Personalized postpartum rehabilitation medical beauty evaluation system and method based on AI

Abstract

The invention relates to the technical field of artificial intelligence and biological characteristic analysis, in particular to an AI-based personalized postpartum rehabilitation medical beauty evaluation system and method. The method comprises the following steps: generating a standardized action sequence according to a rehabilitation stage and a historical record of a user and guiding execution; collecting dynamic biological characteristic data in real time; aligning data and distributing dynamic weights by using a multi-modal fusion model guided by time sequence semantics, constructing a three-dimensional response graph by combining a tension response function, a collaborative exponential function and a stable fluctuation function, and generating individual embedded vectors through a graph neural network; identifying a potential high-risk time period through map potential energy offset analysis and disturbance sensitivity factors; integrating a user state, a historical response and an intervention knowledge graph, recommending a dynamically adaptive intervention scheme by using a multi-channel attention network, and continuously optimizing through a feedback closed loop; precise evaluation and personalized intervention recommendation of the postpartum rehabilitation process are realized through multi-modal fusion and dynamic modeling, and the rehabilitation efficiency and safety are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and biometric analysis technology, and specifically to an AI-based personalized postpartum rehabilitation and medical aesthetics assessment system and method. Background Art

[0002] Postpartum rehabilitation is an important stage for women to recover their physiological functions and reshape their body after childbirth. The demand for personalized assessment and precise intervention is becoming increasingly prominent. With the rapid development of artificial intelligence technology and biometric analysis methods, multimodal data fusion and dynamic modeling technology have provided new technical paths for real-time monitoring and scientific evaluation of the postpartum rehabilitation process.

[0003] A Chinese invention patent application with publication number CN116386795A discloses a method and system for managing obstetric rehabilitation data, which includes: obtaining the original sub-health data of a target patient, determining the target patient's expected postpartum rehabilitation data based on the original sub-health data, generating a rehabilitation strategy formulated for the target patient based on the expected postpartum rehabilitation data, determining the target patient's collection data indicators at each rehabilitation stage based on the rehabilitation strategy, and obtaining the indicator data in real time for storage and management. By determining the target patient's ideal postpartum rehabilitation data based on their sub-health data and then generating a rehabilitation strategy adapted to them for data collection and management, the server can intelligently formulate a rehabilitation strategy for the patient based on their actual physical condition, thereby improving the adaptability of the rehabilitation plan, while also avoiding the influence of human subjectivity and improving objectivity.

[0004] The current demand for intelligent rehabilitation systems in the medical field has gradually evolved from single functional modules to multi-dimensional dynamic collaborative analysis. In particular, the demand for refined modeling of multiple physiological indicators such as tension, coordination, and stability is growing. At the same time, the intelligent generation and dynamic optimization mechanism of personalized intervention strategies have become the key to improving rehabilitation efficiency and safety. In this context, the comprehensive application of multimodal fusion technology guided by temporal semantics, dynamic graph analysis, and structured knowledge graphs has become the core technology development direction for achieving precise, intelligent, and sustainable optimization in the field of postpartum rehabilitation. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology and propose an AI-based personalized postpartum rehabilitation medical beauty evaluation system and method.

[0006] The technical solution of the present invention is a personalized AI-based postpartum rehabilitation and medical aesthetics assessment method, which includes the following specific implementation steps:

[0007] S1, intelligently generates standard action sequences based on the user's postpartum recovery stage and historical intervention records, and guides standardized execution through voice and animation to activate target area responses;

[0008] S2, synchronously collect short-term dynamic individual response characteristics during the user's action;

[0009] S3. Use a temporal semantics-guided multimodal fusion model to dynamically align and weight the encoded data, construct a three-axis temporal response graph modeling network of tension, coordination, and stability, generate a personalized response graph, and output the final individual response embedding vector through a graph neural network;

[0010] S4. Quantify the individual response spectrum response spectrum deviation trend through the spectrum response potential function, use the disturbance sensitivity factor to analyze the mutation intensity of adjacent time segments, construct the energy mutation domain to identify potential risk segments, and set the spectrum deviation rate as the trigger standard to adaptively judge the abnormal state;

[0011] S5. Integrate individual status, historical responses, and group patterns to generate candidate strategies through a multi-path attention model. Combined with a structured knowledge graph, the intervention combination is optimized by verifying synergy and mutual exclusion relationships, and a feedback closed-loop dynamic tuning strategy is implemented.

