A nursing decision support method applied in clinical medicine

Abstract

This invention discloses a nursing auxiliary decision-making method applied in clinical medicine, belonging to the field of nursing auxiliary decision-making technology. Specifically, it includes: acquiring patient health data, daily routine and exercise data, real-time physiological monitoring data, and wound assessment data; generating individualized rehabilitation tolerance characteristics based on these data; retrieving suitable rehabilitation activities from a rehabilitation activity database according to the rehabilitation tolerance characteristics; allocating rehabilitation activities to different time periods in conjunction with the daily routine and exercise data to generate a rehabilitation execution plan; collecting behavioral feedback data and physiological change data during the plan implementation period; and finally, combining these data to correct the rehabilitation tolerance characteristics and update the execution plan. This invention achieves precise matching between rehabilitation activities and the individual patient's condition, establishing a dynamic balance mechanism between promoting functional recovery and avoiding wound damage.

Description

Technical Field

[0001] This invention relates to the field of nursing decision support technology, and more specifically to a nursing decision support method applied in clinical medicine. Background Technology

[0002] Postoperative rehabilitation care is an important component of the modern medical system, affecting the quality of patients' functional recovery and long-term quality of life. With the development of medical technology, more and more medical institutions are adopting standardized rehabilitation programs to guide clinical practice. These programs are developed based on population data and clinical experience, providing healthcare professionals with a systematic nursing framework.

[0003] Among existing technological solutions, several data-driven rehabilitation nursing methods have emerged. These methods collect basic physiological parameters and wound information from patients, combined with a pre-set rehabilitation activity library, to generate a fixed rehabilitation plan for each patient. Some plans also make appropriate adjustments based on the patient's progress, such as fine-tuning the activity intensity based on patient feedback. These methods have achieved certain results in improving the standardization of rehabilitation nursing.

[0004] However, existing technologies, due to their standardized approach, fail to adequately consider individual patient characteristics and real-time changes in their condition, making it difficult to achieve a precise balance between "promoting functional recovery" and "avoiding wound damage." This lack of a balancing mechanism can lead to a mismatch between the intensity of rehabilitation activities and the patient's actual tolerance, resulting in either insufficient or excessive activity. This not only affects rehabilitation outcomes but may also increase the risk of secondary injury. The ability of existing solutions to dynamically adapt to the patient's rehabilitation process needs further improvement. Summary of the Invention

[0005] The purpose of this invention is to provide a nursing decision support method applicable to clinical medicine, and to solve the following technical problems:

[0006] Existing technical solutions, due to their use of uniform standards, fail to fully incorporate the individual characteristics and real-time changes in the patient's condition, which may lead to a mismatch between the intensity of rehabilitation activities and the patient's actual tolerance.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A nursing decision support method for clinical application includes the following steps:

[0009] S1. Obtain the patient's basic information, real-time physiological monitoring data, and wound assessment data. The basic information includes health data and daily activity data.

[0010] S2. Generate individualized rehabilitation tolerance characteristics based on health data, real-time physiological monitoring data, and wound assessment data;

[0011] S3. Based on the rehabilitation tolerance characteristics, retrieve suitable rehabilitation activity items from the rehabilitation activity database;

[0012] S4. By combining rest and exercise data, rehabilitation activities are allocated to different time periods to generate a rehabilitation execution plan;

[0013] S5. During the implementation of the rehabilitation plan, collect patients' behavioral feedback data and physiological change data;

[0014] S6. Based on behavioral feedback data and physiological change data, revise the individualized rehabilitation tolerance characteristics and update the rehabilitation implementation plan.

[0015] As a further aspect of the present invention: in S1, the health data includes age data, medical history data, and medication record data; the activity and rest data includes sleep duration data, daily activity data, and activity preference data; the real-time physiological monitoring data includes heart rate variability data, blood oxygen saturation data, and respiratory rate data; and the wound assessment data includes wound area change data, wound volume change data, and wound surface temperature distribution data.

[0016] As a further aspect of the present invention: in step S2, the specific process of generating individualized rehabilitation tolerance characteristics is as follows:

[0017] A dynamic graph structure is constructed, where nodes represent health data, real-time physiological monitoring data, and wound assessment data, and edges represent the relationships between data. Multi-scale feature propagation is performed on the dynamic graph structure to extract local neighborhood features and global topological features.

[0018] We used multi-time-window analysis to analyze real-time physiological monitoring data, extracted short-term fluctuation patterns and long-term trend patterns, established a spatial feature extraction network for wound assessment data, and calculated the morphological features and texture distribution features of the wound area.

[0019] Perform cross-modal feature interaction operations to splice together health data features, physiological monitoring features, and wound features to construct a hierarchical feature aggregation architecture. Fuse feature representations of different granularities layer by layer, and generate rehabilitation tolerance feature representations with spatiotemporal correlation through feature recombination operations.

[0020] As a further aspect of the present invention: in step S3, the specific process of retrieving suitable rehabilitation activity items from the rehabilitation activity database is as follows:

[0021] A dynamic feature space mapping relationship is constructed to project rehabilitation tolerance features onto a multi-dimensional feature space. A cluster center identification mechanism based on density peaks is established in the feature space, and a prototype representation of rehabilitation features is generated based on the cluster centers.

[0022] The similarity measure between the current individualized rehabilitation tolerance features and each prototype representation is calculated. Based on the similarity measure, an initial recommendation set of rehabilitation activities is constructed. A multimodal attention mechanism is introduced, and the daily routine and exercise data are used as attention query vectors. The priority of each activity in the initial recommendation set is adjusted by attention weights to generate a final recommendation set that considers feature space distribution and daily routine matching.

