Method, device, system, medium and equipment for generating exercise prescription based on clinical data

Abstract

This application belongs to the field of information processing and artificial intelligence technology, and discloses a method, device, system, medium, and equipment for generating exercise prescriptions based on clinical data. The method includes: acquiring clinical data of test subjects; extracting joint physiological features from the clinical data based on a multi-source heterogeneous data fusion model; inputting the extracted joint physiological features into a dynamic randomization engine, and grouping the joint physiological features based on the dynamic randomization engine; constructing and training a health status prediction model; inputting the grouped joint physiological features into the trained health status prediction model; outputting the health status level of the test subject and the physiological status evolution trajectory under different exercise interventions; and generating a personalized exercise prescription. This application can achieve accurate prediction of individual health status and personalized generation of intelligent exercise prescriptions, effectively improving the scientific, targeted, and practical nature of health management for test subjects.

Description

Technical Field

[0001] This application belongs to the field of information processing and artificial intelligence technology, specifically relating to a method, device, system, medium and equipment for generating exercise prescriptions based on clinical data. Background Technology

[0002] Exercise prescriptions are widely used in rehabilitation training and health promotion, but existing systems are often based on static rules or expert experience, lacking dynamic feedback and personalized adjustments to individual clinical data, and the tracking methods are relatively simple and cannot dynamically reflect the evolution of the effects of exercise intervention.

[0003] Currently, although there are explorations of using some machine learning algorithms (such as clustering and SVM) for exercise prescription optimization, the results are difficult to interpret and there is a lack of a "visual-decision linkage" mechanism for medical scenarios.

[0004] Therefore, there is an urgent need for a comprehensive approach that integrates "dynamic simulation algorithms, data tracking, and individual characteristic modeling" to address the adaptation problem between individual differences and clinical practice. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the main objective of this application is to provide a method, device, system, medium, and equipment for generating exercise prescriptions based on clinical data. This application enables interpretable modeling and dynamic visual presentation of the health status and exercise response of test subjects, providing test subjects with more accurate exercise prescription formulation and intervention assessment tools.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A method for generating exercise prescriptions based on clinical data, the method comprising: acquiring clinical data of test subjects, the clinical data including structured data, text data, image data, and audio data; extracting joint physiological features from the clinical data based on a multi-source heterogeneous data fusion model; inputting the extracted joint physiological features into a dynamic randomization engine, and grouping the joint physiological features based on the dynamic randomization engine; constructing and training a health status prediction model, inputting the grouped joint physiological features into the trained health status prediction model, outputting the health status level of the test subjects and the physiological status evolution trajectory under different exercise interventions, and generating personalized exercise prescriptions.

[0008] Optionally, the multi-source heterogeneous data fusion model includes: a structured data encoder, a text data extractor, and an image and audio data encoder, wherein the structured data encoder is used to extract numerical indicators and categorical variable features from the structured data; the text data extractor is used to extract latent semantic features from the text data; and the image and audio data encoder is used to extract key spatiotemporal structure and physiological representation information from the image and audio data.

[0009] Optionally, the multi-source heterogeneous data fusion model further includes a decoupling fusion module, used to realize the interaction and fusion between structured data, text data, image data and audio data, so as to extract joint physiological features across modalities.

[0010] Optionally, the dynamic randomization engine includes: a feature input interface for receiving joint physiological features; a dimensional modeling module for grouping the joint physiological features into static and dynamic dimensional factors; a balance assessment module for assigning corresponding clinical importance weights to the static and dynamic dimensional factors and calculating a complex balance index under different grouping schemes; a minimization allocation module for determining the optimal grouping scheme for the test subjects based on the complex balance index; and an adjustment module for dynamically adjusting the clinical importance weights of each dimensional factor in the optimal grouping scheme when an imbalance between groups is detected, and updating the grouping strategy based on the adjustment results.

[0011] Optionally, the health status prediction model includes: an input layer, an intervention encoding layer, a temporal modeling layer, a health status mapping layer, a state decoding layer, and an exercise prescription generator, arranged sequentially. The input layer is used to input the joint physiological characteristics of the test subject at the current moment and the current exercise intervention vector extracted by the multi-source heterogeneous data fusion model. The intervention encoding layer is used to convert the current exercise intervention vector into a representation with the same dimension as the joint physiological characteristics of the test subject at the current moment. The temporal modeling layer is used to fuse the joint physiological characteristics of the test subject at the current moment and the current exercise intervention vector to obtain a health latent state vector. The health status mapping layer is used to map the health latent state vector into a health status level probability vector. The state decoding layer is used to predict the physiological state of the test subject at the next moment based on the health latent state vector. The exercise prescription generator is used to generate an exercise prescription based on the health status level probability vector and the physiological state of the test subject at the next moment.

[0012] Optionally, the health status prediction model is trained through the following steps: acquiring the health status levels and corresponding exercise prescription records of multiple test subjects, constructing a dataset containing physiological state sequences, intervention information, future states, and health status labels, and dividing the dataset into a training set and a validation set; initializing training parameters, training the model using the training set until the maximum number of iterations is reached; validating the model using the validation set. During the validation process, if the root mean square error (RMSE) of state prediction, the accuracy of health status level prediction, and the F1 score of health status level prediction all meet the thresholds, the model is validated successfully; otherwise, the training parameters are modified and the model is retrained until the model is validated successfully.

[0013] This application also provides a clinical data-based exercise prescription generation device, the device comprising: an acquisition module for acquiring clinical data of test subjects, the clinical data including structured data, text data, image data, and audio data; a feature extraction module for extracting joint physiological features from the clinical data based on a multi-source heterogeneous data fusion model; a grouping module for inputting the extracted joint physiological features into a dynamic randomization engine and grouping the joint physiological features based on the dynamic randomization engine; and an exercise prescription generation module for constructing and training a health status prediction model, inputting the grouped joint physiological features into the trained health status prediction model, outputting the health status level of the test subjects and the physiological status evolution trajectory under different exercise interventions, and generating a personalized exercise prescription.

[0014] This application also provides a clinical data-based exercise prescription generation system, the system including the apparatus described above.

