Individualized rehabilitation intervention prescription generation method fusing hierarchical positioning rehabilitation assessment and multi-stage reinforcement search

Abstract

The application provides an individualized rehabilitation intervention prescription generation method fusing hierarchical positioning rehabilitation evaluation and multi-stage enhanced retrieval, and belongs to the field of rehabilitation medical engineering and intelligent health technology. The method collects motion function data of a subject during the execution of a preset evaluation action, constructs macro kinematics characteristics and micro electromyography characteristics, performs hierarchical evaluation of the motion function state of the subject from the whole to the local, identifies abnormal function partitions and abnormal muscle groups, embeds and represents basic information of the patient, macro motion evaluation information, micro electromyography evaluation information and past training prescription information, and modulates and fuses to form a unified evaluation embedding, completes evaluation-prescription rule retrieval, training item matching and safety criterion verification in sequence based on an enhanced retrieval mechanism, and generates an individualized rehabilitation training prescription. The method can realize direct mapping of the evaluation result and the training prescription, and improve the pertinence, interpretability and safety of the rehabilitation training.

Description

Technical Field

[0001] This invention belongs to the field of rehabilitation medicine engineering and intelligent health monitoring technology, and in particular relates to a method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval. This method achieves an objective assessment of the degree of motor function impairment by performing hierarchical assessment and localization of the impairment site, and is used to guide the formulation of individualized rehabilitation training prescriptions. It is applicable to rehabilitation assessment and training intervention scenarios for patients with neurogenic motor function impairment. Background Technology

[0002] Motor dysfunction is a common type of functional deficit in patients with stroke, brain injury, and other neurological diseases. Rehabilitation typically relies on assessments of the patient's motor function to develop appropriate rehabilitation training programs. Current clinical rehabilitation assessments primarily use functional scales, such as the Fugl-Meyer Assessment Scale, Brunnstrom Staging, and Barthel Index. These assessment methods can grade and assess a patient's motor abilities from a holistic functional perspective and have the advantages of being easy to operate and widely applicable in clinical practice.

[0003] However, the above assessment methods mainly rely on the therapist's subjective observation and experience judgment. The scoring granularity is relatively coarse, making it difficult to conduct a detailed analysis of the functional status of specific limbs, joints or muscle groups. Moreover, the assessment results are mostly presented in the form of grades or total scores, which are difficult to directly reflect the specific abnormal characteristics of patients in terms of motor control, muscle activation and compensatory patterns, thus limiting their ability to guide the development of individualized rehabilitation training prescriptions.

[0004] With the development of wearable sensing technology, motor function assessment methods based on inertial measurement units and surface electromyography (EMG) signals have gradually attracted attention. Related studies have attempted to objectively quantify patients' motor abilities using kinematic parameters or EMG characteristics, improving the objectivity and repeatability of assessments to some extent. However, existing wearable sensing-based assessment methods mostly focus on overall motor performance or single signal feature analysis, lacking a hierarchical assessment mechanism from the whole to the part. This makes it difficult to achieve step-by-step localization of functional impairments and to distinguish whether functional limitations stem from insufficient muscle activation, abnormal synergistic control, or compensatory motor enhancement.

[0005] Furthermore, during the implementation of rehabilitation training, there is often a lack of clear and quantifiable mapping between assessment results and training prescriptions. The content and intensity of rehabilitation training still mainly rely on the therapist's experience for manual formulation, making it difficult to directly translate assessment results into targeted training programs, thus affecting the accuracy and efficiency of rehabilitation interventions. Therefore, current technology lacks a solution that can achieve hierarchical positioning during the motor function assessment stage and automatically generate individualized rehabilitation training prescriptions based on the positioning results. Summary of the Invention

[0006] The purpose of this invention is to address the problems of existing rehabilitation assessment methods, which rely primarily on overall scoring and lack the ability to precisely locate abnormalities in motor function. Furthermore, the correlation between assessment results and rehabilitation training prescriptions depends on manual experience mapping, making it difficult to achieve individualized and safe prescription generation. This invention proposes a method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment with multi-stage enhanced retrieval. By conducting hierarchical assessments of the subject's motor function status from macro to micro levels, abnormal functional zones and abnormal muscle groups at different functional levels are identified, constructing a structured assessment result that reflects the patient's current comprehensive motor function status. Based on this, a multi-source assessment information fusion and embedding representation mechanism is introduced to uniformly represent the patient's basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information. Further, by combining an assessment-prescription rule base, a training program knowledge base, and a safety criterion base, an enhanced retrieval method is used to generate candidate training prescriptions that match the assessment results. Finally, under safety constraints, an executable individualized rehabilitation training prescription is obtained through screening.

[0007] To achieve the objectives of this invention, a method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval is disclosed, comprising the following steps:

[0008] S1. Data Acquisition: This includes inertial motion data acquisition and electromyography (EMG) data acquisition. Macroscopic motion data and microscopic EMG data are collected from the subjects during the execution of preset assessment actions. The collected data are preprocessed to construct motor function characteristic data for subsequent analysis.

[0009] S2. Layered Rehabilitation Assessment: Based on motor function characteristic data, the subject's motor function status is assessed in a layered manner, completing the overall motor function assessment, macroscopic abnormal function zoning and abnormal muscle groups and synergistic control abnormality identification tasks, and outputting structured layered motor function assessment results.

[0010] S3. Multi-source assessment information fusion and embedding: Encode, embed, and modulate the patient's basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information to construct a unified multi-source assessment embedding representation to characterize the patient's current comprehensive motor function status.

[0011] S4. Training prescription generation: Based on multi-source assessment embedding representation, combined with assessment-prescription rule base, training item knowledge base and safety criterion base, the enhanced retrieval mechanism sequentially completes the tasks of rule matching, training item matching and safety verification to generate an individualized rehabilitation training prescription that matches the patient's current motor function status.

[0012] Through the above four steps, this invention uses the results of hierarchical rehabilitation assessment as the basis for training decisions, directly mapping the results of motor function assessment to the rehabilitation training prescription generation process, so that the training prescription can be differentiated for abnormalities in different functional levels and functional areas, thereby improving the pertinence and effectiveness of rehabilitation training.

[0013] Furthermore, in step 1, the configuration and data acquisition of the macroscopic motion data sensor specifically includes the following steps:

[0014] Inertial measurement units (IMUs) were used to collect macroscopic motion data during the subject's movement. One IMU sensor was used for each macroscopic anatomical location. The IMU included a triaxial accelerometer, a triaxial angular velocity meter, and a triaxial magnetometer, which were used to collect acceleration, angular velocity, and spatial orientation information of the limbs during movement.

[0015] IMU sensors are worn on key limb segments of the subject, including at least: upper limbs: upper arm, forearm (optionally worn on the back of the hand); lower limbs: thigh, calf (optionally worn on the dorsum of the foot); the inertial measurement unit (IMU) continuously acquires motion data at a fixed sampling frequency fs, obtaining 12 channels of macroscopic motion data signals, including:

[0016] Triaxial acceleration signal a(t):

[0017]

[0018] Among them, a x (t) represents the acceleration in the x-axis direction, a y (t) represents the acceleration in the y-axis direction, a z (t) represents the acceleration in the z-axis direction;

[0019] Triaxial angular velocity signal w(t):

[0020]

[0021] Among them, w x (t) represents the angular velocity along the x-axis, w y (t) represents the angular velocity along the y-axis, w z (t) represents the angular velocity along the z-axis;

[0022] Triaxial magnetic field intensity signal m(t):

[0023]

[0024] Where, m x (t) represents the magnetic field strength along the x-axis, my (t) represents the magnetic field strength along the y-axis, m z (t) represents the magnetic field strength along the z-axis;

[0025] The collected raw motion data is processed by time synchronization, noise reduction filtering and coordinate system calibration, and the limb posture information is calculated based on sensor fusion algorithms (such as complementary filtering or Kalman filtering) to obtain the posture sequence represented in quaternion or Euler angle form.

[0026]

[0027] Where, θ x (t) represents the rotation angle about the x-axis, θ y (t) represents the rotation angle about the y-axis, θ z (t) represents the rotation angle about the z-axis.

[0028] Furthermore, in step 2, the rehabilitation assessment of macroscopic movement data specifically includes the following steps:

[0029] Macroscopic motion feature construction and evaluation; based on attitude calculation results and angular velocity information, a set of kinematic features for macroscopic evaluation is constructed; for a 12-channel signal acquired by an inertial measurement sensor, the macroscopic motion features include:

[0030] Activity range ROM∈R 12 :

[0031]

[0032] Where max represents the maximum value, min represents the minimum value, and ROM i ROM representing the signal of the i-th channel; average activity ∈R 12 :

[0033]

[0034] in, Represents the signal of the i-th channel ;

[0035] Multi-joint synergy feature matrix C∈R 12×12 :

[0036]

[0037] Where corr represents the correlation coefficient, C ij The multi-joint coordination between the i-th and j-th signals is represented; x(t) represents the signal of a single channel acquired by the IMU, which is any one of the 12 channels of macroscopic motion data signals; xi (t), x j (t) represents two different signals from the 12-channel macroscopic motion data signal;

[0038] Based on macroscopic movement characteristics, a macroscopic assessment of the overall motor function status of the subjects is performed; a macroscopic motor function score S is calculated. macro ∈R:

[0039]

[0040] Where: f k Represents the k-th macroscopic motion characteristic; ω k The corresponding feature weights reflect the contribution of each feature to overall motor function; N represents the number of features; the macroscopic motor function score reflects the overall motor ability level of the subject and serves as the basic input for subsequent stratified assessment; the above features together constitute the macroscopic motor feature set F. macro :

[0041]

[0042] Macroscopic anomaly localization: Based on macroscopic motor function assessment, further analysis is conducted on the distribution differences of macroscopic motor characteristics in different limb segments or functional zones to achieve macroscopic anomaly localization.

[0043] For each macroscopic anatomical region (corresponding to one IMU sensor), calculate the corresponding local macroscopic feature vector:

[0044]

[0045] And calculate its relationship with the set of health reference features. Distance between:

[0046]

[0047] The function dis(a1,a2) represents a vector formed by calculating the Euclidean distance between the elements in sets a1 and a2 in sequence; when the distance D corresponding to a certain functional partition... r When the threshold is exceeded, the functional area is determined to be a macroscopic motor abnormality area. This method enables macroscopic localization of the subject's motor function abnormality areas, guiding the subsequent acquisition of microscopic electromyographic data and the microscopic functional assessment process.