[0023] As a further aspect of the present invention, the specific process of constructing the dynamic feature space mapping relationship is as follows:

[0024] Collect rehabilitation tolerance feature sequences from historical rehabilitation cases, construct an encoder-decoder network structure, the encoder maps the input features to the latent space, and train the encoder-decoder network using the rehabilitation tolerance feature sequences;

[0025] In the latent space, the distance matrix between feature vectors is calculated, and an adjacency graph of the feature vectors is constructed based on the distance matrix. The adjacency graph is converted into a low-dimensional feature representation through a graph embedding algorithm. The correspondence between the low-dimensional feature representation and the rehabilitation activity items is established. When a new rehabilitation tolerance feature is input, the distance matrix and adjacency graph structure are updated, and the coordinate position of the low-dimensional feature representation is adjusted.

[0026] As a further aspect of the present invention: in step S4, the specific process of generating the rehabilitation execution plan is as follows:

[0027] A time series prediction model was constructed by inputting rehabilitation activity item attribute information and daily exercise data, establishing a multi-objective optimization function, and introducing activity intensity distribution parameters and patient daily routine parameters.

[0028] Dynamic programming algorithm is used to solve the time allocation scheme, which allocates rehabilitation activities to different time periods, designs transition methods for activity sequences, inserts buffer periods between adjacent activities, and constructs a conflict detection process to identify activity arrangements that conflict with the fixed schedule.

[0029] Generate an execution plan framework that includes time periods, activities, and transition periods. Adjust the time distribution of the activity sequence through iterative calculations and output a rehabilitation execution plan.

[0030] As a further aspect of the present invention: in step S5, the specific process of collecting the patient's behavioral feedback data and physiological change data is as follows:

[0031] After each rehabilitation activity session, activity completion data was recorded, patient self-reported descriptions of physical sensations were collected, and heart rate monitoring data sequences for that session were simultaneously acquired; blood oxygen saturation monitoring data sequences were recorded, fluctuation characteristics of heart rate monitoring data sequences were calculated, and change trends of blood oxygen saturation monitoring data sequences were calculated.

[0032] The activity completion data and physical sensation descriptions are combined into a behavioral feedback dataset, and the heart rate fluctuation characteristics and blood oxygen change trends are combined into a physiological change dataset. A unified timestamp is added to both the behavioral feedback dataset and the physiological change dataset.

[0033] As a further aspect of the present invention: in step S6, the specific process of modifying individualized rehabilitation tolerance characteristics and updating the rehabilitation execution plan is as follows:

[0034] Construct a multimodal data tensor to integrate the activity completion rate from behavioral feedback data and the heart rate fluctuation characteristics from physiological change data; establish a dynamic correction coordinate system in the rehabilitation tolerance feature space, calculate the projection positions of behavioral feedback data points and physiological change data points in the feature space, and generate feature correction vectors through vector operations in the feature space.

[0035] The modified feature vector is superimposed with the original rehabilitation tolerance feature, and the feature dimension is orthogonalized to obtain the modified rehabilitation tolerance feature. The activity recommendation weights are then recalculated, and a new rehabilitation execution plan is generated using the updated activity recommendation weights.

[0036] As a further aspect of the present invention: the specific process of generating the feature correction vector through vector operations in the feature space is as follows:

[0037] Construct a covariance matrix of behavioral feedback data and physiological change data, calculate the eigenvalues ​​and eigenvectors of the covariance matrix, select the k largest eigenvectors to form a projection subspace, k is a set value, map the behavioral feedback data points to the projection subspace to obtain the behavioral feedback projection, and map the physiological change data points to the projection subspace to obtain the physiological change projection.

[0038] The vector difference between the behavioral feedback projection and the physiological change projection is calculated. The vector difference is standardized to obtain a unit vector. The unit vector is linearly combined with the basis vectors of the original feature space, and the feature correction vector is generated by the linear combination result.

[0039] The beneficial effects of this invention are:

[0040] This invention achieves accurate assessment of patients' rehabilitation status by constructing a personalized rehabilitation tolerance feature generation mechanism, effectively solving the problem of insufficient or excessive activity in rehabilitation activities. Employing dynamic graph structures and multi-scale feature propagation technology, it can extract spatiotemporally correlated feature representations from multi-source data, overcoming the limitations of traditional methods in depicting a single patient state. An activity recommendation model based on dynamic feature space mapping establishes a nonlinear mapping relationship between rehabilitation features and activity items, significantly improving the accuracy of activity recommendations. By integrating behavioral feedback and physiological change data through multimodal data tensors, and generating feature correction vectors using vector operations in the feature space, dynamic optimization of the rehabilitation plan is achieved. A multimodal attention mechanism is also introduced, using daily routine data as query vectors to further optimize the matching degree between activity arrangements and patients' daily routines. By establishing a dynamic correction coordinate system and a basis vector adjustment mechanism, it can respond promptly to changes in patient status, establishing an effective balance between promoting functional recovery and avoiding wound damage, significantly improving the safety and effectiveness of rehabilitation care. Attached Figure Description

[0041] The invention will now be further described with reference to the accompanying drawings.

[0042] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 As shown, this invention is a nursing decision support method applied in clinical medicine, comprising the following steps:

[0045] S1. Obtain the patient's basic information, real-time physiological monitoring data, and wound assessment data. The basic information includes health data and daily activity data.

[0046] S2. Generate individualized rehabilitation tolerance characteristics based on health data, real-time physiological monitoring data, and wound assessment data;

[0047] S3. Based on the rehabilitation tolerance characteristics, retrieve suitable rehabilitation activity items from the rehabilitation activity database;

[0048] S4. By combining rest and exercise data, rehabilitation activities are allocated to different time periods to generate a rehabilitation execution plan;

[0049] S5. During the implementation of the rehabilitation plan, collect patients' behavioral feedback data and physiological change data;

[0050] S6. Based on behavioral feedback data and physiological change data, revise the individualized rehabilitation tolerance characteristics and update the rehabilitation implementation plan.