[0012] S6. Visualize the recommended intervention pathways and obtain logical explanations for the recommendations.

[0013] Preferably, the three-axis temporal response spectrum modeling network construction process is as follows:

[0014] Construct a cross-modal attention network, guided by physiological intent, and calculate dynamic weights by the similarity between the semantic query vector and the feature key vectors of each modality:

[0015]

[0016] Among them, F fused (t) represents the fusion feature vector at time point t; M represents the total number of modalities; α m (t) represents the dynamic semantic weight of modality m at time t; x m (t) represents the response vector of mode m at time t; Represents the semantic query vector q of the current time window t The transpose of k m,t represents the characteristic key vector of mode m, i.e., its physiological meaning embedding; exp(·) represents the natural exponential function;

[0017] Construct a three-axis time series response graph modeling network to capture the subtle trend changes and coordinated abnormal mutation points of individuals during the recovery process:

[0018] Tension response function T r (t): T r (t)=||x 张力 (t)-x 张力 (t-δ)||2;

[0019] Synergy index function C i (t):

[0020] Stable wave function S f (t):

[0021] Among them, x 张力 (t) represents the characteristic vector of the tension mode at the current time t; δ represents the response time interval; cov(·) represents the covariance function; x 声带 (t) represents the eigenvector of the vocal cord modality at the current time t; x 呼吸 (t) represents the characteristic vector of the respiratory mode at the current time t; δ 声带 and δ 呼吸 represent the standard deviation of vocal cord modal features and respiratory modal features respectively; Var(·) represents the variance within the time window; μ(·) represents the mean within the time window; Δ represents the length of the time window; ε represents a small constant.

[0022] Preferably, the final individual response embedding vector generation process is as follows:

[0023] The three-dimensional data points obtained in all time segments are combined to form a personality response map response map RRG i :

[0024] Among them, RRG i represents the three-dimensional coordinates of the individual response in the i-th time period; n represents the total number of time periods in the recovery monitoring cycle;

[0025] Input the three-dimensional atlas into the graph neural network to generate the final individual response embedding vector E uesr :

[0026] E uesr =GAT(RRG i );

[0027] Here, GAT(·) represents a predefined graph neural network.

[0028] Preferably, the process of generating the energy mutation domain of the graph is as follows:

[0029] Construct a graph response potential function, balance stability and deviation rate by using a weighted coefficient, and combine KL divergence to quantify the difference between the current embedding distribution and the healthy baseline to evaluate the individual response activity and deviation trend;

[0030] The perturbation sensitivity factor is used to quantify the degree of mutation of the embedding vector between adjacent time segments, and to identify the potential high-risk segments of individual states in the time dimension.

[0031] The potential risk time slice set is identified by dual determination of the spectrum response potential energy threshold and the disturbance sensitivity threshold, and the spectrum energy mutation domain is generated: Ω alert ={t i |Φ(t i )>θ Φ Λγ(t i )>θ γ};

[0032] Among them, Ω alert represents the energy mutation domain; t i represents the time index; θ Φ represents the spectrum response potential threshold; θ γ represents the disturbance sensitivity threshold; Φ(t i ) represents the spectrum response potential energy function; γ(t i ) represents the disturbance sensitivity factor.

[0033] Preferably, the process of adaptively determining an abnormal state is as follows:

[0034] Define the spectrum deviation rate as the trigger criterion:

[0035]

[0036] If ψ(t) exceeds the upper limit of the warning region θ ψ , and more than d consecutive frames fall in Ω alert , then the execution exception is triggered.

[0037] Preferably, the intervention combination optimization process is:

[0038] For the currently identified risk segment Ω alert And the spectrum shift performance ψ(t), extract features to construct the state vector S i :S i =Concat(ψ(t),H i ,Θ i );

[0039] Among them, S i represents the state feature vector of the i-th user; H i Represents the user's historical intervention response record vector; Θ i represents the individual background parameter set; Concat(·) represents the concatenation operation;

[0040] Construct a multi-channel attention fusion neural network and construct a state vector S by decomposing features i At most semantic subspaces, combined with multi-head attention to extract key information and estimate channel contribution weights, and output personalized intervention strategy probability distribution after weighted aggregation

[0041] A structured intervention knowledge graph is introduced to verify the collaborative, mutually exclusive and enhanced relationships of the strategy, and an optimized intervention combination plan is generated through subgraph search and weighted reasoning.

