Method for automatic generation of individualized care plan based on standardized nursing level and health record

Abstract

The application discloses a personalized nursing plan automatic generation method based on standardized nursing grades and health records, and relates to the technical field of intelligent medical treatment. A deep embedding clustering network is used to process multiple source health records of patients, a dynamic record-grade basic correspondence relationship is constructed, and a random forest algorithm is used to mine non-explicit information such as behavior track and missed medical advice, so as to identify key features and implicit nursing needs. The change of the key features is continuously monitored, the corresponding relationship is updated through incremental learning, and an enhanced index record mapping model is constructed, the change of the disease condition is monitored in real time, and an adjustment trigger signal is automatically generated. An adaptive optimization algorithm is used to iteratively fine-tune the nursing plan parameters, and the correspondence relationship between the needs and the nursing components is continuously optimized through a matching degree evaluation and a feature re-clustering mechanism, so that a closed-loop adaptive intelligent engine is formed, and dynamic personalized management and continuous adaptation of the whole life cycle of the nursing plan are realized.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, specifically a method for automatically generating personalized care plans based on standardized nursing levels and health records. Background Technology

[0002] The development of elderly care plans is directly related to the health, safety, and quality of life of the elderly. Against the backdrop of an increasingly serious aging population, research in this field is of critical significance because it not only affects individual care outcomes but also involves the rational allocation of medical resources and the reduction of social burden. Currently, many care plans still rely heavily on standardized care levels stipulated by the state or industry. While these levels can provide basic protection, they often fail to fully integrate the patient's complete health record information, leading to discrepancies between the plan and the actual situation. In particular, when processing multi-source record data, existing methods struggle to establish a flexible correspondence between indicators and records, making it impossible to specifically reflect the unique disease characteristics of certain patients.

[0003] This lack of correspondence further exacerbates the difficulty of capturing implicit needs. Health records contain a wealth of non-explicit information, such as missed records from previous treatment plans, behavioral patterns, and potential complication risks. If this information is not effectively extracted, care plans may overlook crucial details. For example, a diabetic elderly person living alone with limited joint mobility might only receive routine blood glucose monitoring and turning care under a standardized plan. However, the record shows that they have repeatedly missed turning on time, and their solitary living environment increases the risk of falls. If these implicit factors are not identified in time, the care plan may encounter safety hazards or be ineffective during implementation. The dynamic changes in record data further complicate the issue. A patient's condition may improve or new complications may arise at any time, requiring updates to the health record. However, existing plans often lack corresponding adjustment mechanisms, causing plans to become fixed once generated and unable to iterate with changes. This contradiction between rigidity and the dynamism of records further amplifies the risk of unmet implicit needs. Summary of the Invention

[0004] The purpose of this invention is to provide an automated method for generating personalized care plans based on standardized nursing levels and health records, so as to realize dynamic personalized management and continuous adaptation of care plans throughout their entire life cycle.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This application provides a method for automatically generating personalized care plans based on standardized care levels and health records, including the following steps:

[0007] S1. By extracting multi-source data from patient health records, and using a deep embedding clustering network to group the multi-source data, the basic correspondence between standardized levels and records is obtained.

[0008] S2. Based on the aforementioned basic correspondence, obtain non-explicit information in the archives, such as previously missed execution records and behavioral trajectories, and use the random forest algorithm to mine potential complication risks and determine the key features of implicit needs.

[0009] S3. If the key features exceed the preset threshold, the corresponding relationship is updated for the dynamically changing part. By integrating the updated corresponding relationship with the key features, an enhanced indicator file mapping model is obtained.

[0010] S4. Using the enhanced indicator file mapping model, monitor the file updates for patient condition improvement or new complications, determine whether the plan adjustment mechanism is triggered, and obtain the trigger signal for real-time iteration.

[0011] S5. Using the trigger signal, extract adjustment parameters from the enhanced indicator file mapping model, use the adaptive moment estimation optimization algorithm to fine-tune the nursing plan parameters, and determine the iterative plan version;

[0012] S6. Based on the iterative plan version, obtain the latest implicit requirement features from the multi-source data, determine the matching degree between the features and the plan version, and if the matching degree is lower than a preset threshold, reprocess the latest implicit requirement features to obtain an optimized correspondence update.

[0013] S7. Through the optimized correspondence update, it is integrated into the enhanced indicator file mapping model to determine the final dynamic mapping structure, which is used to support the continuous adaptive adjustment of the care plan.

[0014] The beneficial effects of this invention are as follows:

[0015] By employing a deep embedding clustering network to group and process multi-source patient health record data, and combining feature coding fusion and clustering optimization, a basic correspondence between standardized levels and records is constructed. This effectively solves the problem that existing plans rely on standardized nursing levels and are difficult to integrate with complete health record information. It achieves accurate integration and standardized representation of multi-source heterogeneous record data. Through cluster label mapping and anomaly review mechanism, the standardized level allocation is made more in line with the actual health status of patients, which greatly reduces the deviation between the plan and the actual situation. At the same time, a flexible correspondence between indicators and records is established, which fully reflects the unique disease characteristics of patients and improves the basic adaptability of care plans.