[0015] This application also provides a storage medium including instructions that, when executed on a computer, cause the computer to perform the method as described in the preceding claim.

[0016] This application also provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any of the preceding claims.

[0017] Compared with the prior art, this application can bring the following technical effects:

[0018] This application constructs a health status prediction model that integrates multimodal information such as structured data, text, images, and audio. Combined with a time-series modeling and risk mapping mechanism based on dynamic intervention perception, it achieves high-precision prediction of individual physiological state changes and health status levels. Furthermore, leveraging an intelligent exercise prescription generator, it recommends optimal intervention combinations based on health evolution trajectories and optimization goals, effectively improving the scientific rigor, adaptability, and operability of individualized intervention plans. The overall system possesses high automation, real-time performance, and clinical interpretability, providing strong technical support and decision-making basis for precision medicine and personalized health management. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for generating exercise prescriptions based on clinical data, provided in one embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the structure of a multi-source heterogeneous data fusion model provided in another embodiment of this application;

[0021] Figure 3 This is a schematic diagram of the structure of a health status prediction model provided in another embodiment of this application;

[0022] Figure 4 This is a schematic diagram of a clinical data-based exercise prescription generation device provided in another embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0025] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0026] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0027] Figure 1 This application provides an exemplary embodiment of a method for generating exercise prescriptions based on clinical data, such as... Figure 1 As shown, the method includes the following steps:

[0028] S1: Obtain clinical data of the test subjects, including, for example, structured data (e.g., blood pressure, blood sugar, BMI), text data (e.g., medical records, questionnaires), image data (e.g., MRI, motion videos) and audio data (e.g., speech, breathing sounds).

[0029] S2: Based on a multi-source heterogeneous data fusion model, joint physiological characteristics are extracted from the clinical data;

[0030] S3: Input the extracted joint physiological features into the dynamic randomization engine to group the joint physiological features;

[0031] S4: Construct and train a health status prediction model. Input the grouped joint physiological features into the trained health status prediction model, output the health status level of the test subjects and the physiological status evolution trajectory under different exercise interventions, and generate personalized exercise prescriptions.

[0032] In another exemplary embodiment, in step S2, as Figure 2As shown, the multi-source heterogeneous data fusion model includes: a structured data encoder, a text data extractor, an image and audio data encoder, and a decoupling fusion module.

[0033] In this embodiment, the structured data encoder is used to extract numerical indicators and categorical variable features from structured data. The structured data encoder includes an input layer, a categorical embedding layer, a numerical feature normalization layer, a multi-head self-attention mechanism, a feedforward layer, and an output layer. The input layer is used to input structured data and divide it into categorical fields and numerical fields. The categorical embedding layer is used to map each categorical field (such as gender, medical history, marital status) to a fixed-length embedding vector through a learnable embedding lookup table. The embedding vectors obtained from mapping all categorical fields are finally concatenated to form a categorical embedding matrix.

[0034] For example, firstly, this application defines three categorical fields as shown in Table 1, where each field is embedded as a 4-dimensional vector:

[0035] Table 1. Categorical Fields

[0036]

[0037] Next, an embedding lookup table for a 4-dimensional embedding vector is constructed for each category value, as shown in Tables 2 to 4:

[0038] Table 2 Gender Embedding Table

[0039]

[0040] Table 3. Embedded table of medical history

[0041]

[0042] Table 4 Embedded Table of Marital Status

[0043]

[0044] Furthermore, by looking up an individual's information in a table, we can obtain: gender = female; disease = diabetes.

[0045] Marital status = married; then its corresponding embedding vectors are: gender vector: [0.0,0.1,-0.1,0.2]; medical history vector: [-0.1,0.3,0.0,0.1]; marital status vector: [0.1,-0.1,0.2,0.2].

[0046] Finally, concatenate the vectors of the three fields row by row to obtain the embedding matrix shown below:

[0047]

[0048] In this embedding matrix, the first row corresponds to the embedding vector for "gender = female"; the second row corresponds to the embedding vector for "medical history = diabetes"; and the third row corresponds to the embedding vector for "marital status = married".

[0049] The numerical feature normalization layer is used to compress all numerical fields (such as blood pressure and BMI) to [0,1] through standard normalization (Z-score or Min-Max normalization). Subsequently, the embedding vectors of all categorical fields are concatenated with the normalized numerical fields in terms of dimensions to form a unified feature representation tensor [number of fields, embedding dimension].

[0050] For example, this individual has a systolic blood pressure of 150 mmHg and a BMI of 31.0. The systolic blood pressure is normalized to 0.667 and the BMI is normalized to 0.64. To facilitate splicing, these scalars are converted into vectors. Here, each numerical field is also extended to 4 dimensions (aligned with the embedded vector). The systolic blood pressure vector is represented as: [0.667, 0.667, 0.667, 0.667]; the BMI vector is represented as: [0.64, 0.64, 0.64, 0.64].

[0051] Finally, the categorical and numerical fields of this individual are concatenated into a vector, as shown in Table 5:

[0052] Table 5 Vector Table

[0053]

[0054] This ultimately yields a unified feature representation tensor:

[0055]

[0056] After obtaining the unified feature tensor, it is input into a multi-head self-attention mechanism. In this mechanism, each row vector in the unified feature tensor is treated as a sequence position, i.e., a semantic representation of a structured field. Each position can be mapped to three sets of vectors through a learnable weight matrix: a query vector, a key vector, and a value vector. Subsequently, for each set of fields, its attention score is calculated.

[0057]

[0058] in, express The query vector of the field. Representation field The key vector,

[0059] All scores were normalized using the Softmax normalization function to obtain the field. Attention weights for all fields This weight reflects the field The degree of attention paid to other field information can be interpreted as the semantic dependencies between fields learned by the model.