[0051] In S1 of this invention, health data serves as the foundational input for individualized patient characteristics and is stored in a structured data format. Age data is recorded as a continuous numerical variable, accurate to the year of birth, serving as a basic reference for physiological function assessment in subsequent feature association analysis. Medical history data employs a combination of classification and text encoding, covering three core categories of information: past surgical history, chronic disease history, and allergy history. Surgical history requires noting the type of surgery and postoperative recovery period; chronic disease history requires specifying the diagnosis time and disease control level; and allergy history requires recording the type of allergen and past allergic reactions. Medication records are stored in a time-series structured table, including the generic name of the drug, route of administration, single dose, frequency of use, and start and end dates of use, ensuring that subsequent analysis of the potential impact of drugs on physiological indicators and recovery tolerance is conducted.

[0052] Daily activity data was collected using a multimodal approach to capture patients' daily behavioral patterns. Sleep duration data was recorded with minute-level precision using sleep monitoring devices, including sleep onset time, number of awakenings, and the percentage of light sleep / deep sleep / REM sleep, forming a complete sleep cycle time sequence. Daily activity data was collected using wearable devices, calculating total steps, duration of moderate-to-vigorous intensity activities (MET value ≥ 3.0), and duration of sedentary posture on a daily basis to quantify patients' basic activity abilities. Activity preference data was obtained through standardized questionnaires combined with behavioral log analysis, using categorized tags to label patients' preferred activity types (e.g., upper limb activities, lower limb activities, trunk activities), activity times (e.g., morning, afternoon, evening), and activity intensity tendencies (e.g., mild, moderate), providing a basis for behavioral adaptation in subsequent activity allocation.

[0053] Real-time physiological monitoring data is captured dynamically using high-frequency acquisition equipment. Heart rate variability data is acquired at a sampling frequency of 1 Hz, recording the RR interval sequence to extract the time and frequency domain characteristics of heart rate fluctuations. Blood oxygen saturation data is acquired at a frequency of 1 time / minute, recording real-time arterial blood oxygen saturation values ​​to form a continuous time-series curve. Respiratory rate data is acquired via a chest and abdominal motion sensor at a frequency of 1 time / 30 seconds, calculating the number of breaths per minute and simultaneously recording respiratory rhythm stability indicators. All three types of data are uploaded to the data processing terminal in real time via a wireless transmission module to ensure data timeliness.

[0054] Wound assessment data utilizes visualization and quantitative detection technologies to obtain local healing status. Wound area change data is obtained by capturing frontal images of the wound using high-definition image acquisition equipment. Image processing algorithms automatically segment the wound area, calculate pixel areas, and convert them into actual physical areas, recording the area change differences at different assessment time points. Wound volume change data is obtained by acquiring a three-dimensional point cloud model of the wound using 3D laser scanning equipment. Volume calculation algorithms are used to generate the actual wound volume, and volume data at different time points are compared to obtain healing progress information. Wound surface temperature distribution data is acquired using an infrared thermal imager, generating a wound area temperature heat map. The differences between the temperature of the wound's central and peripheral areas and the surrounding normal skin temperature are extracted to quantify the local inflammatory response status of the wound.

[0055] In S2 of this invention, the specific process of generating individualized rehabilitation tolerance characteristics is as follows:

[0056] 1. Dynamic graph structure construction and multi-scale feature propagation

[0057] First, feature encoding is performed on the collected health data, real-time physiological monitoring data, and wound assessment data to construct a node set with a dynamic graph structure. Structured variables in the health data are converted into low-dimensional dense vectors using an embedding layer as node features. The time-series data of the real-time physiological monitoring data have their local features extracted using a temporal convolutional network and then used as node features. Quantitative indicators and image features of the wound assessment data are mapped to vectors of a unified dimension through a fully connected layer and used as node features.

[0058] The edge weights of the dynamic graph structure are calculated based on the statistical correlation and clinical association rules between data. The linear correlation of continuous data (such as age and heart rate variability) is calculated using the Pearson correlation coefficient, and the non-linear correlation of categorical data (such as medical history type and wound inflammation status) is calculated using mutual information values. Combined with the associations clearly defined in clinical guidelines (such as the correlation between anticoagulant medication records and bleeding risk), the initial values ​​of the edge weights are set, and the edge weights are dynamically updated as new data is input, thereby realizing the dynamic adjustment of the graph structure.

[0059] Multi-scale feature propagation is performed on a dynamic graph structure, and feature fusion at different scales is achieved through graph convolutional networks. Local neighborhood feature extraction employs a 1-3 order neighbor node aggregation strategy, summing the features of the target node and its directly and indirectly adjacent nodes according to edge weights to capture local correlation information between data. Global topological feature extraction calculates the global connectivity, node degree distribution, and clustering coefficients of the graph using the Laplacian matrix. This global topological information is encoded into a vector and then concatenated with local neighborhood features to form a preliminary feature representation that combines local correlations and global structure.

[0060] 2. Multi-time window analysis and wound spatial feature extraction

[0061] For real-time physiological monitoring data, a multi-time-window analysis framework was designed to achieve hierarchical extraction of time-series features. Short-term fluctuation pattern extraction uses a sliding window of 5-15 minutes, calculating time-domain indicators (such as standard deviation and range) and frequency-domain indicators (such as low-frequency power and high-frequency power) for the heart rate variability (RR) interval sequence within the window to capture short-term dynamic changes in physiological indicators. Long-term trend pattern extraction uses a fixed window of 24-72 hours, calculating the mean, median, and trend slope for blood oxygen saturation and respiratory rate data within the window to analyze the long-term variation patterns of physiological indicators. The features of the two types of patterns are integrated into a unified physiological monitoring feature vector through feature concatenation.