[0042] Preferably, the multi-path attention fusion neural network includes:

[0043] Channel decomposition layer: the input state feature vector S i Mapping to K' different semantic subspaces:

[0044] Where, Represents the state feature vector S i The kth semantic subspace of W k represents the linear projection matrix of the kth subspace; b k represents the bias term;

[0045] Multi-head attention encoding layer: for each subspace feature Introducing multi-head attention: After output fusion, the semantic vector Z encoded for each subspace is k ;

[0046] Where Q, K, and V represent query, key, and value matrices, respectively; h represents the attention head number; d k Represents the feature dimension of the kth semantic channel; Attention h (·) indicates attention output;

[0047] Channel importance estimation layer: calculate the semantic contribution α′ of each subspace k :

[0048]

[0049] Among them, α′ k represents the attention fusion weight of the kth semantic channel; u Τ represents the transpose operation of the channel score vector u; W c and b c Represent the weight and bias of the channel fusion scoring layer respectively;

[0050] Result aggregation and recommendation output layer: The weighted fusion features are sent to the final recommendation layer to output the intervention label distribution:

[0051] in, represents the probability distribution of all optional intervention strategies, that is, the predicted probability distribution of all intervention strategy outputs for the i-th user.

[0052] Preferably, the spectrum response potential threshold is adaptively generated by estimating the sliding mean + standard deviation of healthy samples.

[0053] The technical solution of the present invention is an AI-based personalized postpartum rehabilitation and medical aesthetics assessment system, which is used to implement the above-mentioned AI-based personalized postpartum rehabilitation and medical aesthetics assessment method, including:

[0054] The action induction module is used to intelligently generate standard action induction sequences based on the user's postpartum recovery stage and historical intervention records. It guides the user to perform standardized actions through voice and animation, activating the target body area response;

[0055] Multimodal response acquisition module, used to collect dynamic multimodal biometric data in real time;

[0056] The dynamic response modeling module is used to build a multimodal fusion model guided by temporal semantics and align multimodal temporal data. It dynamically assigns weights through a cross-modal attention mechanism, fuses feature vectors, and calculates tension response functions, synergy index functions, and stable fluctuation functions. It constructs individual three-dimensional response maps and generates response embedding vectors through graph neural networks.

[0057] The risk segment determination module is used to identify potential high-risk segments in the current individual recovery cycle through spectral perturbation potential analysis and mutation rate curve deconstruction;

[0058] A flexible intervention feedback engine that integrates user status, historical responses, intervention knowledge graphs, and group patterns, recommending dynamically adapted intervention plans through a multi-channel attention network.

[0059] The interactive explainable module is used to provide visual explanations of intervention strategies, support user interactive feedback, and optimize strategies in conjunction with the elastic intervention feedback engine.

[0060] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0061] This paper designs a personalized AI-based postpartum rehabilitation and aesthetic medical assessment system and method. Through dynamic multimodal biometric acquisition and fusion modeling technology, combined with a temporal semantics-guided multimodal fusion model, it accurately captures the three-dimensional response maps of tension, synergy, and stability during the postpartum rehabilitation process, effectively addressing the shortcomings of traditional assessment methods that rely on single-modal data and ignore dynamic correlations. A two-stage risk assessment framework is constructed using graph potential energy offset analysis and disturbance sensitivity factors, significantly improving the sensitivity and accuracy of identifying potential rehabilitation risk segments. Through the collaborative optimization mechanism of a structured intervention knowledge graph and a multi-pathway attention fusion network, personalized intervention strategies that take into account both synergy effects and taboo constraints are intelligently recommended, reducing the adaptation errors and physiological conflict risks associated with traditional empirical approaches. An intervention effect feedback closed-loop and self-evolutionary learning mechanism are introduced to achieve dynamic tuning and enhanced adaptability of intervention strategies, significantly improving the personalization and safety of the rehabilitation process. A visual interpretation module is combined to enhance user trust and compliance with the intervention logic, constructing a complete intelligent closed-loop system from assessment to intervention, providing an innovative solution for the postpartum rehabilitation field that combines accuracy, safety, and sustainable optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a system architecture diagram of the AI-based personalized postpartum rehabilitation and medical aesthetic assessment system proposed by the present invention;

[0063] Figure 2 This is a flow chart of the method proposed by the present invention, which is based on AI personalized postpartum rehabilitation and medical aesthetic assessment method. DETAILED DESCRIPTION

[0064] Example 1, as Figure 1 As shown, the present invention proposes an AI-based personalized postpartum rehabilitation and medical beauty assessment system, which includes: a motion induction module, a multimodal response acquisition module, a dynamic response modeling module, a risk segment determination module, a flexible intervention feedback engine and an interactive interpretable module.