[0048] Furthermore, in step 1, the configuration and data acquisition of the microscopic electromyography data sensor specifically includes the following steps:

[0049] After completing the macroscopic motor function assessment and abnormality localization, microscopic electromyography (EMG) data is collected for the identified abnormal limb segments or functional zones to analyze the local neuromuscular functional status. The microscopic EMG data is collected using surface electromyography (sEMG) sensors.

[0050] The surface electromyography (EMG) sensor is a bipolar or multipolar surface electrode, which is attached to the surface of key muscle groups within the target functional area and arranged along the muscle fiber direction. In specific implementation, the EMG signal acquisition parameters are set as follows:

[0051] Sampling frequency: femg = 1000~2000Hz

[0052] Analog bandpass filter range: 20–450Hz;

[0053] Power frequency notch filtering: 50Hz or 60Hz;

[0054] The raw electromyographic signals acquired at the abnormal zonal locations are represented as follows:

[0055]

[0056] Where x i (t) represents the electromyographic signal of the i-th target muscle at time t, and M is the number of muscles collected; the original electromyographic signal is denoised, rectified and smoothed to obtain the electromyographic signal sequence used for feature extraction.

[0057] Furthermore, in step 2, the rehabilitation assessment of microscopic electromyographic data specifically includes the following steps:

[0058] Microscopic electromyographic feature extraction and microscopic neuromuscular function assessment: Based on preprocessed electromyographic signals, microscopic electromyographic features that can reflect muscle activation level, fatigue state and synergistic control characteristics are extracted.

[0059] Microscopic electromyographic characteristics include:

[0060] Root mean square value RMS∈R M RMS i RMS representing the i-th lead:

[0061]

[0062] Mean Absolute Value (MAV) ∈ R M MAV i The average absolute value of the i-th lead:

[0063]

[0064] Muscle activation normalization index A∈RM A i The muscle activation normalization index representing the i-th lead normalizes the electromyographic amplitude relative to the maximum voluntary contraction MVC:

[0065]

[0066] Based on microscopic electromyographic feature vectors, a microscopic assessment of neuromuscular functional status within the target functional area is performed; a microscopic muscle function score S is constructed using a weighted method. micro ∈R:

[0067]

[0068] in, These are the weighting coefficients for the corresponding features; the weights are used to reflect the degree of contribution of different muscle groups and features to local motor function; the microscopic assessment results are used to quantify whether the muscle activation level is insufficient, or whether there is abnormal enhancement or fatigue trend.

[0069] Abnormal Muscle Groups and Synergistic Patterns: To further identify specific muscle factors leading to macroscopic motor abnormalities, this invention performs abnormal muscle group localization analysis based on microscopic assessment. By comparing the relative relationships between activation levels of different muscle groups, a muscle group synergistic matrix C is constructed. emg ∈R M×M Its elements Represented as:

[0070]

[0071] When the activation level of a certain muscle group is significantly lower or higher than that of other muscle groups in the same functional area, or when the correlation between it and the main synergistic muscle group is lower than a preset threshold, the muscle group is determined to be an abnormal muscle group.

[0072] At the same time, by combining the abnormal functional zoning information obtained in the macro-assessment stage, the micro-assessment results are constrained, thereby improving the accuracy and stability of abnormal muscle group localization.

[0073]

[0074] Through the process of acquiring and evaluating microscopic electromyographic data, a set of microscopic neuromuscular function evaluation results within the target functional area is output. The results include at least: a set of abnormal muscle groups; activation level indicators of each muscle group; and abnormal characteristics of synergistic control, denoted as F. micro .

[0075] Furthermore, step 3, the fusion and embedding of multi-source evaluation information specifically includes the following steps:

[0076] Step 3-1: Word embedding of patient basic information, i.e., first information embedding;

[0077] The first piece of information is the patient's basic information, which includes the patient's age, gender, affected side, disease stage, and other basic attributes related to rehabilitation training. This basic information is encoded and embedded into a vector inf1 of length L.

[0078]

[0079] Further construct a trainable transformation matrix S1:

[0080]

[0081] And the first information matrix Inf1 is constructed using a weighted method:

[0082]

[0083] Step 3-2: Word embedding of macroscopic motor function information, i.e., second information embedding;

[0084] The second information is macroscopic motor function assessment information, which originates from the macroscopic motor data acquisition and assessment positioning steps, including macroscopic motor characteristics of different limb segments or functional zones; the macroscopic motor function assessment information is embedded into a second information matrix Inf2 of size N×L1:

[0085]

[0086] Where N represents the number of macroscopic functional zones or macroscopic anatomical zones; each row represents the macroscopic motion feature embedding representation corresponding to a functional zone, that is, the set of macroscopic motion features. Each element in the array is expanded, concatenated, and then embedded, with a length of L1.

[0087] Step 3-3: Word embedding of microscopic electromyographic functional information, i.e., third information embedding;

[0088] The third piece of information is microscopic electromyography (EMG) function assessment information, which comes from the microscopic EMG data acquisition and assessment steps, including the activation level, synergistic relationship, and fatigue characteristics of the target muscle group; the microscopic EMG function assessment information is embedded into a third information matrix Inf3 of size M×L1:

[0089]

[0090] Where M represents the number of target muscle groups or microscopic functional units; L1 represents the dimension of the microscopic electromyographic feature embedding, i.e., the set of microscopic neuromuscular function assessment results F micro Each element is expanded and concatenated to form a vector of length L1;

[0091] Steps 3-4: Word embedding of previously trained prescription information, i.e., fourth information embedding;

[0092] The fourth information is the previous training prescription information, which includes the types of training exercises, training intensity, training frequency, and training effect feedback. This previous training prescription information is embedded into a fourth information matrix Inf4 of size K×L1.

[0093]

[0094] Where K represents the number of items in the prescription, and L1 is the length of the execution result embedding for each item;

[0095] Steps 3-5: Multi-source assessment information fusion; modulated fusion of patient basic information and macroscopic motion information to obtain Inf5, and modulated fusion of patient basic information and microscopic electromyographic information to obtain Inf6, and outputting the fusion embedded results.

[0096] Furthermore, steps 3-5 specifically include the following steps:

[0097] Step 3-5-1: Modulation and fusion of patient basic information and macroscopic motion information to obtain Inf5;

[0098] The patient's basic information embedding vector is:

[0099]

[0100] The macroscopic motion information matrix is ​​as follows:

[0101]

[0102] Step 3-5-1-1: Trainable mapping of basic information and generation of gated vectors;

[0103] Introducing a trainable weight matrix W g1 ∈R L1×L1 With bias b g1 ∈R L1 The patient's basic information is mapped to a gating vector g1 (modulation vector):

[0104]

[0105] Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1];

[0106] Step 3-5-1-2: Trainable feature transformation of macroscopic information;

[0107] Introducing the macroscopic feature transformation matrix W m1 ∈R L1×L1 With bias b m1 ∈R L1 By transforming the macroscopic information line by line, we obtain :

[0108]

[0109] Where 1∈R N×1 This is a vector of all 1s, used for broadcasting the bias term;

[0110] Step 3-5-1-3: Gated modulation and residual fusion;

[0111] The gate vector g1 is broadcast row-wise to N×L1 to achieve weighted modulation of macroscopic information by patient information. A residual term is introduced to improve stability, resulting in the fifth information matrix Inf5, which is a modulation and fusion of basic patient information and macroscopic motion information.

[0112]

[0113] in For element-wise multiplication, λ is a configurable or trainable scalar coefficient used to control the residual fusion ratio;

[0114] Step 3-5-2: Modulation and fusion of patient basic information and microscopic electromyographic information to obtain Inf6;

[0115] The patient's basic information embedding vector is:

[0116]

[0117] The macroscopic motion information matrix is ​​as follows:

[0118]

[0119] Step 3-5-2-1: Trainable mapping of basic information and generation of gated vectors;

[0120] Introducing a trainable weight matrix W g2 ∈R L1×L1 With bias b g2 ∈R L1 Patient basic information is mapped to a gating vector (modulation vector):

[0121]

[0122] Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1];

[0123] Step 3-5-2-2: Trainable feature transformation of macroscopic information;

[0124] Introducing the micro-feature transformation matrix W m2 ∈R L1×L1 With bias b m2 ∈R L1The microscopic information is obtained by transforming it line by line. :

[0125]

[0126] Where 1∈R N×1 This is a vector of all 1s, used for broadcasting the bias term;

[0127] Step 3-5-2-3: Gated modulation and residual fusion;

[0128] The gating vector g2 is broadcast row-wise to N×L1 to achieve weighted modulation of microscopic information based on patient information. A residual term is introduced to improve stability. The sixth information matrix Inf6, which fuses the modulation of basic patient information and microscopic electromyographic information, is as follows:

[0129]

[0130] in For element-wise multiplication, λ is a configurable or trainable scalar coefficient used to control the residual fusion ratio;

[0131] Step 3-5-3: Output of Fusion Embedding Results; After completing the modulation and fusion of patient basic information and macroscopic motor function information, and the modulation and fusion of patient basic information and microscopic electromyographic function information, the fifth information matrix and the sixth information matrix are obtained, respectively:

[0132]

[0133]

[0134] Meanwhile, the previously trained prescription information is used to obtain a fourth information matrix after word embedding:

[0135]

[0136] To fully preserve the structural features of assessment information from different sources in the functional partition, muscle group, and training prescription dimensions, pooling or dimensionality reduction of the aforementioned information matrices is not performed. Instead, multi-source information is directly concatenated to construct a unified multi-source fusion representation. The fifth, sixth, and fourth information matrices are vertically concatenated along the row dimension to obtain the fused information matrix.

[0137]

[0138] Where vcon(⋅) represents the vertical concatenation operation, which is used to combine the embedded representations from different information sources into a unified matrix according to their functional unit order;

[0139] To obtain a unified embedding representation for subsequent enhanced retrieval and prescription generation, a trainable mapping network is introduced to perform feature alignment and transformation on the fusion information matrix;

[0140] The mapping network is a fully connected neural network, a linear mapping, or a combination thereof, with an input dimension of L1 and an output dimension of L1. A unified embedding output matrix is ​​obtained by performing a row-by-row mapping transformation on the fused information matrix.

