warning mechanisms.

[0093] In the effect evaluation and feedback module, the follow-up effect evaluation adopts a comprehensive scoring system, and the comprehensive scoring formula is as follows:

[0094] ,

[0095] in, To evaluate the overall follow-up effectiveness, , , These are the completion rate weight, the effect weight, and the feedback weight, respectively. + + =1, To achieve the follow-up completion rate, Rate the health improvement effect. Rate patient feedback.

[0096] In addition, the follow-up completion rate, health improvement effect score, and patient feedback score are weighted and summed according to their weights to obtain a comprehensive follow-up effect score. This quantitatively evaluates the quality of follow-up services, thereby objectively and comprehensively assessing the follow-up effect, clarifying the advantages and disadvantages of follow-up services, providing a clear direction for improvement for the continuous optimization module of the model, and promoting the continuous improvement of follow-up service quality.

[0097] This includes the following steps:

[0098] Step 1: Collect multi-dimensional patient data through the patient data collection module, including physiological indicators, behavioral data, dialogue history, symptom changes, medication response, disease knowledge, treatment experience, behavioral patterns, etc.

[0099] Step 2: The AI ​​twin model construction module uses LSTM neural network and memory network, combined with the information retention priority calculation method, to construct a three-layer memory system consisting of short-term memory, medium-term memory and long-term memory, forming a patient AI twin model;

[0100] Step 3: The personalized follow-up strategy generation module uses the multi-dimensional features of the patient's AI twin, employs reinforcement learning algorithms, and combines the aforementioned reward function to dynamically optimize the follow-up frequency, content, and method. It then combines the aforementioned Euclidean distance similarity calculation to determine the optimal follow-up timing and generate a personalized follow-up plan.

[0101] Step 4: The multimodal risk assessment module integrates multimodal information, uses the ARIMA model to predict the trend of health indicator changes, identifies abnormal signs, and then combines the risk value calculation to establish a graded early warning mechanism to provide early warning of potential health risks.

[0102] Step 5: The intelligent interaction execution module interacts with the patient according to the personalized follow-up plan, and performs voice emotion recognition and text sentiment analysis in combination with the SVM model to understand the patient's emotional state and provide warm and personalized care.

[0103] Step 6: The effect evaluation and feedback module evaluates the follow-up effect according to the comprehensive scoring formula and feeds the evaluation results back to the model continuous optimization module;

[0104] Step 7: The model continuous optimization module adjusts and optimizes the AI ​​twin model and follow-up strategy based on the feedback results. At the same time, the privacy protection module ensures the security of patient data in accordance with the blockchain access permission verification method, and realizes the iterative optimization of the follow-up system.

[0105] It should be noted that the patient data collection, AI twin model construction, personalized follow-up plan generation, multimodal risk warning, intelligent interactive execution, effect evaluation and feedback, and model optimization are completed step by step. Each step is interconnected, forming a closed-loop follow-up workflow. This achieves intelligent and personalized follow-up throughout the entire process, forming a complete closed loop from data collection to service optimization. This significantly improves the follow-up completion rate, the accuracy of risk warning, and patient satisfaction, reduces medical disputes, increases the follow-up rate, and comprehensively improves the level of follow-up services.

[0106] Working principle: The patient data collection module collects information such as the patient's physiological indicators, behavioral data, and dialogue history from multiple dimensions, and transmits it to the AI ​​twin model construction module;

[0107] This module uses LSTM neural networks and memory networks to build a three-layer memory system of short-term, medium-term and long-term, combined with progressive information compression and priority retention mechanisms to extract the patient's disease development trajectory and individual characteristics, and build a unique AI digital twin model.

[0108] The personalized follow-up strategy generation module uses reinforcement learning algorithms to optimize the frequency and content of follow-up based on the multidimensional features of patients output by the model, and combines the K-nearest neighbor algorithm to determine the best follow-up time and generate a personalized plan; the multimodal risk assessment module integrates multi-source information, predicts the trend of health indicators through time series prediction models, and realizes graded early warning based on risk values.