[0060] Furthermore, based on attention weights Calculated fields The final representation vector is a vector of all field values. Weighted summation:

[0061]

[0062] The above process is executed independently at each attention head, generating output representations in multiple subspaces, enabling the capture of interaction patterns between fields from different perspectives. The outputs of all attention heads are concatenated dimensionally and then projected back to the original embedding dimension through a linear transformation, resulting in the final multi-head attention output tensor Z. This tensor preserves the complex, non-linear, and high-dimensional interaction features between structured fields, making it particularly suitable for modeling the potential impact of multiple field combinations on health status. For example, the model can automatically identify hidden feature structures such as "improved health status when high BMI and diabetes co-occur" or "hypertension combined with a family history suggests a stronger need for intervention," without the need for explicit design of cross-features.

[0063] After the unified feature tensor is modeled using a multi-head self-attention mechanism to obtain enhanced field interaction representations, this application further introduces residual connections and layer normalization structures to improve the training stability and generalization ability of the model. Specifically, the output tensor Z of the multi-head attention mechanism and its input tensor (i.e., the unified feature tensor) form a residual connection through element-wise addition. This operation helps to preserve the original field information and can prevent the attention mechanism from overfitting or losing original features in certain training stages. Subsequently, layer normalization is performed on the residually connected tensor (this operation performs mean and variance normalization within each row (i.e., each field vector), standardizing the numerical distribution of each field in different dimensions) to obtain the attention-enhanced representation. This can eliminate numerical instability in the gradient propagation process and improve the convergence speed and quality of the model training process.

[0064] Normalized representation of attention enhancement The input feedforward layer performs nonlinear transformations and compression of the feature space to obtain the integrated intermediate feature tensor. The feedforward layer employs a two-layer structure: the first layer consists of a linear mapping and a ReLU activation function; the second layer consists of a linear mapping and Dropout regularization. The specific computation process is as follows:

[0065]

[0066] in, This represents the output tensor of the feedforward layer; This represents the weight matrix of the first layer; This represents the weight matrix of the second layer; This represents the first-level bias vector; This represents the second-layer bias vector.

[0067] While preserving dimension, the feedforward layer can further extract potential nonlinear combinations of fields, such as capturing health status expressions under multiple conditions such as "high BMI and young age" and "female and diabetic". At this point, the output tensor of the feedforward layer... This represents the semantic enhancement vector for each structured field.

[0068] In summary, the feedforward layer adopts a two-layer structure, consisting of two linear transformation layers with a ReLU activation function and Dropout regularization operation inserted in between. Its main function is to introduce nonlinear transformation capabilities and enhance the model's expressive richness while maintaining consistent output dimensions. The first layer maps the input vector from its original dimension to a higher dimension (e.g., a 4-fold expansion), using the ReLU activation function to enhance feature sparsity and nonlinear modeling capabilities. The second layer then compresses it back to its original dimension, using Dropout to suppress overfitting and improve generalization performance. This structure, through a "first expand, then compress" feature space reconstruction mechanism, enables the model to more effectively extract latent semantic relationships and health status features from complex field combinations, enhancing its adaptability to heterogeneous clinical data and significantly improving the overall network's nonlinear expressiveness and training stability.

[0069] Finally, to achieve effective alignment and fusion of structured data with other modal features (such as text, images, and audio), this application, after completing attention enhancement and feedforward nonlinear mapping, introduces an average pooling layer and performs an average pooling operation on the output tensor. Dimensionality reduction and compression are performed by averaging all field vectors along the same dimension to generate a global structured representation vector of uniform length *d*. This representation vector is further mapped to a predetermined dimension (e.g., 64-dimensional) through an output linear transformation layer, forming the final structured key feature vector. This vector not only highly compresses the information redundancy of the original structured fields but also retains the nonlinear interactions between fields and individual-specific health status features modeled in the attention mechanism. It possesses good expressive power and discriminative performance, comprehensively reflecting clinical semantic relationships, multidimensional feature coupling mechanisms, and individualized health status information, providing a unified and standardized structured input foundation for subsequent multimodal fusion processing and health status modeling.

[0070] The described text data extractor is used to automatically extract potential semantic features with clinical value from unstructured medical text data to assist in multimodal health status modeling and exercise prescription generation. This module is applicable to processing various types of text data sources, including but not limited to electronic medical records (such as diagnostic records, medication history), lifestyle questionnaires (such as diet, daily routine, exercise habits), and patient self-reporting forms, etc. Its structure consists of a text embedding layer, a bidirectional gated recurrent unit encoding layer (BiGRU), an attention mechanism, a feature compression mapping layer, and an output layer in sequence, having strong semantic modeling ability and feature condensation ability.

[0071] First, the text embedding layer is responsible for preprocessing and vectorizing the original text data. The original text data undergoes a standardization process, including removing stop words (words with high frequency but little semantic contribution during text analysis, such as "了", "的"), filtering punctuation and non-medical noise characters, and performing term standardization, abbreviation expansion, and key entity recognition through a medical domain-specific dictionary (such as SNOMED CT, ICD-10, or a self-built entity word library). This step can unify diverse natural language inputs into a standard term system and improve the model's recognition ability of domain knowledge. Subsequently, the text is tokenized, and each word is mapped to a fixed-length vector using a pre-trained word vector model (such as medical Word2Vec, ClinicalBERT, or BioBERT), obtaining a word embedding matrix with a shape of where, represents the length of the word sequence, represents the embedding dimension of each word (such as 100 or 300 dimensions). This embedding matrix is then input into the BiGRU encoding layer. The BiGRU encoding layer consists of two gated recurrent units (GRUs) with opposite directions, traversing and modeling the sequence in the forward (from front to back) and backward (from back to front) directions respectively, and can capture long-term dependencies and context structures in the text. The context-enhanced semantic tensor output by the BiGRU encoding layer has a shape of where, represents the time step, that is, the length of the input text, represents the hidden state dimension of a single direction. This tensor fully integrates the semantic roles and contextual relationships of each word in its context.

[0072] To further enhance the model's ability to focus on key information, the text data extractor introduces a semantic aggregation module based on an attention mechanism. This mechanism calculates the attention score between the context vector output at each time step and the overall semantic center, constructing a normalized weight distribution using Scaled Dot-Product or trainable attention parameters. The attention mechanism then performs a weighted summation of all outputs from the BiGRU encoding layer based on these weights, thereby generating a global semantic representation vector of the text. This process significantly improves the model's ability to perceive and respond to key health state terms or clinical events (such as "recent syncope," "history of hyperglycemia," and "family history of heart disease") in medical texts, demonstrating stronger diagnostic assistance potential.