[0062] A spatial feature extraction network for wound assessment data based on convolutional neural networks was established to mine spatial features from wound area, volume, and temperature distribution data. Wound morphological feature extraction involved progressive downsampling through convolutional and pooling layers to extract geometric features such as perimeter, aspect ratio, and edge irregularity of the wound region. Wound texture distribution features were calculated using the gray-level co-occurrence matrix to quantify the texture changes of the wound surface tissue by measuring energy, entropy, contrast, and correlation indices. Wound temperature distribution features were captured by convolutional kernels to observe temperature gradient changes and hotspot distribution in the temperature heatmap. These three spatial features were fused into a wound spatial feature vector through a fully connected layer, achieving accurate characterization of the wound healing state.

[0063] 3. Cross-modal feature interaction and hierarchical feature aggregation

[0064] To perform cross-modal feature interaction, the first step is to align the dimensions of health data features, physiological monitoring features, and wound features. A feature mapping layer then transforms these three types of features into vectors of the same dimension. An attention mechanism is employed to enable cross-modal feature interaction. A similarity matrix between different modal features is calculated, and attention weights are assigned based on similarity. Features with high correlation (such as wound inflammation status features and heart rate fluctuation features) receive higher weights, highlighting key cross-modal correlation information.

[0065] A hierarchical feature aggregation architecture is constructed to achieve the gradual fusion of features at different granularities. The bottom layer is an intra-modal feature fusion layer, which performs batch normalization and activation function processing on features of different dimensions within a single modality (such as age, medical history, and medication records in health data) to enhance the consistency of features within the modality. The middle layer is a bimodal interaction fusion layer, which performs feature cross-operations on health data-physiological monitoring data and physiological monitoring data-wound data respectively, generating bimodal interaction features through element-wise multiplication and addition operations to capture collaborative information between the two modalities. The top layer is a trimodal global fusion layer, which concatenates bimodal interaction features with third-modal features, eliminating feature redundancy through residual connections and batch normalization operations to improve feature expressive power.

[0066] The rehabilitation tolerance feature representation is generated through feature recombination. The feature vectors after hierarchical aggregation are then subjected to dimensional filtering and sequential recombination. Based on prior clinical knowledge and feature importance assessment (e.g., feature contribution calculated using random forests), feature dimensions strongly correlated with rehabilitation tolerance are retained, while redundant dimensions are eliminated. The feature vectors are recombined in the order of "basic health features - real-time physiological features - wound status features" to ensure logical coherence. Ultimately, personalized rehabilitation tolerance features with spatiotemporal correlation are generated, providing accurate feature input for subsequent rehabilitation activity recommendations.

[0067] In S3 of this invention, the specific process of retrieving suitable rehabilitation activity items from the rehabilitation activity database is as follows:

[0068] 1. Construction process of dynamic feature space mapping relationship

[0069] First, rehabilitation tolerance feature sequences were collected from historical rehabilitation cases. These sequences needed to cover patient data from different diseases and rehabilitation stages. Each sequence contained individualized rehabilitation tolerance feature vectors for multiple consecutive days, and was associated with the execution records and effect evaluation results of rehabilitation activities in the corresponding cases (such as activity completion rate and wound healing progress). Based on this dataset, an encoder-decoder network structure was constructed. The encoder adopted a combination architecture of a multi-layer fully connected network and a temporal convolutional module. The fully connected layer was responsible for dimensionality compression of tolerance features at a single time point, while the temporal convolutional module captured the temporal correlation of the feature sequences. Finally, the high-dimensional input feature sequences were mapped to a low-dimensional latent space, and the output latent feature vectors of fixed dimensions were generated. The decoder adopted a structure symmetrical to the encoder, reconstructing the latent feature vectors into feature sequences of the original dimensions. The network training objective was set to minimize the reconstruction error, i.e., optimizing the parameters of the encoder and decoder through backpropagation, so that the features in the latent space could accurately invert the original feature information, ensuring the effectiveness of the latent features in representing the patient's rehabilitation tolerance status.

[0070] In the latent space output by the trained encoder, a distance matrix is ​​calculated between the latent feature vectors of all historical cases. Euclidean distance is used as the distance metric, and a symmetric distance matrix is ​​generated by calculating the distance between each pair of feature vectors. An adjacency graph of the feature vectors is constructed based on this distance matrix. A distance threshold is set; when the distance between two feature vectors is less than the threshold, an undirected edge is created for the corresponding node in the adjacency graph. The edge weight is set to the reciprocal of the distance; the smaller the distance, the larger the edge weight, reflecting the strength of the association between nodes. A Graph Convolutional Network (GCN) is used as the graph embedding algorithm. The constructed adjacency graph is input into the GCN model, and multi-layer graph convolution operations are used to extract the local and global topological information of the nodes, converting the high-dimensional adjacency graph into a low-dimensional feature representation. This allows the distance between nodes in the low-dimensional space to reflect the correlation of features in the original latent space.

[0071] A correspondence between low-dimensional feature representations and rehabilitation activities is established. Rehabilitation activities in historical cases are categorized and coded according to type (e.g., limb movement, breathing training, balance training) and intensity (e.g., mild, moderate, severe). The performance of rehabilitation activities corresponding to each low-dimensional feature cluster is statistically analyzed, and the activity type and intensity with the best performance are assigned to that feature cluster, forming a mapping table between low-dimensional features and rehabilitation activities. When new individualized rehabilitation tolerance features are input, they are first converted into latent feature vectors by an encoder, then the distance between this vector and historical latent feature vectors is calculated, and the distance matrix is ​​updated. The edge connections and weights of the adjacency graph are adjusted based on the updated distance matrix. The updated adjacency graph is then input into the GCN model to recalculate the coordinate positions of the low-dimensional feature representations, ensuring that the low-dimensional feature space remains synchronized with the latest input patient features.