[0141]

[0142] Where Z represents a unified embedded representation that integrates patient basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information, and NN represents a fully connected network; the unified embedded output matrix Z serves as a structured representation of the patient's current comprehensive motor function status and is used in subsequent enhancement retrieval and individualized training prescription generation steps.

[0143] Furthermore, step 4 specifically includes the following steps:

[0144] Step 4-1: Knowledge Base Establishment and Embedding; To realize the generation of individualized rehabilitation training prescriptions based on multi-source assessment embedding representation, this invention pre-constructs a multi-level knowledge base system related to rehabilitation training and performs vectorized embedding representation on the knowledge base content to support the subsequent enhanced retrieval and decision generation process; The knowledge base system includes at least: an assessment-prescription rule base, a training item knowledge base, and a safety criterion base;

[0145] Step 4-2, Candidate Prescription Generation (using RAG to generate candidate results from three knowledge bases); This invention proposes a phased enhanced retrieval method based on the final assessment result matrix Z; After completing the fusion and embedding of multi-source assessment information, the final assessment result matrix Z for rehabilitation decision-making is obtained. The matrix represents the comprehensive characteristics of the patient's current motor function status in macroscopic functional areas, microscopic muscle groups, and historical training dimensions; Based on the final assessment result matrix Z, candidate rehabilitation training prescriptions are generated in the order of "rule priority - item refinement - safety constraints";

[0146] Step 4-3: Generate the final prescription and output a structured prescription based on triple constraints of rules, items, and security;

[0147] Rule candidate set: R candidate : Represents the set of evaluation-prescription rules that match the final evaluation result; rule-driven candidate training item set: T candidate The training item set obtained through rule-item secondary similarity retrieval; the candidate prescription set for security alerts: S candidate This refers to the set of training items that are deemed risky, require warnings, be downgraded, or prohibited during the safety criteria verification process.

[0148] Based on the final evaluation results, the rule candidate set R is obtained. candidate A rule-driven candidate training item set T is generated through the rule-item mapping relationship. candidate The two are matched one by one; then, safety criteria are introduced to verify the safety of candidate training items, resulting in a set S of candidate prescriptions for safety warnings. candidate S candidate ⊆T candidate ; via T candidate Remove S from the middle candidate The training items in the dataset are used to obtain the final executable training item set D;

[0149]

[0150] Where m represents the final number of training prescriptions, d i The representative training prescription is composed of training item t. i and project rules r i composition.

[0151] Furthermore, step 4-1 specifically includes the following steps:

[0152] Step 4-1-1, Assessment—Establishment and embedding of the prescription rule base (to find the training items and intensity applicable to the patient).

[0153] The evaluation-prescription rule base describes the mapping relationship between evaluation results and training items. Its core function is to transform evaluation results at different levels into constraints or preference information for training prescription generation. Each evaluation-prescription rule includes at least the following structured fields: the mapping rule between evaluation embedded features and training items (evaluation state constraints); the correspondence between macro-evaluation results and the specified training intensity range of training items; the correspondence between micro-evaluation results and the specified training intensity range of training items; and the rule for the j-th evaluation-prescription rule. j Simplified representation is r j The above fields are encoded and embedded respectively:

[0154] Evaluation state constraint embedding: Encode the evaluation embedding feature constraints involved in the rule into a vector r. e (j) And mapped to an embedding vector :

[0155]

[0156] Macro-evaluation—Intensity Mapping Embedding: Encodes the mapping relationship between macro-evaluation results and training intensity intervals into a vector r. m (j) And mapped to an embedding vector :

[0157]

[0158] Microscopic Assessment-Intensity Mapping Embedding: Encodes the mapping relationship between microscopic electromyography assessment results and training intensity ranges into a vector r. u (j) And mapped to an embedding vector :

[0159]

[0160] Subsequently, the three types of rule embedding vectors are concatenated, and dimension alignment is performed using a trainable mapping network to obtain the rule embedding vector. :

[0161]

[0162] Where hcon represents the horizontal splicing operation, W r W e W m W u b is a trainable weight matrix e b m b r b u For trainable bias, Φ represents a nonlinear mapping function (activation function);

[0163] All rule embedding vectors constitute the evaluation-prescription rule base embedding matrix:

[0164]

[0165] Where R represents the number of assessment-prescription rule entries;

[0166] Step 4-1-2: Establishment and embedding of the training project knowledge base (used to find specific operational details of the training project).

[0167] The training program knowledge base stores executable rehabilitation training programs, with the basic unit being a training program entry. Each training program entry includes at least the following attribute information: training program name; target functional area or target muscle group; key points of movement execution. For the i-th training program entry, the above attribute information is embedded as follows:

[0168] The training item names are categorically encoded into action embedding vectors. :

[0169]

[0170] The target functional zones or target muscle groups are encoded into zone / muscle group embedding vectors. :

[0171]

[0172] It structures and encodes the key points of action execution into r c (i) And mapped to key embedding vectors :

[0173]

[0174] The three types of embedding vectors are concatenated and projected onto a unified dimensional space through a trainable mapping network to obtain the training item embedding vectors. :

[0175]

[0176] Where hcon represents the horizontal splicing operation, W t W c b is a trainable weight matrix t b c For trainable bias, Φ represents a nonlinear mapping function (activation function);

[0177] The embedding vectors of all training items constitute the embedding matrix of the training item knowledge base:

[0178]

[0179] Where T represents the number of training items;

[0180] Step 4-1-3: Establishment and embedding of the security criteria library (used to exclude unsafe training items).

[0181] A safety criterion library is used to constrain the training prescription generation process, preventing the generation of training plans that exceed the patient's current exercise capacity or pose risks. Each safety criterion is one of the following: the maximum permissible training intensity corresponding to a certain stage of disease progression (assessment result); contraindicated movements under abnormal conditions of specific muscle groups; individual safety threshold limits based on the patient's basic information; and contraindications for rehabilitation training of a certain complication. The k-th safety criterion is parameterized and encoded to obtain the safety constraint vector s. k And mapped to a security criterion embedding vector through an embedding mapping function:

[0182]

[0183] Among them, W s b is a trainable weight matrix s For trainable bias, Φ represents a nonlinear mapping function (activation function);

[0184] All security criterion embedding vectors constitute the security criterion library embedding matrix:

[0185]

[0186] Where S represents the number of safety criteria.

[0187] Furthermore, step 4-2 specifically includes the following steps:

[0188] Step 4-2-1: Rule base retrieval based on the final evaluation result matrix (determine the applicable training item categories and training frequency and intensity range);

[0189] A unified embedded representation Z that integrates basic patient information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information:

[0190]

[0191] Among them, different rows correspond to historical prescription information, macroscopic functional zoning assessment information, and microscopic muscle group assessment information, respectively.

[0192] To achieve similarity calculation with the evaluation-prescription rule embedding vector, the final evaluation result matrix is ​​first dimensionally aligned and the query vector is constructed.

[0193] Rule query vector construction; mapping the final evaluation result matrix Z to the rule embedding space using a trainable mapping function:

[0194]

[0195] Among them, W q ∈R L1×L2 b is a trainable weight matrix q ∈R L2 For bias, 1∈R (K+N+M)×1 Let q be an all-1 vector, and Φ(⋅) be a non-linear mapping function; construct a rule-based retrieval query vector, where the i-th row has 1 ≤ i ≤ M + N + K, and the query vector is q. i :

[0196]

[0197] Based on rule-based retrieval query vectors, and embedded vectors of the evaluation-prescription rule base. If 1 ≤ j ≤ R, the similarity matrix is ​​calculated to obtain the similarity S. ij rule :

[0198]

[0199] Where sim is the similarity function, which in some embodiments can be the cosine similarity function; to obtain the overall matching degree of the j-th rule (1≤j≤R) in the current evaluation state, the similarity of each row of rules is aggregated to obtain the rule activation degree s. j rule :

[0200]

[0201] Where, α i Weighting coefficients related to the degree of anomaly assessment are used to reflect the importance of different row assessment units to rule matching; scores s are selected. j rule Rule entries that exceed a preset threshold or rank highly (the top k entries, k ≤ R) form a candidate subset R. candidate：

[0202]

[0203] Based on the subset of candidate rules, determine the applicable training item categories, as well as the corresponding training frequency range and training intensity range, under the current evaluation state;

[0204] Step 4-2-2: Knowledge base retrieval of training items based on rule retrieval results; obtaining candidate rule subset R. candidate Next, in order to determine specific executable training items within the constraints of the rules, this step introduces a rule-item secondary similarity retrieval mechanism to achieve a fine mapping from rule-level recommendations to specific training items;

[0205] Candidate rules are obtained by corresponding rule embedding vectors; for a subset R of candidate rules candidate Each assessment-prescription rule in the document j For each rule, 1 ≤ j ≤ k, obtain its corresponding rule embedding vector:

[0206]

[0207] The rule embedding vector is used to characterize the applicable conditions and training parameter constraints of the training item described by the rule;

[0208] The embedding representation of the training project knowledge base; the embedding matrix of the training project knowledge base is:

[0209]

[0210] Rule-based secondary similarity calculation for each item; embedding vector for each candidate rule. Calculate the embedding vector for each candidate rule. The similarity S between the embedding vectors of each training item in the training item knowledge base and the training item knowledge base jttrain :

[0211]

[0212] Where sim(⋅) is the similarity function; the training items corresponding to the rule are determined; for the j-th candidate rule, based on the rule-item similarity calculation result, the training item with the highest similarity score is selected as the training item that best matches the rule:

[0213]

[0214] Perform the above operations on all k candidate rules, thereby establishing a one-to-one correspondence between candidate rules and specific training items, and obtaining a rule-driven candidate training item set T. candidate :

[0215]

[0216] Step 4-2-3: Retrieve safety guidelines for integrating evaluation results and training project information;

[0217] To avoid generating risky training schemes, after obtaining the set of candidate training items, a security criterion library is introduced to verify the security of the candidate training items; for each candidate training item t... i Embed it into vector e train (i) With the final evaluation result matrix after embedding By merging the data, a joint query representation for security verification is constructed:

[0218]

[0219] Based on joint query representation The similarity between the vector and the embedding vector of the security criteria library is calculated to obtain S. ip safe :

[0220]

[0221]

[0222] When the activation level of a safety criterion exceeds a preset risk threshold, the project (based on the strength of the cooperation already achieved) is identified as a warning project; after safety screening, a candidate prescription set S for safety warnings is obtained. candidate , where n is the number of warning items.