[0073] Next, the global semantic vector is fed into a feedforward compression mapping network to further compress the feature dimensions and perform nonlinear transformations. This feedforward compression mapping network includes a linear mapping layer, a ReLU activation function, and Dropout regularization, which enhances the model's nonlinear expressive power, suppresses overfitting, and improves generalization ability. The feedforward compression mapping network ultimately outputs a fixed-length low-dimensional text feature vector (e.g., 64-dimensional), which condenses the semantic focus, domain knowledge, and potential health state signals in the original unstructured text. Finally, this 64-dimensional text feature vector is used as a standardized structural input, fed into a multimodal fusion module along with other modalities (e.g., structured data, image features, audio representations) to achieve joint modeling within a unified semantic space.

[0074] In summary, the text data extractor successfully achieved the efficient transformation of unstructured medical text into structured, semantically enhanced features through the processing chain of "embedding-context modeling-attention aggregation-nonlinear compression," providing the model with a text modal input that is highly expressive, alignable, and medically interpretable.

[0075] The image and audio data encoder is used to extract key spatiotemporal structure and physiological representation information from image and audio data. Specifically, the encoder includes an input transformation layer, a convolutional feature extraction layer, a global pooling layer, a temporal compression layer, a feature compression mapping layer, and an output layer. First, the image and audio data are standardized through the input transformation layer. For example, for image data (such as static images or video frame sequences), the input image is cropped or interpolated to a uniform size (e.g., 224×224 pixels), and pixel values ​​are normalized to [0,1]. For audio data (such as speech or breath sounds), a short-time Fourier transform (STFT) combined with a Mel filter bank is used to convert the one-dimensional audio signal into a two-dimensional Mel spectrogram. This spectrogram preserves the temporal structure and spectral characteristics of the audio signal and has a two-dimensional tensor format similar to that of an image, facilitating the subsequent extraction of local and global features by the convolutional neural network. Subsequently, in the convolutional feature extraction layer, the standardized image data and the 2D Mel spectrogram are input into two consecutive lightweight convolutional modules. Each lightweight convolutional module contains a 3×3 convolutional layer, a batch normalization layer (BatchNorm), a ReLU activation function, and a max pooling layer. First, the standardized image data and the 2D Mel spectrogram pass through the 3×3 convolutional layer, which extracts local spatial features or time-frequency spectral patterns through convolution operations, such as lesion edges and tissue morphology in the image, or speech frequency features in the Mel spectrogram. Next, the batch normalization layer (BatchNorm) is used to standardize the output after convolution, thereby accelerating the training process and improving the stability of the model. Then, the ReLU activation function is applied to a non-linear transformation to increase the expressive power of the model and avoid negative outputs. Finally, the data passes through a max pooling layer, which reduces the spatial dimension of the data by selecting the maximum value in the local region while retaining the most important features, thereby generating a high-dimensional feature map. Next, the high-dimensional feature map obtained after processing by the lightweight convolution module enters the global pooling layer. This layer compresses the two-dimensional features into a one-dimensional channel vector by performing global average pooling on the spatial dimension (image) or spectral dimension (audio). This vector aggregates the global semantic representation of the entire modality, which can significantly reduce the feature dimensionality and retain the key modal expressions. Subsequently, this one-dimensional channel vector enters the temporal compression layer, where temporal modeling is performed through one-dimensional convolution to extract the dynamic evolution features in the one-dimensional channel vector. These dynamic evolution features include, for example, features that can effectively characterize the continuity of speech pronunciation (such as changes in pitch, speed, etc. during the start, duration, and end of pronunciation), changes in respiratory rhythm (such as the trend of changes in respiratory frequency and depth over time), or temporal patterns in motion trajectories (such as changes in stride frequency and speed of an athlete running or walking).Temporal modeling enhances the model's ability to perceive dynamic signals such as behavioral flows (in human motion trajectory analysis, behavioral flows refer to the model of an individual walking, running, or pausing; in speech recognition, behavioral flows refer to the fluctuations and rhythms of a speaker's pronunciation) and pathological rhythms (referring to the physiological periodic changes or abnormal rhythms of the human body or organism under certain pathological states; for example, heart disease patients may exhibit abnormal heart rhythms (such as arrhythmia), or the respiratory rate, depth, and regularity of patients with respiratory diseases may change). Next, to ensure consistency among multimodal features and the discriminative power of subsequent fusion, the temporally compressed feature vectors are fed into a feature compression mapping layer. This layer consists of a fully connected layer, combined with ReLU activation and Dropout regularization, to achieve nonlinear compression and dimensionality unification of the feature vectors. Through this layer, image and audio features are compressed into fixed-length low-dimensional vectors (e.g., 64-dimensional), which not only helps reduce redundant information but also improves the expressive power of cross-modal alignment. Finally, in the output layer, this low-dimensional vector is output as a high-order semantic representation vector of the image and audio modalities. This vector fully integrates the spatial structure, temporal rhythm and semantic content of the input modalities, carrying important information about individuals in morphology, physiological state, language behavior, and action rhythm. It has high discriminativeness and fusion adaptability, and can serve as an important component input in the multimodal health modeling system.

[0076] In summary, the image and audio feature encoder can perform deep modeling and feature extraction on heterogeneous modal data from the perception layer, and output structured, aligned, and highly expressive semantic vectors in a unified manner, providing high-quality modal input support for tasks such as individual health status modeling, behavior recognition, and disease evolution tracking.