[0065] The action induction module intelligently generates standard action induction sequences (including but not limited to abdominal pressure and slow inhalation, and vocal cord pronunciation tests) based on the user's postpartum recovery stage and historical intervention records. It guides the user to perform standardized actions through voice and animation, activating the target body area response;

[0066] Multimodal response acquisition module, which collects dynamic multimodal biometric data in real time, including but not limited to skin displacement image sequences, abdominal micro-expressions (micro-muscle movements), pronunciation sound wave spectra, local temperature infrared changes, etc., covering dimensions such as tension, posture, and vocal cord resonance;

[0067] The dynamic response modeling module builds a temporal semantics-guided multimodal fusion model (TSE-MMFM) and aligns multimodal time series data. It dynamically assigns weights through a cross-modal attention mechanism, fuses feature vectors, and calculates the tension response function, synergy index function, and stable fluctuation function to construct an individual three-dimensional response map and generate a response embedding vector through a graph neural network.

[0068] The risk segment determination module identifies the potential high-risk segments in the current individual recovery cycle through spectral perturbation potential analysis and mutation rate curve deconstruction;

[0069] A flexible intervention feedback engine integrates user status, historical responses, intervention knowledge graphs (verifying strategy coordination / conflict relationships), and group patterns to recommend dynamically adapted intervention plans through a multi-channel attention network.

[0070] The interactive explainable module provides visual explanations of intervention strategies (including but not limited to risk area marking and intervention basis reasoning), supports user interactive feedback, and works in conjunction with the elastic intervention feedback engine to optimize strategies.

[0071] Example 2, as Figure 2 As shown, the present invention proposes an AI-based personalized postpartum rehabilitation and medical aesthetics assessment method, which is applied to the AI-based personalized postpartum rehabilitation and medical aesthetics assessment system proposed in Example 1, and specifically includes the following implementation steps:

[0072] S1. Call the action induction module. According to the user's recovery cycle stage (such as 4 weeks / 8 weeks / 12 weeks after delivery) and historical intervention records, it intelligently selects a matching standard action induction sequence (including but not limited to abdominal pressure and slow inhalation, vocal cord pronunciation test, abdominal touch and rebound), and guides the user to perform each action in a standardized manner through natural language voice feedback and screen animation to ensure the activation of the response of the target body area tissue layer.

[0073] S2. Control the multimodal response acquisition module to synchronously collect short-term dynamic individual response characteristics during the user's movements, including but not limited to: skin surface displacement image sequence, abdominal micro-expressions (composed of micro-muscle movements), pronunciation sound wave spectrum and breath continuity, and local temperature sensing infrared dynamic changes.

[0074] S3, the dynamic response modeling module uses a temporal semantics-guided multimodal fusion model (TSE-MMFM) to perform dynamic alignment and semantic weighted encoding based on the characteristics of different modal data (tension, posture, micro-expression, vocal cord resonance, etc.) with different sampling frequencies and response inertia. Specifically:

[0075] S31. Construct a cross-modal attention network, guided by the core physiological intention in the induced task, and assign dynamic weights to the features of each modality at each time point:

[0076]

[0077] Among them, F fused (t) represents the fusion feature vector at time point t; M represents the total number of modalities; α m (t) represents the dynamic semantic weight of modality m at time t, which is determined by the cross-modal attention mechanism; x m (t) represents the response vector of mode m at time t; Represents the semantic query vector q of the current time window t The transpose of k m,t represents the characteristic key vector of mode m, i.e., its physiological meaning embedding; exp(·) represents the natural exponential function;

[0078] S32. Construct a three-axis temporal response graph modeling network (TAR-Net) to capture the subtle trend changes and coordinated abnormal mutation points of individuals during the recovery process. Specifically:

[0079] Tension response function T r (t): T r (t)=||x 张力 (t)-x 张力 (t-δ)||2;

[0080] Synergy index function C i (t):

[0081] Stable wave function S f (t):