[0072] 2. Cluster center identification and rehabilitation feature prototype generation

[0073] In the dynamically updated low-dimensional feature space, a cluster center identification mechanism based on density peaks is established. First, the local density of each low-dimensional feature vector is calculated, defined as the number of other feature vectors within a defined radius centered on that vector. Then, the distance between each vector and its nearest higher-density vector is calculated; this distance is called the "relative distance." Feature vectors with both high local density and high relative distance are selected as cluster centers, each representing a set of patient features with similar rehabilitation tolerance. Based on the cluster centers, a prototype representation of rehabilitation features is generated. This prototype representation is the low-dimensional feature vector corresponding to the cluster center, while also integrating common features of all feature vectors within the cluster to which the cluster center belongs (such as average physiological index thresholds and wound healing stage), enabling the prototype representation to accurately summarize the rehabilitation tolerance characteristics of a class of patients.

[0074] 3. Similarity Calculation and Multimodal Attention Adjustment

[0075] The similarity metric is calculated between the low-dimensional representation of the current individualized rehabilitation tolerance characteristics and each prototype representation. Cosine similarity is used, determined by calculating the cosine of the angle between two vectors; the closer the cosine value is to 1, the higher the similarity. Based on this similarity metric, an initial recommendation set of rehabilitation activities is constructed. The top N prototype representations with the highest similarity are selected, and the corresponding rehabilitation activities are extracted to form the initial recommendation set. This set includes core information such as activity type, intensity, and suggested duration.

[0076] A multimodal attention mechanism is introduced to convert collected daily routine and exercise data into attention query vectors. First, sleep periods, daily activity levels, and activity preferences in the routine and exercise data are encoded as features. Sleep periods are encoded as time-segment vectors (e.g., morning, afternoon, evening), daily activity levels as numerical vectors, and activity preferences as category vectors. These features are concatenated to form a unified-dimensional attention query vector. The attention weight between this query vector and the feature vectors of each rehabilitation activity in the initial recommendation set (e.g., suggested activity time, required physical exertion, activity type) is calculated. The weight is based on the vector dot product; a larger dot product value indicates a higher weight. The priority of each activity in the initial recommendation set is adjusted according to the attention weights, with higher-weighted activities ranked first. Finally, a final set of recommended activities is generated, considering both feature space distribution (tolerance adaptation) and routine matching (efficiency of execution).

[0077] In S4 of this invention, the specific process of generating the rehabilitation execution plan is as follows:

[0078] First, a time series prediction model is constructed, using a Long Short-Term Memory (LSTM) network as the core architecture. The input data contains two types of core information. The attribute information of rehabilitation activities needs to be structured and encoded, specifically covering activity type (such as limb training, respiratory rehabilitation), standard execution duration, physical exertion level (such as mild 3 MET, moderate 5 MET), and required equipment dependence indicators (such as whether rehabilitation equipment is needed). The daily routine and exercise data need to be converted into time series vectors, including the start and end times of sleep periods, fixed meal times, and the distribution of time periods and durations of daily high-frequency activities (such as commuting, work).

[0079] Based on the above inputs, a multi-objective optimization function is established, which includes two core optimization directions. First, an activity intensity distribution parameter is introduced. This parameter sets intensity thresholds for different time periods each day according to clinical rehabilitation guidelines. For example, the intensity threshold for the morning period (6:00-9:00) is set to moderate or below to avoid high-intensity activities on an empty stomach. Second, a patient's daily routine parameter is introduced. This parameter uses a binary vector to label fixed routine periods (e.g., sleep periods are labeled "1", and allowed activity periods are labeled "0") to ensure that fixed routines are prioritized during optimization. The goal of the optimization function is to maximize the matching degree between activity intensity and time period thresholds, minimize the overlap between activity periods and fixed routines, and simultaneously satisfy the constraint that the total daily activity duration falls within the patient's tolerance range.

[0080] A dynamic programming algorithm is used to solve the time slot allocation scheme. The algorithm's state is defined as "current time slot - set of allocated activities - cumulative intensity value", and the decision variable is "whether to allocate a certain rehabilitation activity to the current time slot". First, the daily time granularity is divided into 15-minute segments to form a discrete time slot sequence. Then, the recommended set of rehabilitation activities is traversed, and the suitability score of each activity in each time slot is calculated (based on intensity matching degree and overlap of work and rest). Through the state transition process, the activity combination with the highest suitability score is gradually selected, and finally the initial time slot allocation scheme is output.

[0081] The design incorporates a transition mechanism for activity sequences, setting buffer periods based on the intensity differences between adjacent activities: when the intensity of a subsequent activity is higher than the preceding activity, the buffer period is set to 10-15 minutes to allow the patient's physiological state to recover; when the intensity of a subsequent activity is lower than or equal to the preceding activity, the buffer period is set to 5-10 minutes to reduce wasted time. A conflict detection process is constructed, implemented through temporal comparison: the activity periods in the initial allocation scheme are compared segment by segment with the fixed periods in the daily activity data. If the overlap between the activity period and the fixed period exceeds a set threshold (e.g., 30%), it is considered a work-rest conflict, and the activity allocation scheme is marked as needing adjustment.