[0077] The decoupled fusion module is used to achieve high-quality interaction and fusion between structured data, text data, image data, and audio data, effectively extracting joint physiological features across modalities while maintaining modal diversity and semantic independence. The decoupled fusion module includes a modality compression layer, a modality orthogonal projection layer, and a semantic co-occurrence attention layer. First, in the intramodal compression layer, feature vectors from each modality encoder (e.g., 64-dimensional structured data, 64-dimensional text features, 64-dimensional image features, and 64-dimensional audio features) are input into independent linear dimensionality reduction modules. Each linear dimensionality reduction module consists of a fully connected layer and an activation function (e.g., ReLU), aiming to compress the original features of each modality into low dimensions and map them uniformly to a common representation space, while simultaneously improving semantic representation capabilities through nonlinear activation. For example, if each modality is initially 64-dimensional, it can be compressed to 32 dimensions, forming a [M×32] fusion starting tensor (M is the number of modalities, typically 4). The purpose of this stage is to remove redundant information within a modality through intramodal dimensionality reduction compression, enhance feature compactness, and unify feature dimensions for subsequent fusion processing.

[0078] Next, in the modal orthogonal projection layer, the compressed feature vector for each modality... This will be achieved through a set of learnable projection matrices. Projecting onto mutually orthogonal subspaces yields the modalities. orthogonal vectors : To ensure that the orthogonality constraints are maintained between the projection matrices, the following orthogonality loss term is added to the loss function during model training:

[0079]

[0080] in, Indicates orthogonality loss; Representing modes The projection matrix, Representing modes The projection matrix; Denotes the square of the Frobenius norm, and represents and The degree of similarity between them.

[0081] This orthogonal regularization term forces the model to keep the representation directions of different modalities in the semantic space as uncorrelated as possible during the learning process, thereby preventing semantic interference or feature coverage during modality fusion.

[0082] The orthogonalized modal representations are input into a semantic co-occurrence attention layer, which is responsible for realizing the cross-interaction and joint representation of information between modalities. First, all projected modal features are assembled into a [M,D] tensor, where each row represents an orthogonal semantic vector of a modality. Then, a semantic co-occurrence attention matrix A ∈ [M,D] is constructed. ^{M×M}, where each element Representing modes For modes The semantic attention level is determined. Attention weights can be calculated using the ScaledDot-Product mechanism. Subsequently, a cross-modal weighted aggregation operation is performed on the semantic vectors of each modality to obtain the modal interaction enhancement vector. :

[0083]

[0084] in, Representing modes For modes The degree of semantic attention; Indicates the number of modes; Representing modes Orthogonal vectors.

[0085] Furthermore, all interaction enhancement vectors are then merged into a final fused vector through concatenation or average pooling. :

[0086]

[0087] This fusion vector embodies the cross-information features of multimodal semantics after semantic resonance, and has higher discriminative power and modeling value.

[0088] Finally, the fusion vector Input into a health status prediction model to infer the physiological evolution trajectory, health status level, or recommended exercise type and intensity of test subjects under different intervention conditions.

[0089] In summary, the innovation of the decoupled fusion module lies in its structured processing mechanism of "compression-orthogonality-reconstruction," which breaks through the problems of information redundancy, semantic interference, and modal bias faced by traditional modal splicing or simple attention mechanisms in the fusion process, and realizes precise control and system optimization from modal representation to cross-modal semantic fusion.

[0090] First, in the modality compression layer, the decoupled fusion module does not adopt the common "uniform linear projection" approach. Instead, it designs an independent dimensionality reduction path for each modality, preserving its unique expressive structure and domain characteristics. While ensuring compression effectiveness, it enhances the resolution of modal semantic expression through a nonlinear activation function, enabling it to maintain strong expressive power in low-dimensional space. This design avoids the learning difficulties and overfitting problems caused by dimensionality explosion in traditional splicing and fusion, creating a balanced foundation for inter-modal alignment.

[0091] Secondly, a "learnable orthogonality constraint" is introduced into the modal orthogonal projection layer. By applying an independent linear projection to each modal compression vector and incorporating the goal of maintaining approximate orthogonality between the projection matrices into the loss function during training, this module can actively regulate the degree of decoupling of different modal representations in the semantic space. This "structural separation" ensures that each modality retains its own domain features in the shared representation space without interference, overlap, or semantic drift, effectively solving the coupling pollution problem of multimodal features in the representation dimension. This not only improves the robustness of fusion but also provides structural guarantees for the interpretability of downstream models.

[0092] Finally, in the semantic co-occurrence attention layer, this layer breaks through the traditional coarse-grained design of "averaging" all modalities and adopts a dynamic semantic weight modeling mechanism between modalities. By constructing an attention weight matrix between modalities, the system can adaptively capture the semantic complementarity between modalities in different sample contexts. For example, when the text modality clearly describes the medical history but structured data is missing, the attention mechanism will automatically increase the weight of the text modality, and vice versa. This fine-grained, multi-sample adaptive co-occurrence modeling mechanism enables modality fusion to shift from "structural superposition" to "semantic synergy," greatly enhancing the model's ability to express heterogeneous clinical data and individual differences.

[0093] From an overall technical perspective, the decoupled fusion module not only constructs a hierarchical, directional, and controllable fusion path, but also achieves a leap from sparse intramodal representation to deep semantic linkage between modalities through a rigorous process of compression → orthogonality → co-occurrence. This design significantly improves the accuracy, generalization, and interpretability of the multimodal health modeling system, enabling the model to maintain stable fusion capabilities even when facing complex clinical data (such as missing, redundant, and semantically contradictory data), and to output predictive evidence with clear semantic sources, meeting the practical needs of "data transparency, traceable paths, and reliable decisions" in smart healthcare.

[0094] In summary, the decoupled fusion module not only has a clear technical division of labor and coupling control strategy in its structural design, but also integrates key technologies in deep learning such as embedding representation, orthogonal constraints and attention mechanisms in its methodological path, forming a novel combination method. It has significant technological progress and application creativity, and represents an important technological breakthrough in the field of multimodal fusion modeling.

[0095] In another exemplary embodiment, in step S3, the dynamic randomization engine includes: a feature input interface for receiving joint physiological features; a dimensional modeling module for grouping the joint physiological features to form static dimensional factors and dynamic dimensional factors; a balance assessment module for setting corresponding clinical importance weights for static and dynamic dimensional factors and calculating complex balance indices under different grouping schemes; a minimization allocation module for determining the optimal grouping scheme for test subjects based on the complex balance index and using the minimization principle; and an adjustment module for dynamically adjusting the clinical importance weights of each dimensional factor in the optimal grouping scheme when an imbalance between groups is detected, and updating the grouping strategy according to the adjustment results.