[0082] Among them, x 张力 (t) represents the characteristic vector of the tension mode at the current time t; δ represents the response time interval, which is the time difference before and after comparison; cov(·) represents the covariance function, which measures the synchronous fluctuation between the two modes; x 声带 (t) represents the eigenvector of the vocal cord modality at the current time t; x 呼吸 (t) represents the characteristic vector of the respiratory mode at the current time t; δ 声带 and δ 呼吸 denote the standard deviation of the vocal cord modal features and the respiratory modal features respectively; Var(·) denotes the variance within the time window; μ(·) denotes the mean within the time window; Δ denotes the length of the time window; ε denotes a small constant, which is set to 1×10 in this embodiment. -8 ;

[0083] S33. Based on the three-axis function, the three-dimensional data points obtained in all time segments are set to form a set of structured personality response maps RRG i :

[0084] Among them, RRG i represents the three-dimensional coordinates of the individual response in the i-th time period, i.e., the three-dimensional atlas dot matrix; n represents the total number of time periods in the recovery monitoring cycle;

[0085] Input the three-dimensional atlas into the graph neural network (GAT) to generate the final individual response embedding vector E uesr :

[0086] E uesr =GAT(RRG i );

[0087] Where GAT(·) represents a predefined graph neural network;

[0088] S34, output the current moment's response three-axis index (T r (t),C i (t),S f (t)), response spectrum response spectrum RRG i and the individual response embedding vector E uesr .

[0089] S4, the risk segment determination module builds a determination framework based on the dual-stage time trajectory disturbance sensitivity + spectrum potential energy offset analysis. It introduces the response potential energy function to quantify the spectrum embedding change trend, uses the disturbance propagation sensitivity factor to compare adjacent segments, constructs the spectrum energy mutation domain to identify nonlinear risk areas, and uses the spectrum offset rate for adaptive trigger judgment. Specifically:

[0090] S41. Construct the spectrum response potential energy function Φ(t) to quantify the activity and deviation of the current individual response spectrum at a specific moment: Φ(t)=||E user (t)||2-λ·KL(P t ||P ref );

[0091] Where λ represents the trade-off coefficient (empirically 0.3 to 0.5), which controls the balance between stability and drift rate; KL(·) represents the Kullback-Leibler divergence, i.e., the KL divergence, which measures the difference between the current distribution and the healthy baseline; P t Represents the embedding feature distribution (local), the normalized probability distribution form of the current embedding vector; P ref Represents the average atlas embedding distribution of healthy samples, which is obtained from the preset or training samples;

[0092] S42. Calculate the disturbance propagation sensitivity of adjacent time segments, characterize the mutation intensity of individual states in the time dimension, and identify potential risk segments:

[0093]

[0094] Among them, γ(t) represents the perturbation sensitivity factor, which measures the degree of embedding mutation between the current and previous windows;

[0095] S43. Define the energy mutation domain Ω alert :Ω alert ={t i |Φ(t i )>θ Φ Λγ(t i )>θ γ};

[0096] Among them, Ω alert represents the energy mutation domain, that is, the set of time slices that are judged to have potential risks; t i Represents the time index, that is, each frame or each time slice; θ Φ represents the potential energy threshold of the spectrum response, which is adaptively generated by the sliding mean + standard deviation estimation of healthy samples; θ γ represents the disturbance sensitivity threshold;

[0097] S44. Define the Embedding Shift Ratio as the trigger criterion:

[0098]

[0099] If ψ(t) exceeds the upper limit of the warning region θ ψ , and more than d consecutive frames fall in Ω alert , then the execution exception is triggered;

[0100] Where ψ(t) represents the spectrum deviation rate, reflecting the degree to which the individual deviates from the healthy spectrum within the last L frames; L represents the length of the backtracking frame; E ref (j) represents the reference atlas embedding vector under the individual’s historical health status. In this embodiment, it is set to the individual’s average atlas over the past 7 days.