[0016] By mining non-explicit information (such as past execution omissions and behavioral trajectories) in health records using the random forest algorithm, and by calculating key features of implicit needs and identifying potential complication risks through dual-dimensional feature importance scores, the pain points of capturing implicit needs and easily overlooking important details have been successfully addressed. This has enabled the effective extraction and value transformation of non-explicit information, and can accurately locate the personalized implicit care needs of patients. For example, it can identify implicit factors such as the risk of falls and missed turning over for elderly people living alone with diabetes and limited joint mobility. This can avoid safety hazards and shortcomings in care plans, and realize the advancement of care plans from routine coverage to precise adaptation.

[0017] By incrementally learning to update correspondences, using an enhanced indicator file mapping model to monitor dynamic changes in files in real time, and employing an adaptive moment estimation optimization algorithm for parameter fine-tuning and matching degree verification in a closed-loop iterative mechanism, this approach resolves the core contradiction of existing solutions being too rigid and unable to keep up with dynamic file updates. It achieves real-time linkage between nursing plans and changes in patient conditions, enabling timely responses to dynamic changes such as improvement in condition or new complications. Automated parameter fine-tuning and plan iteration ensure the continuous adaptability of care plans. Furthermore, by optimizing the cyclical update of correspondences to construct a closed-loop adjustment system, it reduces ineffective investment of medical resources and lowers the risk of unmet implicit needs, further improving the quality of care and alleviating the social medical burden. Attached Figure Description

[0018] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0019] Figure 1 A flowchart illustrating the automated generation method for personalized care plans based on standardized nursing levels and health records provided in Embodiment 1 of this application;

[0020] Figure 2 This is a flowchart illustrating step S2 in the method for automatically generating personalized care plans based on standardized nursing levels and health records provided in Embodiment 1 of this application.

[0021] Figure 3 This is a flowchart illustrating step S3 in the method for automatically generating personalized care plans based on standardized nursing levels and health records provided in Embodiment 1 of this application. Detailed Implementation

[0022] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0023] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0025] Example 1

[0026] Please see Figures 1-3 This embodiment provides a method for automatically generating personalized care plans based on standardized care levels and health records, including the following steps:

[0027] S1. By extracting multi-source data from patient health records, and using a deep embedding clustering network to group the multi-source data, a basic correspondence between standardized levels and records is obtained.

[0028] Further, step S1 specifically includes:

[0029] S101. Automatically extract structured and unstructured multi-source heterogeneous data from the patient health record system through a standardized data interface, and use a pre-trained deep embedding network to perform feature encoding and fusion of the multi-source heterogeneous data to generate a high-dimensional, unified semantic embedding vector for each patient health record, thereby obtaining a standardized representation.

[0030] The deep embedding network employs a multi-channel deep coding architecture consisting of three parallel coding channels: the first channel is a text coding channel, using a finely tuned BioBERT model on medical corpora to process unstructured disease descriptions and nursing records, extracting the 768-dimensional vector corresponding to the [CLS] tag as text semantic features; the second channel is a temporal coding channel, using a long short-term memory network with two hidden layers (64 units per layer) to process vital sign sequences (such as hourly heart rate, blood pressure, and body temperature data for 72 consecutive hours), using the hidden state (64 dimensions) of the last time step as the temporal dynamic feature; the third channel is a structured data coding channel, processing categorical and numerical data such as age, gender, ADL score, and comorbidity index through one-hot encoding or normalization before inputting it into a three-layer fully connected network (128-64-32), outputting a 32-dimensional dense feature vector. The fusion process is as follows: The feature vectors output from the three channels (including 768-dimensional, 64-dimensional, and 32-dimensional vectors) are concatenated to obtain an 864-dimensional concatenated vector. This concatenated vector is then input into a fusion layer with an attention mechanism, which uses a trainable weight matrix. (864×256 dimensions) The concatenated vector is linearly mapped to a 256-dimensional fusion vector. A softmax layer is used to calculate the weights of the features in each of the three channels, and the fusion vector is then weighted and adjusted. Finally, a 256-dimensional unified semantic embedding vector is generated for each patient file.

[0031] S102. Input the embedding vectors of all files into a deep embedding clustering network for joint training. Optimize the clustering loss function to aggregate similar files in the vector space to form several grouped clusters. Then analyze the clinical characteristics and distribution density of files within each cluster. Combine algorithm recommendation and manual judgment to determine the number of standardized health levels and define the cluster label boundaries corresponding to each level.