[0096] In this embodiment, firstly, the feature input interface receives joint physiological features output from the multi-source heterogeneous data fusion model. These features integrate multimodal information such as structured data (e.g., blood pressure, BMI), text semantic features (e.g., medical history description), and image and audio features (e.g., motion video, speech rhythm), representing the overall health status and individual characteristics of the test subject before enrollment.

[0097] Subsequently, the dimensional modeling module performs structured analysis and classification of the input joint physiological features, dividing them into two categories: one is static dimensional factors, such as attributes that do not change drastically over time, including gender, age, disease stage, and medication adherence; the other is dynamic dimensional factors, such as fluctuating parameters that are related to physiological state in real time, including heart rate variability (HRV), diurnal blood pressure rhythm deviation, exercise energy expenditure, and nighttime heart rate decline rate. The role of this module is to build a quantifiable, multi-dimensional comparative foundation for subsequent equilibrium analysis.

[0098] Next, the balance assessment module receives the aforementioned dimensional factors and, based on medical expert experience or system-preset strategies, assigns corresponding clinical importance weights to each dimension. Building upon this, the module simulates the allocation scenarios of current subjects in different alternative groups (e.g., group A, group B) and, based on the intra-group distribution differences of all factors, calculates the complex balance index corresponding to each grouping scheme. :

[0099]

[0100] in, The complex equilibrium index represents the index at the th... The value of each candidate grouping scheme is used to measure the degree of difference between the group and the reference group in terms of feature distribution. The smaller the value, the more balanced the features between the groups. Indicates the first The clinical importance weight of each feature; Indicates the total number of feature dimensions; Indicates the first The standardized cumulative distribution value of the feature in candidate group A; Indicates the first The standardized cumulative distribution value of the feature in reference group B; This represents the weighting adjustment coefficient, used to adjust the degree of influence of reference group B.

[0101] Subsequently, the minimization allocation module takes over the CEI evaluation results and automatically determines the optimal group assignment for the current test subject according to the "minimization principle". If the CEI difference between multiple groups is less than a set threshold (e.g., 5%), a certain degree of randomness is introduced (e.g., selecting the optimal group with an 80% probability and random assignment with a 20% probability) to enhance the naturalness of the distribution between groups and the randomness of the experiment, and to avoid bias caused by excessive determinism.

[0102] Finally, to avoid systemic risks caused by consecutive adverse results (such as low exercise levels or abnormal blood pressure) in a single group during the actual trial, the adjustment module dynamically monitors the trial process. This module analyzes the physiological feedback data of each group after intervention in real time. Once an imbalance between groups or consecutive abnormalities in key indicators are detected (such as three consecutive high-risk indicators), the weights of each dimension factor are automatically and dynamically adjusted. The adjusted weight values ​​are then fed back to the balance assessment module, thereby influencing the subsequent grouping strategy for subjects and forming an intelligent closed-loop control of "prediction-decision-feedback-correction".

[0103] In summary, the dynamic randomization engine, through modeling, grouping evaluation, and feedback regulation of multidimensional features, not only improves the scientific rigor and balance of grouping but also enhances the adaptability and safety control level throughout the entire clinical research process, providing technical support and assurance for exercise prescription intervention research.

[0104] Figure 3 This is a schematic diagram of a health status prediction model provided in another embodiment of this application. The health status prediction model is used to predict the physiological state change trend and health status level of a test subject under different exercise interventions, and based on this, generates personalized intervention recommendations. Figure 3 As shown, the health status prediction model includes, in sequence, an input layer, an intervention encoding layer, a temporal modeling layer, a health status mapping layer, a state decoding layer, and an exercise prescription generator. Specifically, in the input layer, the model receives the joint physiological characteristics of the test subject at the current moment. This vector is extracted by a multi-source data fusion model, including important indicators such as maximum oxygen uptake, heart rate variability, and blood pressure recovery rate; simultaneously, the input layer receives the current exercise intervention vector. This vector characterizes the intervention content the test subject participated in at that moment, such as exercise type (aerobic, resistance), intensity level, duration, and frequency. Next, the intervention encoding layer is entered, which uses an embedding layer or a multi-sensor machine (MLP) to encode the raw intervention information. Transformation with associated physiological characteristics The same-dimensional representation enables the model to recognize the semantics of different intervention combinations (exercise type + intensity + time).

[0105] The temporal modeling layer employs a GRU (Gated Recurrent Unit) with conditional input, which combines the joint physiological features of the current time step. With motion intervention vector Concatenate the vectors to obtain the fused vector. :

[0106]

[0107] The fusion vector As input to the gated recurrent unit (GRU), the output is the healthy hidden state vector at the current time step. :

[0108]

[0109] in, This represents the healthy hidden state vector at the previous moment.

[0110] By leveraging the memory gate mechanism of GRU, the evolutionary trajectory of a test subject's health status over time can be dynamically captured, and the impact of intervention behaviors on status updates can be integrated, thereby establishing a behavior-state joint modeling framework. This process not only models the regularity of physiological indicators changing over time but also perceives the impact of intervention measures on the trajectory, achieving joint modeling of state and behavior.

[0111] The health hidden state vector at the current time point output by the time series modeling layer This is then fed into a health status mapping layer, which uses a set of nonlinear transformation functions (such as ReLU and Softmax) to... Mapped to a multi-category health status level probability vector :

[0112]

[0113] in, This represents the probability distribution of health levels at the current moment; Represents the mapping matrix; Indicates the health level category, such as low, medium, or high; This indicates the bias term.

[0114] at the same time, It is also fed into the state decoding layer to predict the physiological state of the test subject in the next moment. :

[0115]

[0116] in, Represents the state decoding matrix; This indicates the bias term.

[0117] The state decoding layer establishes a mapping relationship between the current health hidden state vector and the future physiological state to achieve trajectory simulation and continuous prediction of an individual's health state.