[0223] Compared with existing technologies, the significant advancements of this invention are: 1) Proposing a "hierarchical localization" integrated method for rehabilitation assessment and prescription generation: This invention expands the rehabilitation assessment process from the traditional single overall scoring to a hierarchical assessment process of macroscopic motor function assessment—abnormal functional area localization—microscopic electromyography function assessment—abnormal muscle group localization, and uses the hierarchical assessment results as direct input for training decisions, achieving deep coupling between assessment and prescription generation, and avoiding the subjective mapping between assessment results and training prescriptions; 2) A multimodal collaborative assessment mechanism based on macroscopic movement and microscopic electromyography: This invention, for the first time, integrates inertial motion sensing data and surface electromyography data in rehabilitation assessment. The assessment stage utilizes the following collaborative methods: macroscopic motion data is used for overall motor ability assessment and abnormal functional area localization, while microscopic electromyography (EMG) data is used for local neuromuscular function assessment and abnormal muscle group identification. These two methods mutually constrain and refine each other step by step, significantly improving the accuracy and stability of abnormal motor function localization; 3) A structured embedding and modulation fusion method for multi-source assessment information is introduced: This invention embeds patient basic information, macroscopic motion assessment information, microscopic EMG assessment information, and previous training prescription information into separate representations. A weighted fusion of macroscopic and microscopic assessment information is achieved through a gating modulation mechanism driven by patient basic information, preserving functional area and muscle group structure information while... This invention achieves differentiated influence of different information sources on prescription generation, improving the model's ability to perceive individual differences; 4) It proposes an evaluation-prescription rule enhancement retrieval method based on row-level queries: Unlike existing methods that generate prescriptions based on overall features or simple rule matching, this invention constructs a row-level rule query vector based on the final evaluation result matrix and performs row-by-row similarity matching with the evaluation-prescription rule base, achieving a fine perception of rules for specific functional areas or muscle group abnormalities, thereby improving the pertinence and interpretability of rule triggering; 5) It adopts a prescription refinement strategy of "rule-item secondary similarity retrieval": This invention determines the training item category and parameters in the rule retrieval process. After defining the scope, a secondary similarity calculation between rule embedding and training item embedding is further introduced to achieve a precise mapping from abstract rules to specific executable training items, avoiding the limitations of relying solely on fixed rules or manually configured items; 6) Constructing a safety criterion-driven prescription screening and elimination mechanism: This invention constructs a safety verification query vector by fusing evaluation results with training item information, and performs similarity matching with the safety criterion library to identify training items with risks, forming a set of safety warning candidate prescriptions, and eliminating unsafe items from the rule-driven candidate items through set difference operation, thereby significantly improving the safety of prescription execution while ensuring the training's relevance.

[0224] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0225] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0226] Figure 1 This is a schematic diagram of the overall system framework. Detailed Implementation

[0227] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0228] This invention enables a direct mapping of rehabilitation assessment results to the training prescription generation process, allowing the generated training prescriptions to be configured differently for abnormalities at different functional levels and functional zones. This improves the relevance, interpretability, and safety of rehabilitation training, reduces reliance on human experience, and is suitable for rehabilitation assessment and training guidance for patients with motor dysfunction.

[0229] 1. Overall Method Flow

[0230] like Figure 1 As shown, the present invention provides a hierarchical positioning rehabilitation assessment and individualized training prescription generation method, including a motor function data acquisition module, a hierarchical rehabilitation assessment module, a multi-source assessment information fusion and embedding module, and a training prescription generation module. The modules cooperate with each other to realize the generation of individualized rehabilitation training prescriptions based on hierarchical assessment results.

[0231] S1. Data Acquisition: This includes inertial motion data acquisition and electromyography (EMG) data acquisition. Macroscopic motion data and microscopic EMG data are collected from the subjects during the execution of preset assessment actions. The collected data are preprocessed to construct motor function characteristic data for subsequent analysis.

[0232] S2. Layered Rehabilitation Assessment: Based on motor function characteristic data, the subject's motor function status is assessed in a layered manner, completing the overall motor function assessment, macroscopic abnormal function zoning and abnormal muscle groups and synergistic control abnormality identification tasks, and outputting structured layered motor function assessment results.

[0233] S3. Multi-source assessment information fusion and embedding: Encode, embed, and modulate the patient's basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information to construct a unified multi-source assessment embedding representation to characterize the patient's current comprehensive motor function status.

[0234] S4. Training prescription generation: Based on multi-source assessment embedding representation, combined with assessment-prescription rule base, training item knowledge base and safety criterion base, the enhanced retrieval mechanism sequentially completes the tasks of rule matching, training item matching and safety verification to generate an individualized rehabilitation training prescription that matches the patient's current motor function status.

[0235] Through the above four modules, this invention uses the results of hierarchical rehabilitation assessment as the basis for training decisions, and directly maps the results of motor function assessment to the rehabilitation training prescription generation process, so that the training prescription can be configured differently for abnormalities in different functional levels and functional areas, thereby improving the pertinence and effectiveness of rehabilitation training.

[0236] 2. Technical Details

[0237] 2.1 Macroscopic motion data acquisition and assessment positioning

[0238] 2.1.1 Sensor Configuration and Macroscopic Motion Data Acquisition

[0239] In the macroscopic motor function assessment phase, this invention employs an inertial measurement unit (IMU) to collect macroscopic motion data during the subject's movement. The IMU sensor includes at least a three-axis accelerometer and a three-axis angular velocity meter, and optionally a three-axis magnetometer, for collecting acceleration, angular velocity, and spatial orientation information of the limbs during movement.

[0240] In practice, the IMU sensors are worn on key limb segments of the subject, including at least: upper limbs: upper arm and forearm (optionally worn on the back of the hand); lower limbs: thigh and calf (optionally worn on the dorsum of the foot). The IMU sensors continuously acquire motion data at a fixed sampling frequency fs, and the acquired 12-channel macroscopic motion data signals include:

[0241] Triaxial acceleration signal a(t):

[0242]

[0243] Among them, a x (t) represents the acceleration in the x-axis direction, a y (t) represents the acceleration in the y-axis direction, a z (t) represents the acceleration along the z-axis; the triaxial angular velocity signal w(t) is:

[0244]

[0245] Among them, w x (t) represents the angular velocity along the x-axis, w y(t) represents the angular velocity along the y-axis, w z (t) represents the angular velocity along the z-axis; the triaxial magnetic field strength signal m(t):

[0246]

[0247] Where, m x (t) represents the magnetic field strength along the x-axis, m y (t) represents the magnetic field strength along the y-axis, m z (t) represents the magnetic field strength in the z-axis direction; the collected raw motion data is processed by time synchronization, noise reduction filtering and coordinate system calibration, and the limb posture information is calculated based on sensor fusion algorithms (such as complementary filtering or Kalman filtering) to obtain the posture sequence represented in quaternion or Euler angle form;

[0248]

[0249] Where, θ x (t) represents the rotation angle about the x-axis, θ y (t) represents the rotation angle about the y-axis, θ z (t) represents the rotation angle about the z-axis.

[0250] 2.1.2 Construction and Evaluation of Macroscopic Motion Characteristics

[0251] Based on the attitude calculation results and angular velocity information, a set of kinematic features for macroscopic evaluation is constructed. For a 12-channel signal acquired by an inertial measurement sensor, the macroscopic motion features include:

[0252] Activity range ROM∈R 12 :

[0253]

[0254] Where max represents the maximum value, min represents the minimum value, ROMi represents the ROM of the i-th channel signal; average activity ∈R12:

[0255]

[0256] in, Represents the signal of the i-th channel ;

[0257] Multi-joint synergy feature matrix C∈R 12×12 :

[0258]

[0259] Where corr represents the correlation coefficient, Cij This represents the multi-joint synergy between the i-th and j-th signals;

[0260] Where x(t) represents the signal from a single channel acquired by the IMU, which is any one of the 12 channels of macroscopic motion data; x i (t), x j (t) represents two different signals from the 12-channel macroscopic motion data signal;

[0261] Based on macroscopic movement characteristics, a macroscopic assessment of the overall motor function status of the subjects is performed; a macroscopic motor function score S is calculated. macro ∈R:

[0262]

[0263] Where: f k Represents the k-th macroscopic motion characteristic; ω k The corresponding feature weights reflect the contribution of each feature to overall motor function; N represents the number of features; the macroscopic motor function score reflects the overall motor ability level of the subject and serves as the basic input for subsequent stratified assessment; the above features together constitute the macroscopic motor feature set F. macro :

[0264]

[0265] 2.1.3 Macroscopic Anomaly Location

[0266] Based on macroscopic motor function assessment, this invention further analyzes the distribution differences of macroscopic motor characteristics in different limb segments or functional zones to achieve abnormal localization at the macroscopic level.

[0267] Specifically, for each macroscopic anatomical region (corresponding to one IMU sensor), the corresponding local macroscopic feature vector is calculated:

[0268]

[0269] And calculate its relationship with the set of health reference features. Distance between:

[0270]

[0271] The function dis(a1,a2) represents a vector formed by calculating the Euclidean distance between the elements in sets a1 and a2. When the distance Dr corresponding to a certain functional region exceeds a preset threshold, that functional region is determined to be a macroscopic motor abnormality region. This method enables the macroscopic localization of abnormal motor function regions in the subject, guiding the subsequent acquisition of microscopic electromyographic data and the microscopic functional assessment process.

[0272] 2.2 Acquisition and Evaluation of Microscopic Electromyographic Data

[0273] 2.2.1 Electromyography (EMG) sensor configuration and data acquisition

[0274] After completing the macroscopic motor function assessment and abnormality localization, this invention further acquires microscopic electromyographic data for the identified abnormal limb segments or functional zones to analyze the local neuromuscular functional state. The microscopic electromyographic data is acquired using surface electromyography (sEMG) sensors.

[0275] The surface electromyography (EMG) sensor is preferably a bipolar or multipolar surface electrode, which is attached to the surface of key muscle groups within the target functional area and arranged along the muscle fiber direction. In specific implementation, the EMG signal acquisition parameters can be set as follows:

[0276] Sampling frequency: f emg =1000~2000Hz

[0277] Analog bandpass filter range: 20–450Hz;

[0278] Power frequency notch filter: 50Hz or 60Hz.