[0032] The joint training process of the deep embedding clustering network specifically involves a network structure consisting of an autoencoder and a clustering layer. The encoder part of the autoencoder contains three fully connected layers (256-128-64), which reduce the dimensionality of the input 256-dimensional embedding vector to a 64-dimensional latent feature space. The decoder then symmetrically reconstructs it back to 256 dimensions. Training loss function. It consists of two parts: .in, To reconstruct the loss, mean squared error is used to ensure that the latent space Z retains the key information of the original data; For clustering loss, KL divergence (KL divergence) is specifically used to measure the difference between the soft-assigned distribution Q and the target distribution P. The soft-assigned distribution Q...ij The student t-distribution is used as the kernel function to measure the latent feature point z of the i-th file. i With the j-th cluster center μ j The similarity is calculated; the target distribution P is obtained by sum-of-squares normalization of Q to reinforce high-confidence assignments. The parameter γ is set to 0.1 to balance the two losses. During joint training, the network is pre-trained for 50 epochs using an autoencoder, and then jointly trained for at least 200 epochs using the Adam optimizer (learning rate = 0.001) until the change in clustering loss Lc is less than 1 × 10⁻⁴ / 1 × 10⁻⁴. After training, each file is assigned to its corresponding cluster based on the maximum soft assignment probability output by the clustering layer.

[0033] S103. Based on the mapping relationship between cluster labels and health levels, automatically assign a corresponding standardized health level to each file in each cluster, and calculate the distance from the vector of each file to the center of its cluster. If the distance exceeds the preset threshold, mark it as an abnormal file for review, and obtain the basic correspondence between the standardized level and the file.

[0034] Specifically, the process of calculating distances and marking abnormal files involves calculating the Mahalanobis distance from the embedding vector v of each file to its cluster center μk.

[0035] Σk−1 is the inverse of the covariance matrix of all file embedding vectors within the cluster, and this distance takes into account the correlation between features. A preset threshold is set at the 95th quantile of the Mahalanobis distance distribution within each cluster. For example, for a cluster of patients in the recovery period, if the cluster center's ADL score is 75, and the heart rate variability feature vector is 0.2,−0.1,...0.2,−0.1,..., and a file's Mahalanobis distance exceeds the 95th quantile of the cluster (e.g., 6.8), then the file's feature pattern is considered to deviate significantly from the cluster's typical features. The system will mark it as an abnormal file and automatically transfer it to the manual review queue to avoid incorrect grading due to individual case variations.

[0036] S2. Based on the aforementioned basic correspondence, obtain non-explicit information in the archives, such as previously missed execution records and behavioral trajectories, and use the random forest algorithm to mine potential complication risks and determine the key features of implicit needs.

[0037] Further, step S2 specifically includes:

[0038] S201. Based on the basic correspondence, extract unstructured, non-explicit information from relevant health records, including missing records of medical order execution and daily behavior trajectories; fill in missing values ​​and remove outliers from the non-explicit information to obtain cleaned trajectory sequence records.

[0039] The processing of non-explicit information specifically involves the following steps: For records of missed medical order execution, regular expressions are used to match keywords (such as not executed, refused, missed) in the nursing record text. The matched records are then sorted by timestamp, and the frequency of missed orders and the number of consecutive missed days for various medical orders (such as turning over, medication, and blood glucose testing) over the past 30 days are calculated. For daily behavior trajectory data, if the file contains data from wearable devices or indoor positioning systems, sequences such as the number of steps per hour, activity duration, and number of times the patient gets out of bed at night are extracted over the past 7 days. Missing values ​​are filled using forward imputation, and outliers are removed after comparison with three times the standard deviation of the patient's historical data.

[0040] S202. Construct time-series and statistical feature vectors based on the cleaned trajectory records to form an initial feature set; train the initial feature set using the random forest algorithm, and obtain a sorted list of feature importance by calculating feature importance scores.

[0041] The process of constructing the initial feature set and training the random forest model is as follows: The constructed features include: 1) Medical order execution: frequency of missed turning over, frequency of missed medication, and average duration of nursing response delay during nighttime (22:00-06:00); 2) Behavioral trajectory: coefficient of variation (standard deviation / mean) of activity steps during the daytime (08:00-20:00), number of times the patient gets out of bed at night, and longest continuous sitting time (>90 minutes). This forms an initial set containing approximately 50 features. The random forest model parameters are set as follows: number of decision trees n_estimators=150, maximum tree depth max_depth=12 to prevent overfitting, and node splitting criterion is Gini. The training data consists of 2000 historical patient files from the hospital information system within the past two years, clearly diagnosed by doctors as complications such as diabetic foot and hypostatic pneumonia. The labels indicate whether the patient ultimately developed complications (1 for occurrence, 0 for non-occurrence). During model training, 5-fold cross-validation is used, and the AUC (Area Under Curve) value is used as the model performance evaluation metric to ensure that the AUC reaches above 0.8.

[0042] S203. Features are ranked by calculating their importance scores, including: calculating an importance score based on the reduction of Gini impurity. During the construction of each decision tree in the random forest, the amount of Gini impurity reduction brought about by each feature during node splitting is recorded, and the results of all decision trees are averaged and normalized to obtain a first importance score; calculating an importance score based on out-of-bag error. For each decision tree, the importance is assessed by randomly shuffling the values ​​of specified features in the out-of-bag samples and observing the increase in model prediction error, and the results of all trees are summarized to obtain a second importance score; combining the first and second importance scores, a standardized comprehensive importance score (range 0-1) is generated for each feature, and the features are sorted in descending order according to the comprehensive importance score to form the feature importance list. For example, the comprehensive importance score for the number of times a person gets out of bed at night might be 0.85, while the comprehensive importance score for the frequency of missed turning over in bed might be 0.79.