[0118] Finally, the model outputs a multi-class health status level probability vector. and the physiological state at the next moment The exercise prescription is input into the exercise prescription generator, which serves as the decision output module of the health status prediction model. The generator aims to automatically generate optimal personalized intervention recommendations based on an individual's current physiological state, predicted trends in state change, and health status level. Its operating principle is based on a multi-objective optimization strategy, integrating the probability vectors of multiple health status levels at the current moment. Physiological state at the next moment and current status The final exercise prescription is generated, and the exercise prescription generator is represented as follows:

[0119]

[0120]

[0121]

[0122] in, This represents a nonlinear motion prescription generation operator; Represents a nonlinear modulation function; Represents an adaptive mapping function for health status; Represents the personality adjustment matrix; This represents the weight mapping matrix of the intervention mode; This represents the final decoding mapping matrix, and the output structure is an intervention scheme vector (containing fields such as motion type, intensity, time, and frequency). Indicates perturbation regularization; This represents the harmonic matrix indicating the health status.

[0123] The exercise prescription generator constructs a nonlinear multi-source fusion mapping function. Based on the probability vector of multi-category health status levels at the current moment Physiological state at the next moment and current status Using these as inputs, the moderating effect of individual health change trends and health status levels on intervention programs is jointly modeled. The core of this formula lies in using state differences... Capture potential for physiological improvement and introduce a personalized modulation matrix. By learning historical response patterns, each individual acquires dynamic sensitivity to different intervention variables; simultaneously, health status information is processed via a function. After mapping, the differences in state are fused using a weighted structure. By regulating the direction and intensity of interventions, a dual objective of "state improvement-driven + health state avoidance optimization" is achieved. Ultimately, a nonlinear activation function... With decoding matrix The fused features are mapped into a multidimensional exercise prescription output, including key parameters such as type, intensity, time, and frequency. This design not only possesses high expressiveness and dynamic adaptability, but also achieves difficult-to-retrieve expression forms through the introduction of structural innovation and symbol nesting, effectively improving the individualization, accuracy, and model safety of exercise prescription generation, and providing a highly intelligent and structurally controllable decision-making basis for clinical health intervention systems.

[0124] In summary, the proposed health status prediction model dynamically integrates an individual's combined physiological characteristics with real-time exercise intervention information to construct a health status evolution trajectory across time scales, thereby enabling health status assessment and personalized exercise intervention recommendations. Its operating principle is based on a deep temporal modeling framework, inputting the current physiological state and intervention vector into a conditional input GRU structure through a fusion mechanism, effectively capturing the dynamic patterns of physiological indicators changing over time and the laws governing intervention responses. The model's output latent state not only carries individualized health change trends but is also further mapped to a probability distribution of health status levels and predicted future status values, thus achieving the dual functions of status assessment and prediction. Finally, an exercise prescription generator integrates the current state, future trends, and health status level to generate the optimal intervention combination through reverse reasoning, forming a closed-loop decision chain. This model demonstrates high individual adaptability, temporal continuity, and causal interpretability in its technical effectiveness, enhancing the scientific rigor and foresight of health intervention strategies and strengthening the model's practicality and decision support value in clinical health management scenarios. It serves as a key modeling foundation for achieving precision medicine and intelligent intervention.

[0125] In another exemplary embodiment, the health status prediction model is trained through the following steps:

[0126] S31: Obtain the health status level and corresponding exercise prescription records of multiple test subjects, construct a dataset containing physiological state sequences, intervention information, future state and health status labels, and further divide it into training set and validation set, for example, a division ratio of 6:4.

[0127] S32: Initialize training parameters, for example, set the maximum number of iterations to 100, the Dropout ratio to 0.2~0.5, and train the model using the training set until the maximum number of iterations is reached;

[0128] S33: Validate the model using a validation set. During the validation process, if the root mean square error (RMSE) for state prediction, the accuracy for health status level prediction, and the F1 score for health status level prediction all meet the thresholds (RMSE ≤ 0.08, accuracy ≥ 0.85, F1 ≥ 0.82), the model is validated. Otherwise, modify the training parameters (e.g., adjust the maximum number of iterations to 200) and retrain the model until the model is validated.

[0129] Below, this application will provide an exemplary description of the scheme described in this application with specific data. The current observation data of a certain test object is shown in Table 6:

[0130] Table 6 Current Observation Data

[0131]

[0132] Based on the scheme proposed in this application, the following prediction output results can be obtained:

[0133] Predict the next state:

[0134] VO2max increased to 35.4 ml / kg / min

[0135] HRV improved to 30ms

[0136] Blood glucose level dropped slightly to 6.2 mmol / L

[0137] Systolic blood pressure recovery increased by 25 mmHg

[0138] Risk level prediction (multi-class Softmax output):

[0139] Low risk: 0.12

[0140] Medium risk: 0.31

[0141] High risk: 0.57

[0142] Based on the target improvement direction (cardiorespiratory endurance, blood glucose control) and high-risk warning signals, this application uses the state-risk coupling function to deduce the optimal exercise prescription, as shown in Table 7:

[0143] Table 7 Optimal Exercise Prescription

[0144]

[0145] In summary, the current test subjects are predicted to be at high risk. The recommended personalized exercise prescription is: moderate-intensity aerobic + resistance training 4 times a week, each session lasting 45 minutes. It is recommended to monitor blood glucose response after training.

[0146] In another exemplary embodiment, this application also provides an apparatus for generating exercise prescriptions based on clinical data, such as... Figure 4As shown, the device includes: an acquisition module 100 for acquiring clinical data of the test subjects, wherein the clinical data includes structured data, text data, image data, and audio data; a feature extraction module 200 for extracting joint physiological features from the clinical data based on a multi-source heterogeneous data fusion model; a grouping module 300 for inputting the extracted joint physiological features into a dynamic randomization engine and grouping the joint physiological features based on the dynamic randomization engine; and an exercise prescription generation module 400 for constructing and training a health status prediction model, inputting the grouped joint physiological features into the trained health status prediction model, outputting the health status level of the test subjects and the physiological status evolution trajectory under different exercise interventions, and generating a personalized exercise prescription.

[0147] In another exemplary embodiment, this application also provides a clinical data-based exercise prescription generation system, the system including a clinical data-based exercise prescription generation device as described in the previous embodiment.