[0279] The raw electromyographic signals acquired at the abnormal zonal locations are represented as follows:

[0280]

[0281] Where x i (t) represents the electromyographic signal of the i-th target muscle at time t, and M is the number of muscles collected. The original electromyographic signals are denoised, rectified, and smoothed to obtain the electromyographic signal sequence used for feature extraction.

[0282] 2.2.2 Algorithm for Microscopic Electromyographic Feature Extraction and Microscopic Neuromuscular Function Assessment

[0283] Based on the preprocessed electromyographic signals, microscopic electromyographic features that reflect muscle activation level, fatigue state, and synergistic control characteristics are extracted.

[0284] Microscopic electromyographic characteristics include at least one or more of the following:

[0285] Root mean square value RMS∈R M RMS i RMS representing the i-th lead:

[0286]

[0287] Mean Absolute Value (MAV) ∈ R M MAV iMAV representing the i-th lead:

[0288]

[0289] Muscle activation normalization index A∈R M A i A, representing the i-th lead, normalizes the electromyographic amplitude relative to the maximum voluntary contraction MVC:

[0290]

[0291] Based on microscopic electromyographic feature vectors, a microscopic assessment of neuromuscular functional status within a target functional area is performed. In one implementation, a microscopic muscle function score S is constructed using a weighted method. micro ∈R:

[0292]

[0293] in: These are the weighting coefficients for the corresponding features; the weights are used to reflect the degree of contribution of different muscle groups and features to local motor function. The microscopic assessment results are used to quantify whether the muscle activation level is insufficient, or whether there is abnormal enhancement or fatigue trend.

[0294] 2.2.3 Localization of Abnormal Muscle Groups and Synergistic Patterns

[0295] To further identify the specific muscle factors leading to macroscopic motor abnormalities, this invention performs abnormal muscle group localization analysis based on microscopic assessment. In one embodiment, a muscle group synergy matrix C is constructed by comparing the relative relationships between the activation levels of different muscle groups. emg ∈R M×M Its elements Represented as:

[0296]

[0297] When the activation level of a certain muscle group is significantly lower or higher than that of other muscle groups in the same functional area, or when the correlation between it and the main synergistic muscle group is lower than a preset threshold, the muscle group is determined to be an abnormal muscle group.

[0298] Simultaneously, the abnormal functional zoning information obtained in the macroscopic assessment stage can be combined to constrain the microscopic assessment results, thereby improving the accuracy and stability of abnormal muscle group localization. Through the above-mentioned microscopic electromyography data acquisition and assessment process, a set of microscopic neuromuscular function assessment results within the target functional zoning is output, and the results include at least: a set of abnormal muscle groups;

[0299] Indicators of activation levels in various muscle groups; abnormal features of synergistic control F micro , can be represented as:

[0300]

[0301] 2.3 Multi-source assessment information fusion and embedding

[0302] 2.3.1 Word embedding of patient basic information (first information embedding)

[0303] The first piece of information is the patient's basic information, which includes at least the patient's age, gender, affected side, disease stage, and other basic attributes related to rehabilitation training. This basic patient information is encoded and embedded into a vector inf1 of length L.

[0304]

[0305] Further construct a trainable transformation matrix S1:

[0306]

[0307] And the first information matrix Inf1 is constructed using a weighted method:

[0308]

[0309] 2.3.2 Word Embedding of Macroscopic Motion Function Information (Second Information Embedding)

[0310] The second information is macroscopic motor function assessment information, which originates from the macroscopic motor data acquisition and assessment positioning steps, including macroscopic motor characteristics of different limb segments or functional zones. This macroscopic motor function assessment information is embedded into a second information matrix Inf2 of size N×L1:

[0311]

[0312] Where N represents the number of macroscopic functional zones or macroscopic anatomical zones; each row represents the macroscopic motion feature embedding representation corresponding to a functional zone, that is, the set of macroscopic motion features. Each element is expanded and concatenated before being embedded, with a length of L1.

[0313] 2.3.3 Word Embedding of Microscopic Electromyographic Functional Information (Third Information Embedding)

[0314] The third piece of information is microscopic electromyography (EMG) function assessment information, which originates from the microscopic EMG data acquisition and assessment process, including the activation level, synergistic relationships, and fatigue characteristics of the target muscle group. This microscopic EMG function assessment information is embedded into a third information matrix Inf3 of size M×L1:

[0315]

[0316] Where M represents the number of target muscle groups or microscopic functional units; L1 represents the dimension of the microscopic electromyographic feature embedding, i.e., the set of microscopic neuromuscular function assessment results F micro Each element is expanded and concatenated to form a vector of length L1.

[0317] 2.3.4 Word embedding of previously trained prescription information (fourth information embedding)

[0318] The fourth piece of information is the previous training prescription information, which includes the types of training exercises, training intensity, training frequency, and feedback on training effects. This previous training prescription information is embedded into a fourth information matrix Inf4 of size K×L1.

[0319]

[0320] Where K represents the number of items in the prescription, and L1 is the length of the execution result embedding for each item.

[0321] 2.3.5 Multi-source assessment information fusion

[0322] (1) Modulation and fusion of patient basic information and macroscopic motion information (to obtain Inf5)

[0323] The patient's basic information embedding vector is:

[0324]

[0325] The macroscopic motion information matrix is ​​as follows:

[0326]

[0327] 1) Trainable mapping of basic information and generation of gated vectors

[0328] Introducing a trainable weight matrix W g1 ∈R L1×L1 With bias b g1 ∈R L1 The patient's basic information is mapped to a gating vector g1 (modulation vector):

[0329]

[0330] Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1].

[0331] 2) Trainable feature transformation of macroscopic information

[0332] Introducing the macroscopic feature transformation matrix W m1 ∈R L1×L1 With bias b m1 ∈R L1By transforming the macroscopic information line by line, we obtain :

[0333]

[0334] Where 1∈R N×1 It is an all-1 vector used for broadcasting bias terms.

[0335] 3) Gated modulation and residual fusion

[0336] The gate vector g1 is broadcast row-wise to N×L1 to achieve weighted modulation of macroscopic information by patient information. A residual term is introduced to improve stability, resulting in the fifth information matrix Inf5, which is a modulation and fusion of basic patient information and macroscopic motion information.

[0337]

[0338] in For element-wise multiplication, λ is a configurable or trainable scalar coefficient used to control the residual fusion ratio.

[0339] (2) Modulation and fusion of patient basic information and microscopic electromyographic information (to obtain Inf6)

[0340] The patient's basic information embedding vector is:

[0341]

[0342] The macroscopic motion information matrix is ​​as follows:

[0343]

[0344] 1) Trainable mapping of basic information and generation of gated vectors

[0345] Introducing a trainable weight matrix W g2 ∈R L1×L1 With bias b g2 ∈R L1 Patient basic information is mapped to a gating vector (modulation vector):

[0346]

[0347] Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1].

[0348] 2) Trainable feature transformation of macroscopic information

[0349] Introducing the micro-feature transformation matrix W m2 ∈R L1×L1 With bias b m2 ∈R L1 The microscopic information is obtained by transforming it line by line. :

[0350]

[0351] Where 1∈R N×1 It is an all-1 vector used for broadcasting bias terms.

[0352] 3) Gated modulation and residual fusion

[0353] The gating vector g2 is broadcast row-wise to N×L1 to achieve weighted modulation of microscopic information based on patient information. A residual term is introduced to improve stability. The sixth information matrix Inf6, which fuses the modulation of basic patient information and microscopic electromyographic information, is as follows:

[0354]

[0355] in For element-wise multiplication, λ is a configurable or trainable scalar coefficient used to control the residual fusion ratio.

[0356] (3) Output of fusion and embedding results

[0357] After modulating and fusing patient basic information with macroscopic motor function information, and modulating and fusing patient basic information with microscopic electromyographic function information, the fifth and sixth information matrices are obtained, as follows:

[0358]

[0359]

[0360] Meanwhile, the previously trained prescription information is used to obtain a fourth information matrix after word embedding:

[0361]

[0362] To fully preserve the structural features of assessment information from different sources in the functional partition, muscle group, and training prescription dimensions, this invention does not perform pooling or dimensionality reduction on the aforementioned information matrices. Instead, it directly concatenates the multi-source information to construct a unified multi-source fusion representation. Specifically, the fifth, sixth, and fourth information matrices are vertically concatenated along the row dimension to obtain the fused information matrix:

[0363]

[0364] Here, vcon(⋅) represents the vertical concatenation operation, which is used to combine the embedded representations from different information sources into a unified matrix according to their functional unit order.

[0365] In one implementation, to obtain a unified embedding representation for subsequent enhanced retrieval and prescription generation, a trainable mapping network is introduced to perform feature alignment and transformation on the fused information matrix.

[0366] The mapping network is a fully connected neural network, a linear mapping, or a combination thereof, with an input dimension of L1 and an output dimension of L1. A unified embedding output matrix is ​​obtained by performing a row-by-row mapping transformation on the fused information matrix.

[0367]

[0368] Where Z represents a unified embedded representation that integrates patient basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information, and NN represents a fully connected network; the unified embedded output matrix Z serves as a structured representation of the patient's current comprehensive motor function status and is used in subsequent enhancement retrieval and individualized training prescription generation steps.

[0369] 2.4 Enhanced Search and Prescription Generation

[0370] 2.4.1 Knowledge Base Establishment and Embedding

[0371] To achieve personalized rehabilitation training prescription generation based on multi-source assessment embedding representation, this invention pre-constructs a multi-level knowledge base system related to rehabilitation training and performs vectorized embedding representation on the knowledge base content to support subsequent enhanced retrieval and decision generation processes. The knowledge base system includes at least: an assessment-prescription rule base, a training item knowledge base, and a safety criterion base.