[0101] S5. The flexible intervention feedback engine automatically recommends and dynamically adapts personalized intervention strategies based on the individual's current state evolution trajectory, past intervention responses, intervention strategy knowledge graph, and similar group response patterns. It introduces a multi-dimensional attention fusion model and a closed-loop mechanism for intervention effect feedback to achieve continuous iterative optimization of the state-solution-response process. Specifically:

[0102] S51, for the currently identified risk segment Ω alert And the spectrum shift performance ψ(t) extracts features to construct the state vector S i :S i=Concat(ψ(t),H i ,Θ i );

[0103] Among them, S i represents the state feature vector of the i-th user, which serves as the input of the AI ​​recommendation model and comprehensively describes the current rehabilitation status and historical response; H i Represents the user's historical intervention response record vector, including but not limited to the recovery amplitude, allergic event records, and effect duration brought about by each intervention method; Θ i represents an individual background parameter set (including but not limited to age, cesarean section / vaginal delivery, breast condition, and skin condition recovery period), reflecting physiological differences; Concat(·) represents a concatenation operation, which is used to construct the above three parts into a high-dimensional feature input;

[0104] S52. Construct a Multi-Attentive Fusion Network (MAFN) based on the Transformer architecture extension. Through the multi-channel attention focusing mechanism, it mines the feature subspaces with the strongest correlation with the intervention plan in the data of different dimensions and obtains the predicted probability distribution of the intervention strategy label. Output candidate intervention strategy set

[0105] S53, due to the output candidate intervention strategy set In the process of intervention, there are often synergistic effects or conflicting taboos between different means. If directly implemented, it may lead to over-intervention or physiological adaptation mismatch. Therefore, a structured intervention knowledge graph G is introduced. interv , used to perform semantic constraints on candidate tag sets, verify combination legitimacy, and enhance path reasoning. The specific operation process is as follows:

[0106] Perform entity node matching on the top N candidate strategy labels output by the model;

[0107] Find co-apply, boosted-by, and exclusive edges in the graph;

[0108] Generate the optimal intervention combination plan P through subgraph search and weight reasoning i :P i ={(a j ,ω j )};

[0109] Among them, a j is the selected intervention; j is the strategy weight; G intervRepresents the constructed intervention knowledge graph, where nodes are intervention entities and edges are semantic relationships (e.g., enhancement, conflict, synergy); Co-apply indicates that two strategies can be implemented together (e.g., thermal therapy and psychological relaxation guidance) without interference; Exclusive indicates that two strategies have physiological / psychological conflicts and cannot be used in parallel (e.g., skin tightening micro-electric stimulation and deep fat thermolysis); Boosted-by indicates that there is a mutual enhancement effect between the strategies (e.g., medium-frequency electrotherapy enhanced by nutritional infusion for recovery);

[0110] It should be noted that the structured intervention knowledge graph G interv The structure includes:

[0111] Node type: intervention method, applicable population, physiological site, and contraindication;

[0112] Edge types: co-apply, boosted-by, exclude-with, etc.

[0113] Attribute information: implementation intensity, compliance rate, response delay, etc. of each intervention measure;

[0114] S54, intervention combination plan P i After being pushed to the client, the individual response indicators (including but not limited to temperature response lag, muscle tension recovery changes, local metabolic parameters) during the execution process are continuously monitored to form a feedback vector F i , used to adjust the subsequent state vector:

[0115] in, Represents the updated state vector, which serves as the new input for the next round of recommendations, enabling the system’s self-evolutionary learning; represents the state feature vector of the i-th user before the t-th intervention; ΔF i The difference vector representing the state response after the intervention is constructed by the difference in the monitored intervention effects (including but not limited to a slower rate of skin temperature drop and a shorter muscle recovery time);

[0116] According to ΔF i direction and magnitude, update the strategy recommendation logic in real time, complete the fine dynamic self-optimization of individual intervention effects, and continuously perform multiple rounds of closed-loop processing of state update - strategy re-reasoning - combination matching to improve the level of intervention personalization and response sensitivity.

[0117] S6. The interactive explainable module visualizes the recommended intervention path and obtains the recommendation logic explanation.

[0118] In the third embodiment, the present invention proposes an AI-based personalized postpartum rehabilitation and medical aesthetic assessment method, which also includes a multi-channel attention fusion neural network (Multi-Attentive Fusion Network, MAFN), whose specific core modules are:

[0119] Channel decomposition layer (Multi-view Projection): The input state feature vector S i Mapping to K' different semantic subspaces:

[0120] Where, Represents the state feature vector S i The kth semantic subspace of W k represents the linear projection matrix of the kth subspace; b k represents the bias term;

[0121] Multi-Head Attention encoding layer: for each subspace feature Introducing a multi-head attention mechanism to model key information: The output results are fused to form the semantic vector Z after encoding each subspace. k ;

[0122] Where Q, K, and V represent query, key, and value matrices, respectively; h represents the attention head number; d k Represents the feature dimension (subspace dimension) of the kth semantic channel; Attention h (·) indicates attention output;