[0043] S204. Based on the importance of features, select features with a comprehensive importance score greater than 0.7 to form a key feature subset. If the proportion of features in the key feature subset that are related to the pathological mechanisms of known complications (such as falls and pressure injuries) exceeds a preset threshold (such as 60%), then the file is determined to have potential complication risks, and a risk marker is generated. Finally, by combining the risk marker and the key feature subset, the key features of implicit needs associated with the patient's implicit care needs are analyzed and inferred.

[0044] The process of inferring key features of implicit needs is illustrated with an example: Suppose the subset of key features extracted from the file of a diabetic patient living alone includes: frequency of missed turning over (importance 0.79), number of times getting out of bed at night (importance 0.85), and duration of post-meal activity (importance 0.72). Since these features are highly correlated with the pathological mechanisms of pressure injury risk and fall risk, and account for more than 60%, the system generates two risk labels: potential pressure injury risk and potential fall risk. Subsequently, through association rule mining, a combination pattern of missed turning over frequency > 3 times / week and number of times getting out of bed at night > 4 times / night is discovered. The system maps this pattern and infers two specific key features of implicit needs: the need for enhanced nighttime turning over supervision and the need for bedside assistive facilities. Each need is assigned a confidence level between 0 and 1, such as 0.92 for the former and 0.88 for the latter.

[0045] S3. If the key features exceed the preset threshold, the corresponding relationship is updated for the dynamically changing part. By integrating the updated corresponding relationship with the key features, an enhanced indicator file mapping model is obtained.

[0046] Furthermore, step S3 specifically includes:

[0047] S301. Continuously monitor key characteristics of latent needs. By comparing their current values ​​with historical baseline values, calculate the magnitude of change. When the magnitude of change of any key characteristic exceeds the preset dynamic stability threshold for that type of characteristic, it is determined that the patient's state has undergone a meaningful evolution, and the subsequent update process is automatically triggered. Specifically, the method for calculating the magnitude of change is as follows: establish a 30-day sliding time window for each key characteristic, and calculate the moving average of the characteristic value within this window. and standard deviation The magnitude of change in the current value xcurr is quantified using the Z-score, i.e. The dynamic stability threshold is set according to the feature type: for vital sign-related features (such as heart rate variability), the threshold is set to |Z|>2.0; for behavioral habit-related features (such as the number of times a patient gets out of bed at night), the threshold is set to |Z|>2.5. For example, if a patient's 30-day average frequency of missed turning over is 2 times / week with a standard deviation of 0.5, and this frequency suddenly increases to 4 times this week, then Z=(4−2) / 0.5=4.0>2.5, the system determines that the feature has undergone a meaningful change and triggers an update.

[0048] S302. For key features marked as changed, intelligently extract dynamic parameters that are strongly correlated with key features from the basic correspondence between health level and archives. Then, through incremental learning algorithm, incorporate the latest key feature data to recalibrate and optimize the dynamic parameters, and generate an updated standardized correspondence between health level and archives.

[0049] The incremental learning algorithm is specifically implemented as follows: an online passive-attack algorithm is used to update the linear weight vector w in the basic correspondence. The weight vector w1 = (0,…,0) is initialized. When the t-th new sample (i.e., the key feature data that has changed) (xt,yt) arrives, the model predicts its level label as follows: If the prediction is correct, then If the prediction is incorrect, the model update magnitude is minimized by solving the following constrained optimization problem:

[0050] Where ℓ is the hinge loss function. Set the step size. This method can update parameters locally only for the samples that caused the error, and quickly adapt to new changes in the patient's condition without forgetting historical knowledge.

[0051] S303. The updated correspondence obtained is fused and aligned with the patient's complete indicator file in multiple dimensions. The integrated high-quality dataset is used to retrain or fine-tune the original indicator file mapping model to obtain an enhanced indicator file mapping model that reflects the patient's real-time health status and implicit needs. The fusion and alignment process includes: 1) Time alignment: The updated correspondence (reflecting the current status) is associated with the patient's historical file data (such as test reports and vital sign records from the past 3 months) through timestamps. Cubic spline interpolation is used to resample data with different sampling frequencies (such as hourly heart rate and weekly weight) to a unified daily frequency. 2) Feature Space Alignment: The aligned dataset was divided into a training set (80%) and a validation set (20%). Based on the original indicator profile mapping model (a classifier consisting of a three-layer fully connected network with a 128-64-4 structure, outputting four health levels), the parameters of the first two layers were frozen, and only the weights and biases of the last classification layer were fine-tuned over 10 epochs using the Adam optimizer (learning rate = 0.0001). After fine-tuning, an enhanced indicator profile mapping model that more accurately reflects the real-time status of patients was obtained.