[0372] (1) Assessment - Establishment and embedding of the prescription rule base (to find the appropriate training items and intensity for the patient)

[0373] The evaluation-prescription rule base describes the mapping relationship between evaluation results and training items. Its core function is to transform evaluation results at different levels into constraints or preference information for training prescription generation. Each evaluation-prescription rule includes at least the following structured fields: the mapping rule between evaluation embedded features and training items (evaluation state constraints); the correspondence between macro-level evaluation results and the specified training intensity range for training items; and the correspondence between micro-level evaluation results and the specified training intensity range for training items. For the j-th evaluation-prescription rule... j Simplified representation is r j The above fields are encoded and embedded respectively:

[0374] 1) Evaluation state constraint embedding: Encode the evaluation embedding feature constraints involved in the rule into a vector r e (j) And mapped to an embedding vector :

[0375]

[0376] 2) Macro-evaluation-intensity mapping embedding: Encode the mapping relationship between the macro-evaluation results and the training intensity intervals into a vector r. m (j) And mapped to an embedding vector :

[0377]

[0378] 3) Microscopic Assessment-Intensity Mapping Embedding: The mapping relationship between microscopic electromyography assessment results and training intensity ranges is encoded into a vector r. u (j) And mapped to an embedding vector :

[0379]

[0380] Subsequently, the three types of rule embedding vectors are concatenated, and dimension alignment is performed using a trainable mapping network to obtain the rule embedding vector. :

[0381]

[0382] Where hcon represents the horizontal splicing operation, W r W e W m W u b is a trainable weight matrix e b m b r b u For trainable bias, Φ represents the nonlinear mapping function (activation function).

[0383] All rule embedding vectors constitute the evaluation-prescription rule base embedding matrix:

[0384]

[0385] Where R represents the number of assessment-prescription rule entries.

[0386] (2) Establishment and embedding of the training project knowledge base (used to find specific operational details of the training project)

[0387] The training program knowledge base is used to store executable rehabilitation training programs, and its basic unit is the training program entry.

[0388] Each training item entry includes at least the following attribute information: training item name; target functional area or target muscle group; key points of movement execution. For the i-th training item entry, the above attribute information is embedded as follows:

[0389] The training item names are categorically encoded into action embedding vectors:

[0390]

[0391] Encode the target functional area or target muscle group into a area / muscle group embedding vector:

[0392]

[0393] It structures and encodes the key points of action execution into r c (i) And mapped to key embedding vectors :

[0394]

[0395] The three types of embedding vectors are concatenated and projected onto a unified dimensional space through a trainable mapping network to obtain the training item embedding vectors. :

[0396]

[0397] Where hcon represents the horizontal splicing operation, W t W c b is a trainable weight matrix t b c For trainable bias, Φ represents the nonlinear mapping function (activation function).

[0398] The embedding vectors of all training items constitute the embedding matrix of the training item knowledge base:

[0399]

[0400] Where T represents the number of training items.

[0401] (3) Establishment and embedding of a security criteria library (to exclude unsafe training items)

[0402] A safety criterion library is used to constrain the training prescription generation process, preventing the generation of training plans that exceed the patient's current exercise capacity or pose risks. Each safety criterion may be one of the following: the maximum permissible training intensity corresponding to a certain stage of disease progression (assessment result); contraindicated movements under abnormal conditions of specific muscle groups; individual safety threshold limits based on the patient's basic information; or contraindications for rehabilitation training of a certain complication. The k-th safety criterion is parameterized and encoded to obtain the safety constraint vector s. k And mapped to a security criterion embedding vector through an embedding mapping function:

[0403]

[0404] Among them, W s b is a trainable weight matrix s For trainable bias, Φ represents the nonlinear mapping function (activation function).

[0405] All security criterion embedding vectors constitute the security criterion library embedding matrix:

[0406]

[0407] Where S represents the number of safety criteria.

[0408] 2.4.2 Candidate Prescription Generation (using RAG to generate candidate results from the three major knowledge bases, but requiring more creativity)

[0409] This invention proposes a phased enhanced retrieval method based on the final assessment result matrix Z. After completing the fusion and embedding of multi-source assessment information, this invention obtains the final assessment result matrix Z for rehabilitation decision-making. The matrix represents the comprehensive characteristics of the patient's current motor function status in terms of macroscopic functional zones, microscopic muscle groups, and historical training dimensions. Based on the final assessment result matrix Z, this invention generates candidate rehabilitation training prescriptions in the order of "rule priority—item refinement—safety constraints," specifically including the following steps.

[0410] (1) Rule base retrieval based on the final evaluation result matrix (determining the applicable training project categories and training frequency and intensity range)

[0411] A unified embedded representation Z that integrates basic patient information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information:

[0412]

[0413] Among them, different rows correspond to historical prescription information, macroscopic functional zoning assessment information, and microscopic muscle group assessment information, respectively.

[0414] To achieve similarity calculation with the evaluation-prescription rule embedding vector, the final evaluation result matrix is ​​first dimensionally aligned and the query vector is constructed.

[0415] 1) Construction of rule query vectors

[0416] The final evaluation matrix Z is mapped to the rule embedding space using a trainable mapping function:

[0417]

[0418] Among them, W q ∈R L1×L2 b is a trainable weight matrix q ∈R L2 For bias, 1∈R (K+N+M)×1 Let q be an all-1 vector, and Φ(⋅) be a non-linear mapping function; construct a rule-based retrieval query vector, where the query vector q is in the i-th row (1≤i≤M+N+K). i :

[0419]

[0420] 2) Rule similarity calculation and rule subset selection

[0421] Based on rule-based retrieval query vectors, and embedded vectors of the evaluation-prescription rule base. The similarity S is obtained by calculating the similarity matrix (1≤j≤R). ij rule :

[0422]

[0423] Where sim is the similarity function, which in some implementations can be the cosine similarity function. To obtain the overall matching degree of the j-th (1≤j≤R) rule in the current evaluation state, the similarity of each rule is aggregated to obtain the rule activation degree s. j rule :

[0424]

[0425] Where, α i The weighting coefficients are related to the degree of anomaly assessment and are used to reflect the importance of different row assessment units to rule matching.

[0426] Select score s j rule Rule entries that exceed a preset threshold or rank highly (preset k, k≤R) form a candidate subset R. candidate：

[0427]

[0428] Based on a subset of candidate rules, determine the applicable training item categories, as well as the corresponding training frequency range and training intensity range, under the current evaluation state.

[0429] (2) Knowledge base retrieval of training items based on rule retrieval results

[0430] In obtaining the candidate rule subset R candidate Subsequently, in order to determine specific executable training items within the constraints of the rules, this step introduces a rule-item secondary similarity retrieval mechanism to achieve a fine mapping from rule-level recommendations to specific training items.

[0431] 1) Obtaining the embedding vector of the corresponding candidate rule

[0432] For the candidate rule subset R candidate Each assessment-prescription rule in the document j (1≤j≤k), obtain the corresponding rule embedding vector:

[0433]

[0434] The rule embedding vector is used to characterize the applicable conditions and training parameter constraints of the training project described by the rule.

[0435] 2) Training project knowledge base embedding representation

[0436] The embedding matrix of the training project knowledge base is as follows:

[0437]

[0438] 3) Rule - Item Secondary Similarity Calculation

[0439] For each candidate rule embedding vector e, calculate the similarity matrix S∈R between it and the embedding vectors of each training item in the training item knowledge base. k×T Matrix element S jt train :

[0440]

[0441] Wherein, sim(⋅) is the similarity function, which in one implementation is the cosine similarity function.

[0442] 4) Determining the training items corresponding to the rules

[0443] For the j-th candidate rule, based on the rule-item similarity calculation results, the training item with the highest similarity score is selected as the training item that best matches the rule:

[0444]

[0445] Perform the above operations on all k candidate rules, thereby establishing a one-to-one correspondence between candidate rules and specific training items, and obtaining a rule-driven candidate training item set T. candidate :

[0446]

[0447] (3) Safety criteria retrieval of integrated assessment results and training program information

[0448] To avoid generating risky training schemes, after obtaining the set of candidate training items, a security criterion library is introduced to perform security verification on the candidate training items. For each candidate training item t... i Embed it into vector e train (i) The matrix Z, after embedding, is fused to construct a joint query representation for security verification:

[0449]

[0450] Based on joint query representation The similarity between the vector and the embedding vector of the security criteria library is calculated to obtain S. ip safe :

[0451]

[0452] When the activation level of a safety criterion exceeds a preset risk threshold, the project (based on the strength to which the project has already been implemented) is designated as a warning project. After safety screening, a candidate prescription set S for safety warnings is obtained. candidate , where n is the number of warning items:

[0453]

[0454] 2.4.3 Final Prescription Generation (Structured Prescription Output Based on Rules, Items and Safety Triple Constraints)

[0455] Rule candidate set: R candidate : Represents the set of evaluation-prescription rules that match the final evaluation result. Rule-driven candidate training item set: T candidate The training item set obtained through rule-item secondary similarity retrieval. Candidate prescription set for security alerts: S candidate This refers to the set of training programs that are deemed to pose a risk, require warning, be downgraded, or be prohibited during the safety criteria verification process.

[0456] Based on the final evaluation results, the rule candidate set R is obtained. candidate A rule-driven candidate training item set T is generated through the rule-item mapping relationship.candidate The two are matched one-to-one. Subsequently, safety criteria are introduced to verify the safety of candidate training items, resulting in a set S of candidate prescriptions for safety warnings. candidate S candidate ⊆T candidate By using T candidate Remove S from the middle candidate The training projects in the dataset are used to obtain the final executable training project set D.

[0457]

[0458] Where m represents the final number of training prescriptions, d i The representative training prescription is composed of training item t. i and project rules r i composition.