[0082] A framework for generating an execution plan is created, comprising three core fields: time period (accurate to start and end minutes), activity items (including type, intensity, and key execution points), and transition period (start and end minutes and functional description). Iterative calculations are initiated based on conflict detection results. For activities marked as conflicting, their fit scores in other unoccupied time periods are recalculated, and the second-best fit time period is selected for reallocation. Simultaneously, the transition periods for related activities are adjusted to ensure overall sequence coherence. During iteration, the total daily activity intensity and total duration are continuously verified to ensure compliance with the optimization function constraints until no conflicts exist and constraints are satisfied, at which point the final rehabilitation execution plan is output.

[0083] In S5 of this invention, the specific process of collecting the patient's behavioral feedback data and physiological change data is as follows:

[0084] After each rehabilitation activity session, the rehabilitation management terminal automatically records the activity completion data. The recording dimensions include the ratio of actual completion time to standard time, specific indicators of incomplete parts (such as a set of movements not completed), and reasons for interruption (such as physical exhaustion or pain). Simultaneously, a standardized questionnaire pops up on the terminal to collect patients' self-reported descriptions of their physical sensations. The descriptions are quantified through grading options, including pain level (e.g., 1-5), fatigue level (e.g., 1-5), and the presence or absence of abnormal sensations (e.g., dizziness, wound stinging).

[0085] During the activity period, real-time physiological monitoring equipment collected heart rate monitoring data sequences at a frequency of 1 Hz and blood oxygen saturation monitoring data sequences at a frequency of 1 time per minute. The data was synchronized to the data processing terminal via a wireless transmission module. The fluctuation characteristics of the heart rate monitoring data sequences were calculated, specifically including the maximum, minimum, and mean values ​​within the sequence, as well as the statistical differences between adjacent data points. The trend of change of the blood oxygen saturation monitoring data sequences was calculated, and the overall slope of the sequence was obtained through linear fitting to determine whether the blood oxygen level was rising, stable, or declining.

[0086] Activity completion data and bodily sensation descriptions are combined to form a behavioral feedback dataset. Each record in this dataset includes an activity ID, completion percentage, reason code for incomplete completion, pain level, fatigue level, and abnormal sensation identifier, ensuring that each field is directly associated with the corresponding activity. Heart rate fluctuation characteristics and blood oxygen change trends are combined to form a physiological change dataset. Each record includes an activity ID, maximum heart rate, minimum heart rate, mean heart rate, heart rate difference amplitude, and blood oxygen trend slope, achieving a precise correspondence between physiological characteristics and activities.

[0087] A unified timestamp identifier is added to the two types of datasets. The timestamp adopts the format of "year-month-day, hour:minute:second" and takes the value as the end time of the corresponding rehabilitation activity period. This ensures that the behavioral feedback data and physiological change data of the same activity can be associated and matched through timestamps, providing time-consistent multimodal data support for the subsequent feature correction of S6.

[0088] In S6 of this invention, the specific process of modifying individualized rehabilitation tolerance characteristics and updating the rehabilitation execution plan is as follows:

[0089] 1. Construction of multimodal data tensors and establishment of dynamically corrected coordinate systems

[0090] First, a multimodal data tensor is constructed to integrate core features from collected behavioral feedback data and physiological change data. For behavioral feedback data, key indicators of activity completion are selected, including the ratio of actual completion time to standard time, the number of incomplete steps, and the number of interruptions. For physiological change data, core dimensions of heart rate fluctuation characteristics are selected, including maximum heart rate, minimum heart rate, heart rate fluctuation amplitude, and mean heart rate deviation (the difference from resting heart rate). Both types of data undergo standardization preprocessing, mapping all indicator values ​​to the [0,1] interval to eliminate dimensional differences. The data tensor's dimensional design includes a time dimension, a feature modality dimension, and a feature indicator dimension. The time dimension corresponds to the execution time of rehabilitation activities (e.g., morning, afternoon, and evening periods each day). The feature modality dimension is divided into "behavioral feedback" and "physiological changes." The feature indicator dimension includes the standardized specific indicators mentioned above. Tensor construction achieves structured integration of multi-source feedback data, ensuring temporal consistency and modal integrity of data associations.

[0091] A dynamically modified coordinate system is established in the rehabilitation tolerance feature space. This coordinate system uses the dimensions of the original individualized rehabilitation tolerance features as the reference axes, and the range of values ​​for the axes is set according to the clinical rehabilitation tolerance threshold. For example, the coordinate axis range for the heart rate tolerance dimension refers to the safe heart rate range for patients of the same age, and the wound tolerance dimension refers to the safe activity intensity range corresponding to the wound healing stage. The projection positions of behavioral feedback data points and physiological change data points in this coordinate system are calculated: the standardized indicators of the behavioral feedback data are converted into vectors consistent with the coordinate system dimensions through a feature mapping layer, and their projected coordinates on each coordinate axis are determined by vector dot product operations; similarly, the same operation is performed on the standardized indicators of the physiological change data to obtain the corresponding projected coordinates. The projection process must ensure that the coordinate values ​​of the data points are positively correlated with the rehabilitation tolerance ability. For example, the higher the activity completion rate, the larger the projected coordinate value; the greater the heart rate fluctuation exceeds the safe range, the smaller the projected coordinate value, providing an intuitive representation of tolerance deviation for subsequent vector operations.

[0092] 2. The process of generating feature correction vectors

[0093] A covariance matrix is ​​constructed for behavioral feedback data and physiological change data. The rows and columns of the matrix correspond to the integrated multimodal feature indicators (such as completion rate, number of interruptions, maximum heart rate, and heart rate fluctuation amplitude), respectively. Matrix elements are the covariance between any two feature indicators, used to quantify the degree of linear correlation between the indicators. The eigenvalues ​​and eigenvectors of this covariance matrix are calculated. The eigenvalues ​​represent the contribution of the corresponding eigenvectors, and the eigenvectors reflect the direction of the linear combination of feature indicators. A value k is set based on the importance of the clinical rehabilitation features (usually 3-5, determined according to the number of feature dimensions and clinical experience). The top k eigenvectors with the largest eigenvalues ​​are selected; these vectors encompass the main variation information of the multimodal data, forming a low-dimensional projection subspace.