[0123] Channel-wiseAttention: Calculate the semantic contribution α′ of each subspace k (i.e., the influence weights of different feature domains in the current state):

[0124]

[0125] Among them, α′ k represents the attention fusion weight of the kth semantic channel, that is, the influence of this dimension on the current recommendation decision; u Τ represents the transpose operation of the channel score vector u, which is used to capture the importance of the subspace; W c and b c Represent the weight and bias of the channel fusion scoring layer respectively;

[0126] Result aggregation and recommendation output layer: The weighted fusion features are sent to the final recommendation layer to output the intervention label distribution:

[0127] in, Represents the probability distribution of all optional intervention strategies, and the predicted probability distribution of all intervention strategies output for the i-th user (including but not limited to exercise intervention, physical therapy intervention, and nutritional plan).

[0128] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A personalized postpartum rehabilitation and medical aesthetics assessment method based on AI, characterized by: The specific implementation steps include the following: S1, intelligently generates standard action sequences based on the user's postpartum recovery stage and historical intervention records, and guides standardized execution through voice and animation to activate target area responses; S2, synchronously collect short-term dynamic individual response characteristics during the user's action; S3. Use a temporal semantics-guided multimodal fusion model to dynamically align and weight the encoded data, construct a three-axis temporal response graph modeling network of tension, coordination, and stability, generate a personalized response graph, and output the final individual response embedding vector through a graph neural network; S4. Quantify the individual response spectrum response spectrum deviation trend through the spectrum response potential function, use the disturbance sensitivity factor to analyze the mutation intensity of adjacent time segments, construct the energy mutation domain to identify potential risk segments, and set the spectrum deviation rate as the trigger standard to adaptively judge the abnormal state; S5. Integrate individual status, historical responses, and group patterns to generate candidate strategies through a multi-path attention model. Combined with a structured knowledge graph, the intervention combination is optimized by verifying synergy and mutual exclusion relationships, and a feedback closed-loop dynamic tuning strategy is implemented. S6. Visualize the recommended intervention pathways and obtain logical explanations for the recommendations.

2. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 1, characterized in that: The process of constructing the three-axis temporal response spectrum modeling network is as follows: Construct a cross-modal attention network, guided by physiological intent, and calculate dynamic weights by the similarity between the semantic query vector and the feature key vectors of each modality: Among them, F fused (t) represents the fusion feature vector at time point t; M represents the total number of modalities; α m (t) represents the dynamic semantic weight of modality m at time t; x m (t) represents the response vector of mode m at time t; Represents the semantic query vector q of the current time window t The transpose of k m,t represents the characteristic key vector of mode m, i.e., its physiological meaning embedding; exp(·) represents the natural exponential function; Construct a three-axis time series response graph modeling network to capture the subtle trend changes and coordinated abnormal mutation points of individuals during the recovery process: Tension response function T r (t): T r (t)=||x 张力 (t)-x 张力 (t-δ)||2; Synergy index function C i (t): Stable wave function S f (t): Among them, x 张力 (t) represents the characteristic vector of the tension mode at the current time t; δ represents the response time interval; cov(·) represents the covariance function; x 声带 (t) represents the eigenvector of the vocal cord modality at the current time t; x 呼吸 (t) represents the characteristic vector of the respiratory mode at the current time t; δ 声带 and δ 呼吸 represent the standard deviation of vocal cord modal features and respiratory modal features respectively; Var(·) represents the variance within the time window; μ(·) represents the mean within the time window; Δ represents the length of the time window; ε represents a small constant.

3. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 2, characterized in that: The final individual response embedding vector generation process is as follows: The three-dimensional data points obtained in all time segments are combined to form a personality response map response map RRG i : Among them, RRG i represents the three-dimensional coordinates of the individual response in the i-th time period; n represents the total number of time periods in the recovery monitoring cycle; Input the three-dimensional atlas into the graph neural network to generate the final individual response embedding vector E uesr : IN uesr =GAT(RRG i ); Here, GAT(·) represents a predefined graph neural network.

4. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 3, characterized in that: The process of generating the spectrum energy mutation domain is as follows: Construct a graph response potential function, balance stability and deviation rate by using a weighted coefficient, and combine KL divergence to quantify the difference between the current embedding distribution and the healthy baseline to evaluate the individual response activity and deviation trend; The perturbation sensitivity factor is used to quantify the degree of mutation of the embedding vector between adjacent time segments, and to identify the potential high-risk segments of individual states in the time dimension. The potential risk time slice set is identified by dual determination of the spectrum response potential energy threshold and the disturbance sensitivity threshold, and the spectrum energy mutation domain is generated: Ω alert ={t i |Φ(t i )>θ Φ Λγ(t i )>θ γ }; Among them, Ω alert represents the energy mutation domain; t i Represents a time index; θ Φ represents the spectrum response potential threshold; θ γ represents the disturbance sensitivity threshold; Φ(t i ) represents the spectrum response potential energy function; γ(t i ) represents the disturbance sensitivity factor.

5. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 4, characterized in that: The process of adaptively judging abnormal states is as follows: Define the spectrum deviation rate as the trigger criterion: If ψ(t) exceeds the upper limit of the warning region θ ψ , and more than d consecutive frames fall in Ω alert , then the execution exception is triggered.

6. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 5, characterized in that: The intervention combination optimization process is: For the currently identified risk segment Ω alert And the spectrum shift performance ψ(t), extract features to construct the state vector S i :S i =Concat(ψ(t),H i ,Θ i ); Among them, S i represents the state feature vector of the i-th user; H i Represents the user's historical intervention response record vector; Θ i represents the individual background parameter set; Concat(·) represents the concatenation operation; Construct a multi-channel attention fusion neural network and construct a state vector S by decomposing features i At most semantic subspaces, combined with multi-head attention to extract key information and estimate channel contribution weights, and output personalized intervention strategy probability distribution after weighted aggregation A structured intervention knowledge graph is introduced to verify the collaborative, mutually exclusive and enhanced relationships of the strategy, and an optimized intervention combination plan is generated through subgraph search and weighted reasoning.

7. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 6, characterized in that: The multi-channel attention fusion neural network includes: Channel decomposition layer: the input state feature vector S i Mapping to K' different semantic subspaces: Where, Represents the state feature vector S i The kth semantic subspace of W k represents the linear projection matrix of the kth subspace; b k represents the bias term; Multi-head attention encoding layer: for each subspace feature Introducing multi-head attention: After output fusion, the semantic vector Z encoded for each subspace is k ; Where Q, K, and V represent query, key, and value matrices, respectively; h represents the attention head number; d k Represents the feature dimension of the kth semantic channel; Attention h (·) indicates attention output; Channel importance estimation layer: calculate the semantic contribution α′ of each subspace k : Among them, α′ k represents the attention fusion weight of the kth semantic channel; u Τ represents the transpose operation of the channel score vector u; W c and b c Represent the weight and bias of the channel fusion scoring layer respectively; Result aggregation and recommendation output layer: The weighted fusion features are sent to the final recommendation layer to output the intervention label distribution: in, represents the probability distribution of all optional intervention strategies, that is, the predicted probability distribution of all intervention strategy outputs for the i-th user.

8. The AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to claim 5, characterized in that: The spectral response potential threshold is adaptively generated by estimating the sliding mean + standard deviation of healthy samples.

9. An AI-based personalized postpartum rehabilitation and medical aesthetics assessment system, which is used to execute the AI-based personalized postpartum rehabilitation and medical aesthetics assessment method according to any one of claims 1 to 8, characterized in that: include: The action induction module is used to intelligently generate standard action induction sequences based on the user's postpartum recovery stage and historical intervention records. It guides the user to perform standardized actions through voice and animation, activating the target body area response; Multimodal response acquisition module, used to collect dynamic multimodal biometric data in real time; Dynamic response modeling module, used to build a multimodal fusion model guided by temporal semantics and align multimodal temporal data, dynamically assigning weights through a cross-modal attention mechanism, The eigenvectors are fused and the tension response function, synergy index function, and stable fluctuation function are calculated to construct an individual three-dimensional response map. The response embedding vector is generated through a graph neural network. The risk segment determination module is used to identify potential high-risk segments in the current individual recovery cycle through spectral perturbation potential analysis and mutation rate curve deconstruction; A flexible intervention feedback engine that integrates user status, historical responses, intervention knowledge graphs, and group patterns, recommending dynamically adapted intervention plans through a multi-channel attention network. The interactive explainable module is used to provide visual explanations of intervention strategies, support user interactive feedback, and optimize strategies in conjunction with the elastic intervention feedback engine.

Images