[0052] S4. Using the enhanced indicator file mapping model, monitor the file updates for patient condition improvement or new complications, determine whether the plan adjustment mechanism is triggered, and obtain the trigger signal for real-time iteration.

[0053] Further, step S4 specifically includes: S401, using the enhanced indicator file mapping model, performing real-time mapping analysis on the continuously updated patient files, outputting the dynamically changing current indicator file status, extracting and aggregating indicator change data reflecting the improvement of the condition (such as improvement of key indicators) and the occurrence of new complications (such as the emergence of related risk characteristics) based on the current indicator file status, forming a structured set of condition changes.

[0054] S402. For the set of changes in the condition, calculate the comprehensive deviation between the condition and the preset health status baseline in the mapping model, obtain a quantified deviation metric, compare the deviation metric with the clinically set dynamic adjustment threshold, and if it exceeds the threshold, determine that the current change is clinically significant and generate a significant change marker.

[0055] The specific method for calculating the comprehensive deviation is as follows: a weighted Mahalanobis distance is used for measurement. The health status baseline is defined as the cluster center vector of its standardized health level. For an index vector x in the set of disease changes, its weighted Mahalanobis distance .in, W is the inverse of the covariance matrix of the cluster vectors, and W is a diagonal weight matrix with weights on its diagonal. This corresponds to the comprehensive importance score of the i-th feature calculated by the random forest algorithm in step S2. The dynamically adjusted threshold is set to the 90th percentile of the historical deviation values ​​of all patients in this cluster. For example, if the weighted Mahalanobis distance between a patient's current indicator vector and the baseline is 5.2, while the historical 90th percentile of its cluster is 4.1, then the deviation exceeds the threshold, and a significant change marker is generated.

[0056] S403. Match and logically judge the changes in the patient's condition marked with significant changes with the predefined care plan adjustment trigger rule base. If the triggering conditions of a specific adjustment mechanism are met, generate a real-time iterative trigger signal.

[0057] The rule matching process is executed within the Drools rule engine. The fact object is a Java object containing the patient ID, change type (e.g., risk of new-onset hypoglycemia), and deviation metric (e.g., 5.2). When a fact successfully matches a rule condition, the rule engine triggers an action, generating a structured trigger signal object containing specific adjustment instructions.

[0058] S5. Using the trigger signal, extract adjustment parameters from the enhanced indicator file mapping model, use the adaptive moment estimation optimization algorithm to fine-tune the nursing plan parameters, and determine the iterative plan version.

[0059] Furthermore, step S5 specifically includes:

[0060] S501. Upon receiving the generated real-time iterative trigger signal, analyze the health status change type and degree information contained in the trigger signal, and use the analyzed specific change information as an index to intelligently match and extract the corresponding set of nursing intervention dynamic adjustment parameters from the enhanced indicator file mapping model.

[0061] S502. Using the extracted set of dynamically adjustable parameters as the optimization guide and initial update amount, the adaptive moment estimation optimization algorithm is used to automatically and iteratively fine-tune the adjustable parameters in the current nursing plan in multiple rounds. By adaptively adjusting the learning step size, the optimal solution is efficiently found. When the parameter update amount is less than the preset convergence threshold, the iteration stops and the optimized stable parameter set is output.

[0062] The specific process of fine-tuning the parameters using the Adam optimization algorithm is as follows: Define the loss function:

[0063] ,

[0064] Where θ is the adjustable parameter vector in the nursing plan (such as the frequency of turning over f, the advance amount of medication reminder t, etc.), and Constraint(θ) is the clinical rule constraint (such as the frequency of turning over should not exceed 6 times / day). Initialize the first-order moments. Second moment Exponential decay rate Learning rate α = 0.001, stability constant In the t-th iteration:

[0065] a. Calculate the gradient:

[0066] b. Update the partial first-order moment estimate: .

[0067] c. Update the partial second-moment estimate: .

[0068] d. Calculate the first moment after deviation correction: .

[0069] e. Calculate the second moment after deviation correction: .

[0070] f. Update parameters: .

[0071] g. The L2 norm of parameter update When the algorithm is considered to have converged, it stops iterating and outputs the final parameter set. .

[0072] S503. Based on the optimized stable parameter set, automatically update all relevant structured data fields and execution logic in the nursing plan template to generate an iterated personalized care plan version with version identifier (e.g., v2.1.3-20260326), timestamp, and adjustment summary (e.g., due to a significant increase in the number of times people get out of bed at night, the turning frequency is adjusted from q4h to q3h).

[0073] S6. Based on the iterative plan version, obtain the latest implicit requirement features from the multi-source data, determine the matching degree between the features and the plan version, and if the matching degree is lower than a preset threshold, reprocess the latest implicit requirement features to obtain an optimized correspondence update.