[0094] Behavioral feedback data points and physiological change data points are mapped to the projection subspace respectively. For the behavioral feedback data point vector, a low-dimensional representation in the projection subspace, i.e., the behavioral feedback projection, is obtained through a linear combination operation with the first k feature vectors. The same linear combination operation is used to convert the physiological change data point vector into the physiological change projection in the projection subspace. The projection process must retain the tolerance deviation information of the original data points to ensure that the relative positions of the two types of projections in the subspace reflect the consistency between behavioral feedback and physiological changes (e.g., when the completion rate is low and the heart rate fluctuation is large, both types of projections are located in the low tolerance region of the subspace).

[0095] The vector difference between the behavioral feedback projection and the physiological change projection is calculated. The direction and magnitude of this difference indicate whether the tolerance deviations reflected by the two types of data are consistent: if the magnitude of the difference vector is small, it indicates that the behavioral feedback and physiological changes represent tolerance consistently; if the magnitude is large, the validity of data collection needs to be further verified (e.g., excluding patient self-reporting bias or equipment monitoring errors). The vector difference is standardized by using the Min-Max normalization method to map the values ​​of each dimension of the difference vector to the interval [-1, 1], resulting in a unit vector. The direction of this vector clarifies the direction of tolerance correction (a positive direction indicates that tolerance assessment needs to be improved, and a negative direction indicates that it needs to be reduced), while the magnitude reflects the strength of the correction.

[0096] The unit vector is linearly combined with the basis vectors of the original rehabilitation tolerance feature space to generate a feature correction vector. The weights of the linear combination are allocated according to the clinical contribution of the feature dimensions. For example, the basis vectors of the wound-related dimension have higher weights than those of the baseline health dimension (if the wound is in the inflammatory phase), and the basis vectors of the heart rate-related dimension have higher weights than those of the respiratory rate dimension (if heart rate fluctuations are abnormally significant). The combination process is achieved through weighted summation to ensure that the correction vector is precisely applied to the dimension of the original features that is most strongly correlated with the current tolerance deviation, avoiding feature distortion caused by indiscriminate correction.

[0097] 3. Feature Correction and Generation of New Rehabilitation Implementation Plans

[0098] The modified feature vector is element-wise superimposed with the original individualized rehabilitation tolerance features. The superposition direction is determined by the sign of the modified vector: if a dimension of the modified vector is positive, it indicates that the original feature's tolerance assessment for that dimension is too low and needs to be improved through superposition; if it is negative, it indicates that the assessment is too high and needs to be reduced through superposition. The superimposed feature vector is then orthogonalized using the Schmidt orthogonalization method, processing each feature dimension sequentially to eliminate linear correlations between different dimensions. This ensures that each dimension of the modified feature can independently represent the patient's rehabilitation tolerance attributes, avoiding subsequent recommendation biases caused by dimensional redundancy. After orthogonalization, the modified individualized rehabilitation tolerance features are output, which reflect the patient's current actual tolerance capacity in real time.

[0099] Based on the revised individualized rehabilitation tolerance characteristics, the recommended weights of rehabilitation activities are recalculated. The dynamic feature space mapping relationship constructed in S3 is invoked to project the revised features into a low-dimensional feature space, and the cosine similarity between these features and the prototype representations of each rehabilitation feature is recalculated. The recommended weights of corresponding rehabilitation activities are adjusted according to the similarity, with higher similarity activities receiving greater weights. Simultaneously, the multimodal attention mechanism in S3 is retained, and the weights are fine-tuned using daily routine and exercise data to ensure that activity recommendations simultaneously meet tolerance adaptation and routine matching requirements.

[0100] A new rehabilitation execution plan is generated using the updated activity recommendation weights. The S4 time series prediction model and multi-objective optimization function are reused, with the rehabilitation activities corresponding to the new weights as input, to resolve the time slot allocation scheme. An activity sequence transition design and conflict detection process is implemented to adjust the buffer period length and activity time allocation, ensuring no work-rest conflicts and that the intensity distribution conforms to the tolerable range of the revised characteristics. The final output is a new rehabilitation execution plan containing the updated activity items, time slots, and transition periods, achieving dynamic synchronization between the execution plan and the patient's real-time tolerance status.