[0459] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0460] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval, characterized in that, Includes the following steps: S1. Data Acquisition: This includes inertial motion data acquisition and electromyography (EMG) data acquisition. Macroscopic motion data and microscopic EMG data are collected from the subjects during the execution of preset assessment actions. The collected data are preprocessed to construct motor function characteristic data for subsequent analysis. S2. Layered Rehabilitation Assessment: Based on motor function characteristic data, the subject's motor function status is assessed in a layered manner, completing the overall motor function assessment, macroscopic abnormal function zoning and abnormal muscle groups and synergistic control abnormality identification tasks, and outputting structured layered motor function assessment results. S3. Multi-source assessment information fusion and embedding: Encode, embed, and modulate the patient's basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information to construct a unified multi-source assessment embedding representation to characterize the patient's current comprehensive motor function status. S4, Training Prescription Generation; Based on multi-source assessment embedding representation, and combining assessment-prescription rule base, training item knowledge base, and safety criterion base, an enhanced retrieval mechanism sequentially completes rule matching, training item matching, and safety verification tasks to generate an individualized rehabilitation training prescription that matches the patient's current motor function status; step 4 specifically includes the following steps: Step 4-1: Knowledge Base Establishment and Embedding; A multi-level knowledge base system related to rehabilitation training is pre-constructed, and the knowledge base content is vectorized and embedded to support subsequent enhanced retrieval and decision generation processes; the knowledge base system includes at least: an assessment-prescription rule base, a training program knowledge base, and a safety criterion base; Step 4-2: Candidate prescription generation; After completing the fusion and embedding of multi-source assessment information, the final assessment result matrix Z for rehabilitation decision-making is obtained. The matrix represents the comprehensive characteristics of the patient's current motor function status in terms of macroscopic functional areas, microscopic muscle groups, and historical training dimensions. Based on the final assessment result matrix Z, candidate rehabilitation training prescriptions are generated in the order of "rule priority - item refinement - safety constraints". Step 4-3: Generate the final prescription and output a structured prescription based on triple constraints of rules, items, and security; Rule candidate set: R candidate : Represents the set of evaluation-prescription rules that match the final evaluation result; rule-driven candidate training item set: T candidate The training item set obtained through rule-item secondary similarity retrieval; the candidate prescription set for security alerts: S candidate This refers to the set of training items that are deemed risky, require warnings, be downgraded, or prohibited during the safety criteria verification process. Based on the final evaluation results, the rule candidate set R is obtained. candidate A rule-driven candidate training item set T is generated through the rule-item mapping relationship. candidate The two are matched one by one; then, safety criteria are introduced to verify the safety of candidate training items, resulting in a set S of candidate prescriptions for safety warnings. candidate S candidate ⊆T candidate ; via T candidate Remove S from the middle candidate The training items in the dataset are used to obtain the final executable training item set D; Where m represents the final number of training prescriptions, d i The representative training prescription is composed of training item t. i and project rules r i composition.

2. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 1, characterized in that, Step 1, the configuration and data acquisition of the macroscopic motion data sensor, specifically includes the following steps: An inertial measurement unit (IMU) was used to collect macroscopic motion data of the subject during movement. One IMU sensor was used for each macroscopic anatomical position. The IMU included a triaxial accelerometer, a triaxial angular velocity meter, and a triaxial magnetometer to collect acceleration, angular velocity, and spatial orientation information of the limbs during movement. The inertial measurement unit (IMU) continuously acquires motion data at a fixed sampling frequency fs, obtaining 12 channels of macroscopic motion data signals, including: Triaxial acceleration signal a(t): Among them, a x (t) represents the acceleration in the x-axis direction, a y (t) represents the acceleration in the y-axis direction, a z (t) represents the acceleration in the z-axis direction; Triaxial angular velocity signal w(t): Among them, w x (t) represents the angular velocity along the x-axis, w y (t) represents the angular velocity along the y-axis, w z (t) represents the angular velocity along the z-axis; Triaxial magnetic field intensity signal m(t): Where, m x (t) represents the magnetic field strength along the x-axis, m y (t) represents the magnetic field strength along the y-axis, m z (t) represents the magnetic field strength along the z-axis; The collected raw motion data is processed by time synchronization, noise reduction filtering and coordinate system calibration, and the limb posture information is calculated based on the sensor fusion algorithm to obtain the posture sequence represented in quaternion or Euler angle form. Where, θ x (t) represents the rotation angle about the x-axis, θ y (t) represents the rotation angle about the y-axis, θ z (t) represents the rotation angle about the z-axis.

3. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 2, characterized in that, Step 2, the rehabilitation assessment of macroscopic movement data, specifically includes the following steps: Macroscopic motion feature construction and evaluation; based on attitude calculation results and angular velocity information, a set of kinematic features for macroscopic evaluation is constructed; for a 12-channel signal acquired by an inertial measurement sensor, the macroscopic motion features include: Activity range ROM∈R 12 : Where max represents the maximum value, min represents the minimum value, and ROM i ROM representing the signal of the i-th channel; average activity ∈R 12 : in, Represents the signal of the i-th channel ; Multi-joint synergy feature matrix C∈R 12×12 : Where corr represents the correlation coefficient, C ij The multi-joint coordination between the i-th and j-th signals is represented; x(t) represents the signal of a single channel acquired by the IMU, which is any one of the 12 channels of macroscopic motion data signals; x i (t), x j (t) represents two different signals from the 12-channel macroscopic motion data signal; Based on macroscopic movement characteristics, a macroscopic assessment of the overall motor function status of the subjects is performed; a macroscopic motor function score S is calculated. macro ∈R: Where: f k Represents the k-th macroscopic motion characteristic; ω k The corresponding feature weights reflect the contribution of each feature to overall motor function; N represents the number of features; the macroscopic motor function score reflects the overall motor ability level of the subject and serves as the basic input for subsequent stratified assessment; the above features together constitute the macroscopic motor feature set F. macro : Macroscopic anomaly localization: Based on macroscopic motor function assessment, further analysis is conducted on the distribution differences of macroscopic motor characteristics in different limb segments or functional zones to achieve macroscopic anomaly localization. Calculate the corresponding local macroscopic feature vector for each macroscopic anatomical region: And calculate its relationship with the set of health reference features. Distance between: The function dis(a1,a2) represents a vector formed by calculating the Euclidean distance between the elements in sets a1 and a2 in sequence; when the distance D corresponding to a certain functional partition... r When the threshold is exceeded, the functional area is determined to be a macroscopic motion abnormality area.

4. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 1, characterized in that, Step 1, the configuration and data acquisition of the microscopic electromyography data sensor, specifically includes the following steps: After completing the macroscopic motor function assessment and abnormality localization, microscopic electromyography (EMG) data is collected for the identified abnormal limb segments or functional zones to analyze the local neuromuscular functional status; the microscopic EMG data is collected through surface electromyography (sEMG) sensors. Surface electromyography (EMG) sensors are bipolar or multipolar surface electrodes, which are attached to the surface of key muscle groups within the target functional area and arranged along the direction of muscle fibers. The raw electromyographic signals acquired at the abnormal zonal locations are represented as follows: Where x i (t) represents the electromyographic signal of the i-th target muscle at time t, and M is the number of muscles collected; the original electromyographic signal is denoised, rectified and smoothed to obtain the electromyographic signal sequence used for feature extraction.

5. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 1, characterized in that, Step 2, the rehabilitation assessment of microscopic electromyography data, specifically includes the following steps: Microscopic electromyographic feature extraction and microscopic neuromuscular function assessment: Based on preprocessed electromyographic signals, microscopic electromyographic features that can reflect muscle activation level, fatigue state and synergistic control characteristics are extracted. Microscopic electromyographic characteristics include: Root mean square value RMS∈R M RMS i RMS representing the i-th lead: Mean Absolute Value (MAV) ∈ R M MAV i The average absolute value of the i-th lead: Muscle activation normalization index A∈R M A i The muscle activation normalization index representing the i-th lead normalizes the electromyographic amplitude relative to the maximum voluntary contraction MVC: Based on microscopic electromyographic feature vectors, a microscopic assessment of neuromuscular functional status within the target functional area is performed; a microscopic muscle function score S is constructed using a weighted method. micro ∈R: in, These are the weighting coefficients for the corresponding features; the weights are used to reflect the degree of contribution of different muscle groups and features to local motor function; the microscopic assessment results are used to quantify whether the muscle activation level is insufficient, or whether there is abnormal enhancement or fatigue trend. Localization of abnormal muscle groups and synergistic patterns; constructing a muscle synergistic matrix C by comparing the relative relationships between activation levels of different muscle groups. emg ∈R M×M Its elements Represented as: When the activation level of a certain muscle group is significantly lower or higher than that of other muscle groups in the same functional area, or when the correlation between it and the main synergistic muscle group is lower than a preset threshold, the muscle group is determined to be an abnormal muscle group. At the same time, the abnormal functional zoning information obtained in the macro-assessment phase is used to constrain the micro-assessment results; Through the process of acquiring and evaluating microscopic electromyographic data, a set of microscopic neuromuscular function evaluation results within the target functional area is output. The results include at least: a set of abnormal muscle groups; activation level indicators of each muscle group; and abnormal characteristics of synergistic control, denoted as F. micro .

6. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 1, characterized in that, Step 3, the fusion and embedding of multi-source evaluation information, specifically includes the following steps: Step 3-1: Word embedding of patient basic information, i.e., first information embedding; The first piece of information is the patient's basic information. After encoding the patient's basic information, it is embedded into a vector inf1 of length L: Further construct a trainable transformation matrix S1: And the first information matrix Inf1 is constructed using a weighted method: Step 3-2: Word embedding of macroscopic motor function information, i.e., second information embedding; The second information is macroscopic motor function assessment information, which originates from the macroscopic motor data acquisition and assessment positioning steps, including macroscopic motor characteristics of different limb segments or functional zones; the macroscopic motor function assessment information is embedded into a second information matrix Inf2 of size N×L1: Where N represents the number of macroscopic functional zones or macroscopic anatomical zones; each row represents the macroscopic motion feature embedding representation corresponding to a functional zone, that is, the set of macroscopic motion features. Each element in the array is expanded, concatenated, and then embedded, with a length of L1. Step 3-3: Word embedding of microscopic electromyographic functional information, i.e., third information embedding; The third piece of information is microscopic electromyography (EMG) function assessment information, which comes from the microscopic EMG data acquisition and assessment steps, including the activation level, synergistic relationship, and fatigue characteristics of the target muscle group; the microscopic EMG function assessment information is embedded into a third information matrix Inf3 of size M×L1: Where M represents the number of target muscle groups or microscopic functional units; L1 represents the dimension of the microscopic electromyographic feature embedding, i.e., the set of microscopic neuromuscular function assessment results F micro Each element is expanded and concatenated to form a vector of length L1; Steps 3-4: Word embedding of previously trained prescription information, i.e., fourth information embedding; The fourth information is the previous training prescription information, which includes the types of training exercises, training intensity, training frequency, and training effect feedback. This previous training prescription information is embedded into a fourth information matrix Inf4 of size K×L1. Where K represents the number of items in the prescription, and L1 is the length of the execution result embedding for each item; Steps 3-5: Multi-source assessment information fusion; modulated fusion of patient basic information and macroscopic motion information to obtain Inf5, and modulated fusion of patient basic information and microscopic electromyographic information to obtain Inf6, and outputting the fusion embedded results.

7. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 6, characterized in that, Steps 3-5 specifically include the following steps: Step 3-5-1: Modulation and fusion of patient basic information and macroscopic motion information to obtain Inf5; The patient's basic information embedding vector is: The macroscopic motion information matrix is ​​as follows: Step 3-5-1-1: Trainable mapping of basic information and generation of gated vectors; Introducing a trainable weight matrix W g1 ∈R L1×L1 With bias b g1 ∈R L1 Patient basic information is mapped to a gating vector g1: Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1]; Step 3-5-1-2: Trainable feature transformation of macroscopic information; Introducing the macroscopic feature transformation matrix W m1 ∈R L1×L1 With bias b m1 ∈R L1 By transforming the macroscopic information line by line, we obtain : Where 1∈R N×1 This is a vector of all 1s, used for broadcasting the bias term; Step 3-5-1-3: Gated modulation and residual fusion; The gate vector g1 is broadcast row-wise to N×L1 to achieve weighted modulation of macroscopic information by patient information. A residual term is introduced to improve stability, resulting in the fifth information matrix Inf5, which is a modulation and fusion of basic patient information and macroscopic motion information. Where ⊗ represents element-wise multiplication, and λ is a scalar coefficient that can be set or trained to control the residual fusion ratio; Step 3-5-2: Modulation and fusion of patient basic information and microscopic electromyographic information to obtain Inf6; The patient's basic information embedding vector is: The macroscopic motion information matrix is ​​as follows: Step 3-5-2-1: Trainable mapping of basic information and generation of gated vectors; Introducing a trainable weight matrix W g2 ∈R L1×L1 With bias b g2 ∈R L1 Patient basic information is mapped to a gating vector: Where σ(⋅) is the Sigmoid function, used to restrict the gating coefficient to [0,1]; Step 3-5-2-2: Trainable feature transformation of macroscopic information; Introducing the micro-feature transformation matrix W m2 ∈R L1×L1 With bias b m2 ∈R L1 The microscopic information is obtained by transforming it line by line. : Where 1∈R N×1 This is a vector of all 1s, used for broadcasting the bias term; Step 3-5-2-3: Gated modulation and residual fusion; The gating vector g2 is broadcast row-wise to N×L1 to achieve weighted modulation of microscopic information based on patient information. A residual term is introduced to improve stability. The sixth information matrix Inf6, which fuses the modulation of basic patient information and microscopic electromyographic information, is as follows: Where ⊗ represents element-wise multiplication, and λ is a scalar coefficient that can be set or trained to control the residual fusion ratio; Step 3-5-3: Output of Fusion Embedding Results; After completing the modulation and fusion of patient basic information and macroscopic motor function information, and the modulation and fusion of patient basic information and microscopic electromyographic function information, the fifth information matrix and the sixth information matrix are obtained, respectively: Meanwhile, the previously trained prescription information is used to obtain a fourth information matrix after word embedding: The multi-source information is directly concatenated to construct a unified multi-source fusion representation; the fifth, sixth, and fourth information matrices are then vertically concatenated along the row dimension to obtain the fused information matrix. Where vcon(⋅) represents the vertical concatenation operation, which is used to combine the embedded representations from different information sources into a unified matrix according to their functional unit order; A trainable mapping network is introduced to perform feature alignment and transformation on the fused information matrix; The mapping network is a fully connected neural network, a linear mapping, or a combination thereof, with an input dimension of L1 and an output dimension of L1. A unified embedding output matrix is ​​obtained by performing a row-by-row mapping transformation on the fused information matrix. Where Z represents a unified embedded representation that integrates patient basic information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information, and NN represents a fully connected network; the unified embedded output matrix Z serves as a structured representation of the patient's current comprehensive motor function status and is used in subsequent enhancement retrieval and individualized training prescription generation steps.

8. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 1, characterized in that, Step 4-1 specifically includes the following steps: Step 4-1-1: Evaluation—Establishment and embedding of the prescription rule base; The evaluation-prescription rule base is used to describe the mapping relationship between evaluation results and training items. Each evaluation-prescription rule includes at least the following structured field information: the mapping rule between evaluation embedded features and training items; the correspondence between macro-evaluation results and the specified training intensity range of training items; the correspondence between micro-evaluation results and the specified training intensity range of training items; and the rule for the j-th evaluation-prescription rule. j Simplified representation is r j The above fields are encoded and embedded respectively: Evaluation state constraint embedding: Encode the evaluation embedding feature constraints involved in the rule into a vector r. e (j) And mapped to an embedding vector : Macro-evaluation—Intensity Mapping Embedding: Encodes the mapping relationship between macro-evaluation results and training intensity intervals into a vector r. m (j) And mapped to an embedding vector : Microscopic Assessment-Intensity Mapping Embedding: Encodes the mapping relationship between microscopic electromyography assessment results and training intensity ranges into a vector r. u (j) And mapped to an embedding vector : Subsequently, the three types of rule embedding vectors are concatenated, and dimension alignment is performed using a trainable mapping network to obtain the rule embedding vector. : Where hcon represents the horizontal splicing operation, W r W e W m W u b is a trainable weight matrix e b m b r b u For trainable bias, Φ represents a nonlinear mapping function; All rule embedding vectors constitute the evaluation-prescription rule base embedding matrix: Where R represents the number of assessment-prescription rule entries; Step 4-1-2: Establishment and embedding of the training project knowledge base; The training program knowledge base stores executable rehabilitation training programs, with the basic unit being a training program entry. Each training program entry includes at least the following attribute information: training program name; target functional area or target muscle group; key points of movement execution. For the i-th training program entry, the above attribute information is embedded as follows: The training item names are categorically encoded into action embedding vectors. : The target functional zones or target muscle groups are encoded into zone / muscle group embedding vectors. : It structures and encodes the key points of action execution into r c (i) And mapped to key embedding vectors : The three types of embedding vectors are concatenated and projected onto a unified dimensional space through a trainable mapping network to obtain the training item embedding vectors. : Where hcon represents the horizontal splicing operation, W t W c b is a trainable weight matrix t b c For trainable bias, Φ represents a nonlinear mapping function; The embedding vectors of all training items constitute the embedding matrix of the training item knowledge base: Where T represents the number of training items; Step 4-1-3: Establishment and embedding of the security rules library; Each safety criterion is one of the following: the maximum permissible training intensity corresponding to a certain stage of disease progression; prohibited movements under abnormal conditions of specific muscle groups; individual safety threshold limits based on the patient's basic information; contraindications for rehabilitation training of a certain complication; the k-th safety criterion is parameterized and encoded to obtain the safety constraint vector s. k And mapped to a security criterion embedding vector through an embedding mapping function: Among them, W s b is a trainable weight matrix s For trainable bias, Φ represents a nonlinear mapping function; All security criterion embedding vectors constitute the security criterion library embedding matrix: Where S represents the number of safety criteria.

9. The method for generating individualized rehabilitation intervention prescriptions that integrates hierarchical localization rehabilitation assessment and multi-stage enhanced retrieval according to claim 8, characterized in that, Step 4-2 specifically includes the following steps: Step 4-2-1: Rule base retrieval based on the final evaluation result matrix; A unified embedded representation Z that integrates basic patient information, macroscopic motor function assessment information, microscopic electromyographic function assessment information, and previous training prescription information: Among them, different rows correspond to historical prescription information, macroscopic functional zoning assessment information, and microscopic muscle group assessment information, respectively. Perform dimension alignment and query vector construction on the final evaluation result matrix; Rule query vector construction; mapping the final evaluation result matrix Z to the rule embedding space using a trainable mapping function: Among them, W q ∈R L1×L2 b is a trainable weight matrix q ∈R L2 For bias, 1∈R (K+N+M)×1 Let q be an all-1 vector, and Φ(⋅) be a non-linear mapping function; construct a rule-based retrieval query vector, where the i-th row has 1 ≤ i ≤ M + N + K, and the query vector is q. i : Based on rule-based retrieval query vectors, and embedded vectors of the evaluation-prescription rule base. If 1 ≤ j ≤ R, the similarity matrix is ​​calculated to obtain the similarity S. ij rule : Where sim is the similarity function; to obtain the overall matching degree of the j-th rule (1≤j≤R) in the current evaluation state, the similarity of each rule is aggregated to obtain the rule activation degree s. j rule : Where, α i Weighting coefficients related to the degree of anomaly assessment are used to reflect the importance of different row assessment units to rule matching; scores s are selected. j rule Rule entries that exceed a preset threshold or rank highly (the top k entries, k ≤ R) form a candidate subset R. candidate： Based on the subset of candidate rules, determine the applicable training item categories, as well as the corresponding training frequency range and training intensity range, under the current evaluation state; Step 4-2-2: Knowledge base retrieval of training items based on rule retrieval results; obtaining candidate rule subset R. candidate Then, a rule-based secondary similarity retrieval mechanism for projects is introduced; Candidate rules are obtained by corresponding rule embedding vectors; for a subset R of candidate rules candidate Each assessment-prescription rule in the document j For each rule, 1 ≤ j ≤ k, obtain its corresponding rule embedding vector: The rule embedding vector is used to characterize the applicable conditions and training parameter constraints of the training item described by the rule; The embedding representation of the training project knowledge base; the embedding matrix of the training project knowledge base is: Rule-based secondary similarity calculation for each item; embedding vector for each candidate rule. Calculate the embedding vector for each candidate rule. The similarity S between the embedding vectors of each training item in the training item knowledge base and the training item knowledge base jt train : Where sim(⋅) is the similarity function; the training items corresponding to the rule are determined; for the j-th candidate rule, based on the rule-item similarity calculation result, the training item with the highest similarity score is selected as the training item that best matches the rule: Perform the above operations on all k candidate rules, thereby establishing a one-to-one correspondence between candidate rules and specific training items, and obtaining a rule-driven candidate training item set T. candidate : Step 4-2-3: Retrieve safety guidelines for integrating evaluation results and training project information; After obtaining the set of candidate training items, a security criterion library is introduced to perform security verification on the candidate training items; for each candidate training item t i Embed it into vector e train (i) With the final evaluation result matrix after embedding By merging the data, a joint query representation for security verification is constructed: Based on joint query representation The similarity between the vector and the embedding vector of the security criteria library is calculated to obtain S. ip safe : When the activation level of a safety criterion exceeds a preset risk threshold, the project is identified as a warning project; after safety screening, a set S of candidate prescriptions for safety warnings is obtained. candidate , where n is the number of warning items.

Images