[0074] Furthermore, step S6 specifically includes:

[0075] S601. Based on the iterative personalized care plan version, the latest implicit demand features are extracted from the continuous data stream of the patient's health record. The latest implicit demand features are encoded to obtain a feature vector, and the cosine similarity between the feature vector and the current plan version (which is also vectorized) is calculated as a quantitative matching degree value. If the matching degree value is lower than the system's preset adaptability threshold, it is determined that the plan is out of sync with the latest demand, and the optimization process is initiated.

[0076] The specific method for calculating the matching degree is as follows: First, the various nursing measures in the current plan version (such as turning over every 3 hours, daily blood pressure monitoring) are processed through a pre-trained nursing measure-feature association matrix. (n is the number of measures, m is the feature dimension) is mapped to a 512-dimensional plan representation vector. Simultaneously, the latest implicit demand features extracted in step S601 are mapped to the same 512-dimensional demand representation vector through a feature embedding layer. The matching degree is calculated using weighted cosine similarity:

[0077] (This formula specifically means the following in this application: If the matching degree ≥ the preset threshold (75 points) → the plan and requirements match well, maintain the current version; if the matching degree < the preset threshold (75 points) → the plan is determined to be out of sync with the latest requirements, triggering the S602 optimization process.) This is a diagonal weight matrix based on the importance of the features in step S2. The preset fit threshold is 75 points. If the calculated matching score is lower than 75 points, it is considered a mismatch.

[0078] S602. For the latest implicit demand feature set with insufficient matching degree, a hierarchical clustering algorithm is used for secondary analysis and clustering to identify its internal structural pattern.

[0079] The specific parameters and process of the hierarchical clustering are as follows: A bottom-up AGNES algorithm is used, with weighted cosine distance as the similarity measure between features. Initially, each feature is treated as an independent cluster. The algorithm iteratively merges the two closest clusters. and The distance between them is defined as the average link distance:

[0080] .

[0081] The process stops when any of the following conditions are met: the number of current clusters is ≤ 5, or the distance between the two nearest clusters is > 0.4. This method can be used to cluster features such as frequent nighttime urination, increased daytime water intake, and slight weight loss into a single cluster associated with fluid and metabolic abnormalities.

[0082] S603. The adjusted feature clusters generated by clustering are remapped and correlated with the various nursing components in the iterative plan version to construct a more accurate feature-nursing component optimization correspondence. The remapping process involves using a trainable bilinear transformation model to evaluate the correlation strength between feature clusters and nursing components. The centroid vector fc of the feature clusters and the encoding vector of the nursing components are used. Correlation score , where M is a parameter matrix trained using historical successful care case data. The system selects the correlation score. The pairings are used to construct new feature-care component optimization correspondences. For example, the aforementioned fluid and metabolic abnormality feature clusters may generate high correlation scores with care components such as enhanced intake and output recording and adjusted diuretic medication reminders, thus forming new correspondences.

[0083] S7. Through the optimized correspondence update, it is integrated into the enhanced indicator file mapping model to determine the final dynamic mapping structure, which is used to support the continuous adaptive adjustment of the care plan.

[0084] Further, step S7 specifically includes: S701, the obtained feature-nursing component optimization correspondence is used as a key dynamic association rule and injected into the core knowledge base of the current enhanced indicator file mapping model to complete the incremental update of the model logic and the reconstruction of its internal structure.

[0085] S702. Using the updated model, perform remapping calculations on all patient records to generate a global view of the relationship between records, levels, key features, and nursing components. Based on the stability and consistency of the global view, determine and solidify the dynamic mapping structure that accurately depicts complex nursing relationships.

[0086] S703. The solidified dynamic mapping structure is encapsulated as the core adaptive engine of the system. Based on real-time input of file and feature changes, it automatically drives the fine-tuning of the mapping relationship and the updating of the nursing plan, realizing closed-loop and continuous adaptive adjustment of the nursing plan. The closed-loop adjustment workflow is as follows: The engine runs in an event-driven manner, listening for patient data update events in the Kafka message queue. Upon receiving a new data event, the engine executes the following closed-loop steps:

[0087] Perception: Call the deviation calculation module in step S4 to calculate the weighted Mahalanobis distance between the current state and the baseline.

[0088] Decision: If the deviation exceeds the threshold, the Adam optimizer in step S5 is triggered to generate updated power plan parameters under the condition that clinical constraints (such as medication intervals) are met.

[0089] Implementation: Push the updated plan to the nurse workstation or patient app and start tracking implementation.

[0090] Feedback: Collect changes in patient physiological indicators and nurse / patient feedback scores within 72 hours after plan implementation to calculate a new effectiveness evaluation value. This evaluation value will serve as a new training sample, and the mapping model will be fine-tuned using the online passive-attack algorithm in step S3, thus completing a full intelligent closed-loop iteration.

[0091] Example 2

[0092] This embodiment provides another method for automating the generation of personalized care plans based on standardized nursing levels and health records. Through dynamic deduction and simulation verification of health digital twins, it breaks through the limitations of static analysis and passive adjustment in traditional methods, realizes the proactive identification of care needs and the active optimization of plans, and significantly improves the accuracy and feasibility of personalized care plans.