[0101] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A nursing decision support method applied in clinical medicine, characterized in that, Includes the following steps: S1. Obtain the patient's basic information, real-time physiological monitoring data, and wound assessment data. The basic information includes health data and daily activity data. S2. Generate individualized rehabilitation tolerance characteristics based on health data, real-time physiological monitoring data, and wound assessment data; S3. Based on the rehabilitation tolerance characteristics, retrieve suitable rehabilitation activity items from the rehabilitation activity database; S4. By combining rest and exercise data, rehabilitation activities are allocated to different time periods to generate a rehabilitation execution plan; S5. During the implementation of the rehabilitation plan, collect patients' behavioral feedback data and physiological change data; S6. Combine behavioral feedback data and physiological change data to revise individualized rehabilitation tolerance characteristics and update the rehabilitation implementation plan; In step S6, the specific process of modifying individualized rehabilitation tolerance characteristics and updating the rehabilitation execution plan is as follows: Construct a multimodal data tensor to integrate the activity completion rate from behavioral feedback data and the heart rate fluctuation characteristics from physiological change data; establish a dynamic correction coordinate system in the rehabilitation tolerance feature space, calculate the projection positions of behavioral feedback data points and physiological change data points in the feature space, and generate feature correction vectors through vector operations in the feature space. The feature correction vector is superimposed with the original rehabilitation tolerance feature, and the feature dimension orthogonalization operation is performed on the superimposed feature to obtain the corrected rehabilitation tolerance feature. The activity recommendation weight is then recalculated. A new rehabilitation implementation plan will be generated using the updated activity recommendation weights; The specific process of generating the feature correction vector through vector operations in the feature space is as follows: Construct a covariance matrix of behavioral feedback data and physiological change data, calculate the eigenvalues ​​and eigenvectors of the covariance matrix, select the k largest eigenvectors to form a projection subspace, k is a set value, map the behavioral feedback data points to the projection subspace to obtain the behavioral feedback projection, and map the physiological change data points to the projection subspace to obtain the physiological change projection. The vector difference between the behavioral feedback projection and the physiological change projection is calculated. The vector difference is standardized to obtain a unit vector. The unit vector is linearly combined with the basis vectors of the original feature space, and the feature correction vector is generated by the linear combination result.

2. The nursing decision support method applied in clinical medicine according to claim 1, characterized in that, In S1, the health data includes age data, medical history data, and medication record data, and the daily routine and exercise data includes sleep period data, daily activity level data, and activity preference data. Real-time physiological monitoring data includes heart rate variability data, blood oxygen saturation data, and respiratory rate data; wound assessment data includes wound area change data, wound volume change data, and wound surface temperature distribution data.

3. The nursing decision support method applied in clinical medicine according to claim 1, characterized in that, In step S2, the specific process of generating individualized rehabilitation tolerance characteristics is as follows: A dynamic graph structure is constructed, where nodes represent health data, real-time physiological monitoring data, and wound assessment data, and edges represent the relationships between data. Multi-scale feature propagation is performed on the dynamic graph structure to extract local neighborhood features and global topological features. We used multi-time-window analysis to analyze real-time physiological monitoring data, extracted short-term fluctuation patterns and long-term trend patterns, established a spatial feature extraction network for wound assessment data, and calculated the morphological features and texture distribution features of the wound area. Perform cross-modal feature interaction operations to splice together health data features, physiological monitoring features, and wound features to construct a hierarchical feature aggregation architecture. Fuse feature representations of different granularities layer by layer, and generate rehabilitation tolerance feature representations with spatiotemporal correlation through feature recombination operations.

4. The nursing decision support method applied in clinical medicine according to claim 1, characterized in that, In step S3, the specific process of retrieving suitable rehabilitation activity items from the rehabilitation activity database is as follows: A dynamic feature space mapping relationship is constructed to project rehabilitation tolerance features onto a multi-dimensional feature space. A cluster center identification mechanism based on density peaks is established in the feature space, and a prototype representation of rehabilitation features is generated based on the cluster centers. The similarity measure between the current individualized rehabilitation tolerance features and each prototype representation is calculated. Based on the similarity measure, an initial recommendation set of rehabilitation activities is constructed. A multimodal attention mechanism is introduced, and the daily routine and exercise data are used as attention query vectors. The priority of each activity in the initial recommendation set is adjusted by attention weights to generate a final recommendation set that considers feature space distribution and daily routine matching.

5. A nursing decision support method applied in clinical medicine according to claim 4, characterized in that, The specific process for constructing the dynamic feature space mapping relationship is as follows: Collect rehabilitation tolerance feature sequences from historical rehabilitation cases, construct an encoder-decoder network structure, the encoder maps the input features to the latent space, and train the encoder-decoder network using the rehabilitation tolerance feature sequences; In the latent space, the distance matrix between feature vectors is calculated, and an adjacency graph of the feature vectors is constructed based on the distance matrix. The adjacency graph is converted into a low-dimensional feature representation through a graph embedding algorithm. The correspondence between the low-dimensional feature representation and the rehabilitation activity items is established. When a new rehabilitation tolerance feature is input, the distance matrix and adjacency graph structure are updated, and the coordinate position of the low-dimensional feature representation is adjusted.

6. The nursing decision support method applied in clinical medicine according to claim 1, characterized in that, In step S4, the specific process of generating the rehabilitation execution plan is as follows: A time series prediction model was constructed by inputting rehabilitation activity item attribute information and daily exercise data, establishing a multi-objective optimization function, and introducing activity intensity distribution parameters and patient daily routine parameters. Dynamic programming algorithm is used to solve the time allocation scheme, which allocates rehabilitation activities to different time periods, designs transition methods for activity sequences, inserts buffer periods between adjacent activities, and constructs a conflict detection process to identify activity arrangements that conflict with the fixed schedule. Generate an execution plan framework that includes time periods, activities, and transition periods. Adjust the time distribution of the activity sequence through iterative calculations and output a rehabilitation execution plan.

7. The nursing decision support method applied in clinical medicine according to claim 1, characterized in that, In step S5, the specific process of collecting the patient's behavioral feedback data and physiological change data is as follows: After each rehabilitation activity session, activity completion data was recorded, patient self-reported descriptions of physical sensations were collected, and heart rate monitoring data sequences for that session were simultaneously acquired; blood oxygen saturation monitoring data sequences were recorded, fluctuation characteristics of heart rate monitoring data sequences were calculated, and change trends of blood oxygen saturation monitoring data sequences were calculated. The activity completion data and physical sensation descriptions are combined into a behavioral feedback dataset, and the heart rate fluctuation characteristics and blood oxygen change trends are combined into a physiological change dataset. A unified timestamp is added to both the behavioral feedback dataset and the physiological change dataset.

Images