[0148] In another exemplary embodiment, this application also provides a storage medium including instructions that, when executed on a computer, cause the computer to perform a clinical data-based exercise prescription generation method as described in any of the preceding embodiments.

[0149] In another exemplary embodiment, this application also provides an electronic device, the electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for generating exercise prescriptions based on clinical data as described in any of the preceding embodiments.

[0150] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for generating exercise prescriptions based on clinical data, characterized in that, The method includes: Acquire clinical data of the test subjects, including structured data, text data, image data, and audio data; Joint physiological features were extracted from the clinical data based on a multi-source heterogeneous data fusion model. The extracted joint physiological features are input into a dynamic randomization engine, and the joint physiological features are grouped based on the dynamic randomization engine. A health status prediction model is constructed and trained. The grouped joint physiological features are input into the trained health status prediction model, which outputs the health status level of the test subjects and the physiological status evolution trajectory under different exercise interventions, and generates personalized exercise prescriptions. The health status prediction model includes: The system consists of an input layer, an intervention encoding layer, a time-series modeling layer, a health state mapping layer, a state decoding layer, and an exercise prescription generator, arranged sequentially. The input layer is used to input the joint physiological characteristics of the test subject at the current moment and the current motion intervention vector extracted by the multi-source heterogeneous data fusion model. The current motion intervention vector is used to characterize the intervention content that the test subject participates in at this moment, including the type of motion, intensity level, duration and frequency. The intervention coding layer is used to convert the current motion intervention vector into a representation with the same dimension as the joint physiological characteristics of the test subject at the current moment; The temporal modeling layer is used to fuse the joint physiological characteristics of the test subject at the current moment with the current motion intervention vector to obtain the health latent state vector; The health status mapping layer is used to map the health hidden state vector to a health status level probability vector; The state decoding layer is used to predict the physiological state of the test subject at the next moment based on the health hidden state vector; The exercise prescription generator is used to generate exercise prescriptions based on the probability vector of health status level and the physiological state of the test subject at the next moment.

2. The method for generating exercise prescriptions based on clinical data according to claim 1, characterized in that, The multi-source heterogeneous data fusion model includes: Structured data encoder, text data extractor, image and audio data encoder, among which, The structured data encoder is used to extract numerical indicators and categorical variable features from structured data; Text data extractors are used to extract latent semantic features from text data; Image and audio data encoders are used to extract key spatiotemporal structure and physiological characterization information from image and audio data.

3. The method for generating exercise prescriptions based on clinical data according to claim 2, characterized in that, The multi-source heterogeneous data fusion model also includes: The decoupling and fusion module is used to enable the interaction and fusion of structured data, text data, image data, and audio data in order to extract joint physiological features across modalities.

4. The method for generating exercise prescriptions based on clinical data according to claim 1, characterized in that, The dynamic randomization engine includes: Feature input interface, used to receive combined physiological features; The dimensional modeling module is used to group joint physiological characteristics into static and dynamic dimensional factors. The balance assessment module is used to assign corresponding clinical importance weights to static and dynamic dimension factors and calculate the complex balance index under different grouping schemes. The minimization allocation module is used to determine the optimal grouping scheme for test objects based on a complex equilibrium index. The adjustment module is used to dynamically adjust the clinical importance weights of static and dynamic dimension factors in the optimal grouping scheme when an imbalance between groups is detected, and to update the grouping strategy based on the adjustment results.

5. The method for generating exercise prescriptions based on clinical data according to claim 1, characterized in that, The health status prediction model is trained through the following steps: Obtain the health status levels and corresponding exercise prescription records of multiple test subjects, construct a dataset containing physiological state sequences, intervention information, future states and health status labels, and divide the dataset into training set and validation set; Initialize the training parameters and train the model using the training set until the maximum number of iterations is reached. The model is validated using a validation set. During the validation process, if the root mean square error (RMSE) of state prediction, the accuracy of health status level prediction, and the F1 score of health status level prediction all meet the thresholds, the model is validated. Otherwise, the training parameters are modified and the model is retrained until the model is validated.

6. A device for generating exercise prescriptions based on clinical data, characterized in that, The device includes: The acquisition module is used to acquire the clinical data of the test subjects, including structured data, text data, image data, and audio data. The feature extraction module is used to extract joint physiological features from the clinical data based on a multi-source heterogeneous data fusion model; The grouping module is used to input the extracted joint physiological features into the dynamic randomization engine, and to group the joint physiological features based on the dynamic randomization engine. The exercise prescription generation module is used to build and train a health status prediction model. It inputs the grouped joint physiological features into the trained health status prediction model, outputs the health status level of the test subject and the physiological status evolution trajectory under different exercise interventions, and generates a personalized exercise prescription. The health status prediction model includes: The system consists of an input layer, an intervention encoding layer, a time-series modeling layer, a health state mapping layer, a state decoding layer, and an exercise prescription generator, arranged sequentially. The input layer is used to input the joint physiological characteristics of the test subject at the current moment and the current motion intervention vector extracted by the multi-source heterogeneous data fusion model. The current motion intervention vector is used to characterize the intervention content that the test subject participates in at this moment, including the type of motion, intensity level, duration and frequency. The intervention coding layer is used to convert the current motion intervention vector into a representation with the same dimension as the joint physiological characteristics of the test subject at the current moment; The temporal modeling layer is used to fuse the joint physiological characteristics of the test subject at the current moment with the current motion intervention vector to obtain the health latent state vector; The health status mapping layer is used to map the health hidden state vector to a health status level probability vector; The state decoding layer is used to predict the physiological state of the test subject at the next moment based on the health hidden state vector; The exercise prescription generator is used to generate exercise prescriptions based on the probability vector of health status level and the physiological state of the test subject at the next moment.

7. A system for generating exercise prescriptions based on clinical data, characterized in that, The system includes the apparatus as described in claim 6.

8. A storage medium, characterized in that, It includes instructions that, when run on a computer, cause the computer to perform the exercise prescription generation method based on clinical data as described in any one of claims 1 to 5.

9. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the exercise prescription generation method based on clinical data as described in any one of claims 1 to 5.

Images