[0109] The intelligent interaction execution module uses voice and text emotion recognition technology to judge the patient's emotions, generate empathetic responses and adjust communication methods to complete personalized interactions. After the follow-up, the effect evaluation and feedback module quantifies the follow-up effect with a comprehensive scoring system and feeds the results back to the model continuous optimization module, which updates the AI ​​twin model parameters and follow-up strategies accordingly.

[0110] Meanwhile, the privacy protection module ensures data security throughout the process through end-to-end encryption, federated learning, and blockchain access control, ultimately forming a complete closed loop of "data collection - model building - strategy generation - risk warning - interactive execution - effect feedback - optimization iteration", realizing personalized, predictive, and secure intelligent follow-up management throughout the entire process.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0112] 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 intelligent follow-up system based on patient AI digital twins, characterized in that, include: The system comprises a patient data acquisition module, an AI twin model construction module, a personalized follow-up strategy generation module, a multimodal risk assessment module, an intelligent interaction execution module, an effect evaluation and feedback module, and a model continuous optimization module. The patient data acquisition module collects multi-dimensional data from patients, including physiological indicators, behavioral data, dialogue history, symptom changes, medication responses, disease knowledge, treatment experience, and behavioral patterns. The AI ​​twin model construction module uses an LSTM neural network and a memory network to construct a three-layer memory system: short-term memory, medium-term memory, and long-term memory. Short-term memory stores immediate information such as recent dialogue history, symptom changes, and medication responses. Medium-term memory integrates stage-specific data such as disease knowledge, treatment experience, and behavioral patterns. Long-term memory constructs a model of the patient's disease progression and individual characteristics based on long-term historical data. Through progressive information compression and priority retention mechanisms, it achieves unlimited contextual memory. The information retention priority calculation method is as follows: ,in, For the first Priority of retaining each piece of information 、 、 These are time weight, information importance weight, and information relevance weight, respectively. + + =1， This is a normalized value representing the information's present time. Score the importance of the information. The system assigns a correlation score between information and the patient's current disease status; the personalized follow-up strategy generation module dynamically optimizes the follow-up frequency, content, and method based on the multidimensional features of the patient's AI twin using reinforcement learning algorithms to determine the optimal follow-up timing; the multimodal risk assessment module integrates multimodal information such as physiological indicators, behavioral data, and emotional states, uses a time-series prediction model to analyze the changing trends of health indicators, identifies abnormal signs based on the historical patterns of the patient's AI twin, and establishes a graded early warning mechanism; the intelligent interaction execution module combines technologies such as voice emotion analysis and text emotion recognition to understand the patient's emotional state and provide personalized care and communication; the effect evaluation and feedback module evaluates the follow-up effect and feeds the evaluation results back to other modules; the model continuous optimization module continuously optimizes the AI ​​twin model and follow-up strategy based on the feedback results.

2. The intelligent follow-up system based on patient AI digital twin as described in claim 1, characterized in that: In the personalized follow-up strategy generation module, when dynamically optimizing the follow-up frequency using a reinforcement learning algorithm, patient follow-up compliance, degree of health improvement, and satisfaction are used as key parameters of the reward function. The reward function formula is as follows: ,in, To enhance the immediate reward value of learning, 、 、 These are, respectively, the weights for compliance, health improvement, and satisfaction. + + =1， To ensure patient follow-up compliance, To the extent of improvement in the patient's health Rate patient satisfaction.

3. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the multimodal risk assessment module, a time-series forecasting model is used to analyze the changing trends of health indicators, and an ARIMA model is used to predict short-term health indicators. The prediction formula is as follows: ,in, For the first Predicted values ​​of health indicators at any given time. ， ， ， These are the autoregressive coefficients. ， ， ， Historical health indicator values, ， ， ， The moving average coefficient is... ， ， ， This is the historical error term. This represents the error term at the current time.

4. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the intelligent interaction execution module, the voice emotion recognition extracts features such as tone, speech rate, and pauses, and uses a support vector machine model for emotion classification. The classification decision function is as follows: ,in, For the emotion classification results, The speech feature vector to be classified. For Lagrange multipliers, For the emotion labels of the training samples, For kernel function, For classification threshold, This represents the number of training samples.

5. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: It also includes a privacy protection module, which uses end-to-end encryption technology to encrypt and store patient data, employs federated learning technology to optimize the model while protecting privacy, and ensures patient data security through blockchain-based access control and audit trails. The blockchain access verification formula is as follows: ,in, This is the access permission verification value. For user identification, Preset permission levels for users. This is the current access timestamp. For hash functions, For XOR operation, Preset a key for the system, when When the value matches the verification value stored in the blockchain, the user gains the corresponding access privileges.

6. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the AI ​​twin model construction module, when constructing the patient's disease progression trajectory using long-term memory, a temporal convolutional network is employed to process the long-term history. Feature extraction is performed on historical data, and the feature extraction formula is as follows: ,in, This represents the feature vector of the disease progression trajectory in long-term memory. ， ， ， For patients Historical data for each time period, for The model's weight parameters, This is the feature extraction function for temporal convolutional networks.

7. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the personalized follow-up strategy generation module, the personalized time window prediction uses the K-nearest neighbor algorithm to determine the optimal follow-up timing based on the patient's lifestyle and response patterns, and the similarity calculation uses the Euclidean distance formula. ,in, The feature vector of the time window to be predicted With historical time window feature vector Euclidean distance, For feature dimension, 、 They are respectively 、 The value at the k-th feature dimension is selected based on the minimum distance. Using the follow-up timing corresponding to a historical time window as a reference, the current optimal follow-up timing is determined.

8. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the tiered early warning mechanism of the multimodal risk assessment module, the classification of mild alert, moderate attention, and severe intervention is based on risk value, which is calculated as follows: ,in, For the patient's health risk value, For the first Risk weights of each health indicator For the first The degree to which a health indicator deviates from the normal range, The number of health indicators used in risk assessment; when A mild alert is triggered when the value is less than 0.3; a warning is triggered when the value is less than or equal to 0.

3. When <0.7, moderate attention is triggered; when A value ≥0.7 triggers a severe intervention.

9. The intelligent follow-up system based on patient AI digital twin according to claim 1, characterized in that: In the effect evaluation and feedback module, the follow-up effect evaluation adopts a comprehensive scoring system, and the comprehensive scoring formula is as follows: ,in, To evaluate the overall follow-up effectiveness, 、 、 These are respectively the completion rate weight, the effect weight, and the feedback weight, and + + =1， To achieve the follow-up completion rate, Rate the health improvement effect. Rate patient feedback.

10. An intelligent follow-up method based on patient AI digital twins, characterized in that, The process includes the following steps: Step 1: Collect multi-dimensional patient data through the patient data acquisition module, including physiological indicators, behavioral data, dialogue history, symptom changes, medication response, disease knowledge, treatment experience, and behavioral patterns; Step 2: The AI ​​twin model construction module uses LSTM neural networks and memory networks, combined with the aforementioned information retention priority calculation method, to construct a three-layer memory system consisting of short-term memory, medium-term memory, and long-term memory, forming a patient AI twin model; Step 3: The personalized follow-up strategy generation module, based on the multi-dimensional features of the patient AI twin, uses reinforcement learning algorithms, combined with the aforementioned reward function, to dynamically optimize the follow-up frequency, content, and method, and then, combined with the aforementioned Euclidean distance similarity calculation, determines the optimal follow-up timing and generates a personalized follow-up plan; Step 4: The multimodal risk assessment module integrates multimodal information, uses the ARIMA model to predict the trend of health indicator changes, identifies abnormal signs, and then combines the risk value calculation to establish a graded early warning mechanism to provide early warning of potential health risks. Step 5: The intelligent interaction execution module interacts with the patient according to the personalized follow-up plan, and uses the SVM model to perform voice emotion recognition and text sentiment analysis to understand the patient's emotional state and provide warm, personalized care. Step 6: The effect evaluation and feedback module evaluates the follow-up effect according to the comprehensive scoring formula and feeds the evaluation results back to the model continuous optimization module. Step 7: The model continuous optimization module adjusts and optimizes the AI ​​twin model and follow-up strategy based on the feedback results. Simultaneously, the privacy protection module ensures patient data security using the blockchain access permission verification method, achieving iterative optimization of the follow-up system.

Images