[0093] The details are as follows:

[0094] This research employs a multimodal temporal attention fusion approach to model digital twins of health records. Based on a standardized nursing level classification benchmark, it extracts structured vital sign temporal data, unstructured behavioral trajectory text, medical order execution records, and imaging reports from multi-source heterogeneous data within the patient's health record. A multimodal temporal attention network is used to assign dynamic weights to data across different dimensions. For temporal data (such as continuous 72-hour blood pressure fluctuations and daily activity duration sequences), an attention mechanism is used to focus on features of data from abnormal time periods. For unstructured text data (including descriptions of frequent nighttime urination in nursing records), semantic enhancement encoding is performed using a clinical domain knowledge graph. The encoded features are then input into a digital twin modeling engine to construct a health digital twin that maps the patient's physiological state, behavioral habits, and care response characteristics. This is simultaneously linked to a set of basic care indicators corresponding to the standardized nursing level, forming a three-dimensional record model encompassing the level benchmark, individual characteristics, and dynamic states.

[0095] Digital twin-driven dynamic simulation of implicit needs and risks, based on a 3D file model, utilizes digital twins to conduct multi-scenario simulations, simulating the trend of patient status changes under different care interventions. Simultaneously, a weighted random forest algorithm is employed to fuse the dynamic trend data output by the twin simulation with non-explicit information in the file (such as the frequency of past medication omissions and the coefficient of variation of behavioral trajectories) for training, optimizing the feature importance calculation logic. Higher weights are assigned to temporal features strongly correlated with complications (such as the delay in postprandial blood glucose peaks and the duration of morning joint stiffness), thus identifying key features of implicit needs. By simulating the risk evolution path corresponding to feature changes through twin simulation, such as simulating the correlation between increased medication omission frequency and the risk of sudden blood pressure spikes, risk warning markers with a time dimension are generated, accurately identifying the patient's potential care needs.

[0096] A human-machine collaborative threshold self-optimization mechanism for dynamic priority ranking of needs is constructed, abandoning the fixed threshold judgment logic: using clinical expert experience as the initial input and combining it with historical care plan execution effect data, a threshold optimization model is trained; key features of implicit needs and risk warning markers are input into the model to automatically calculate the priority scores of different needs; the scoring dimensions include the degree of risk urgency, the extent to which the needs are met and the improvement of the patient's condition, and the cost of care resources consumed; for priority conflict scenarios (such as the simultaneous existence of fall prevention needs and medication monitoring needs), the execution effect of different priority care plans is simulated through digital twins, and the optimal ranking result is output to form a structured needs-priority-resource matching list;

[0097] The generation and iteration of care plans through closed-loop simulation validation using a digital twin involves integrating basic interventions at standardized nursing levels based on a priority list of needs to generate an initial personalized care plan. This plan is then input into a health digital twin for closed-loop simulation validation, simulating patient status feedback during plan execution (e.g., whether regular turning intervention reduces pressure ulcer risk, and whether medication reminder intervention improves adherence), and calculating the plan's fit with the patient's actual needs. If the fit does not reach a threshold, parameters such as intervention frequency and execution time in the plan are automatically adjusted based on the feedback data from the digital twin simulation, and the simulation validation is repeated. This iteration continues until the fit reaches the target, ultimately outputting a personalized care plan with a version identifier and simulation validation report. The iterative data is then fed back to the threshold optimization model to achieve continuous self-optimization of the model.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for automatically generating personalized care plans based on standardized nursing levels and health records, characterized by: Includes the following steps: S1. By extracting multi-source data from patient health records, and using a deep embedding clustering network to group the multi-source data, the basic correspondence between standardized levels and records is obtained. S2. Based on the aforementioned basic correspondence, obtain non-explicit information in the archives, such as previously missed execution records and behavioral trajectories, and use the random forest algorithm to mine potential complication risks and determine the key features of implicit needs. S3. If the key features exceed the preset threshold, the corresponding relationship is updated for the dynamically changing part. By integrating the updated corresponding relationship with the key features, an enhanced indicator file mapping model is obtained. S4. Using the enhanced indicator file mapping model, monitor the file updates for patient condition improvement or new complications, determine whether the plan adjustment mechanism is triggered, and obtain the trigger signal for real-time iteration. S5. Using the trigger signal, extract adjustment parameters from the enhanced indicator file mapping model, use the adaptive moment estimation optimization algorithm to fine-tune the nursing plan parameters, and determine the iterative plan version; S6. Based on the iterative plan version, obtain the latest implicit requirement features from the multi-source data, determine the matching degree between the features and the plan version, and if the matching degree is lower than a preset threshold, reprocess the latest implicit requirement features to obtain an optimized correspondence update.

2. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S1 specifically includes: By automatically extracting structured and unstructured multi-source heterogeneous data from the patient health record system through a standardized data interface, and using a pre-trained deep embedding network to perform feature encoding and fusion of the multi-source heterogeneous data, a high-dimensional, unified semantic embedding vector is generated for each patient health record to obtain a standardized representation. The embedding vectors of all files are input into a deep embedding clustering network for joint training. By optimizing the clustering loss function, similar files are grouped together in the vector space to form several clusters. Then, the clinical characteristics and distribution density of files within each cluster are analyzed. Combining algorithm recommendation and manual judgment, the number of standardized health levels is determined, and the cluster label boundaries corresponding to each level are defined. Based on the mapping relationship between cluster labels and health levels, a corresponding standardized health level is automatically assigned to each file within a cluster, and the distance from the vector of each file to the center of its cluster is calculated. If the distance exceeds a preset threshold, the file is marked as an abnormal file for review, thus obtaining the basic correspondence between the standardized level and the file.

3. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S2 specifically includes: S21. Based on the basic correspondence, extract unstructured, non-explicit information from relevant health records, including missing records of medical order execution and daily behavior trajectories; fill in missing values ​​and remove outliers from the non-explicit information to obtain cleaned trajectory sequence records. S22. Construct time-series and statistical feature vectors based on the cleaned trajectory records to form an initial feature set; train the initial feature set using the random forest algorithm, and obtain a sorted list of feature importance by calculating feature importance scores; S23. Select a subset of key features based on the importance of the features; if the proportion of features related to the pathological mechanisms of known complications in the subset of key features exceeds a preset threshold, the file is determined to have potential complication risks and a risk marker is generated; combine the risk marker and the subset of key features to analyze and infer the key features of implicit needs related to the patient's implicit care needs.

4. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 3, characterized in that: By calculating feature importance scores, including: To calculate the importance score based on the reduction of Gini impurity, during the construction of each decision tree in the random forest, the amount of Gini impurity reduction brought about by each feature when splitting at the node is recorded, and the results of all decision trees are summarized, averaged and normalized to obtain the first importance score. Calculate the importance score based on out-of-bag error. For each decision tree, use out-of-bag samples. Randomly shuffle the values ​​of specified features in the out-of-bag samples and observe the increase in model prediction error to assess importance. Summarize the results of all trees to obtain the second importance score. By combining the first importance score and the second importance score, a standardized comprehensive importance score is generated for each feature, and the features are sorted in descending order according to the comprehensive importance score to form the feature importance list.

5. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S3 specifically includes: S31. Continuously monitor key features of latent needs, calculate the magnitude of change by comparing their current values ​​with historical benchmark values, and determine that the patient's condition has undergone a meaningful evolution when the magnitude of change of any key feature exceeds the preset dynamic stability threshold. S32. For key features marked as changing, intelligently extract dynamic parameters that are strongly correlated with key features from the basic correspondence between health level and archives. Then, through incremental learning algorithm, incorporate the latest key feature data to recalibrate and optimize the dynamic parameters, and generate an updated correspondence between standardized levels and archives. S33. The updated correspondence obtained is fused and aligned with the patient's complete indicator file in multiple dimensions. The integrated high-quality dataset is used to retrain or fine-tune the original indicator file mapping model to obtain an enhanced indicator file mapping model that reflects the patient's real-time health status and implicit needs.

6. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S4 specifically includes: Using the enhanced indicator profile mapping model, real-time mapping analysis is performed on continuously updated patient profiles, outputting the dynamic changes in the current indicator profile status. Based on the current indicator profile status, indicator change data reflecting the improvement of the condition and the occurrence of new complications are extracted and aggregated to form a structured set of condition changes. For the set of disease condition changes, the comprehensive deviation between the deviation and the preset health status baseline in the mapping model is calculated to obtain a quantified deviation metric. The deviation metric is compared with the clinically set dynamic adjustment threshold. If it exceeds the threshold, the current change is determined to be clinically significant, and a significant change marker is generated. The changes in patient condition marked with significant changes are matched and logically judged against a predefined base of rules for adjusting care plans. If the triggering conditions of a specific adjustment mechanism are met, a trigger signal is generated for real-time iteration.

7. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S5 specifically includes: Upon receiving the generated real-time iterative trigger signal, the system analyzes the type and degree of health status changes contained in the trigger signal, and uses the analyzed specific change information as an index to intelligently match and extract the corresponding set of dynamic adjustment parameters for nursing intervention from the enhanced indicator file mapping model. Using the extracted set of dynamically adjustable parameters as the optimization guide and initial update amount, the adaptive moment estimation optimization algorithm is used to automatically and iteratively fine-tune the adjustable parameters in the current nursing plan in multiple rounds. By adaptively adjusting the learning step size, the optimal solution is found. When the parameter update amount is less than the preset convergence threshold, the iteration stops and the optimized stable parameter set is output. Based on the optimized and stable parameter set, all relevant structured data fields and execution logic in the care plan template are automatically updated to generate an iterative personalized care plan version with version identifier, timestamp, and adjustment summary.

8. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Step S6 specifically includes: Based on the iterative version of the personalized care plan, the latest implicit needs features are extracted from the continuous data stream of the patient's health record. The latest implicit needs features are encoded to obtain feature vectors, and the cosine similarity between the vectors and the current plan version is calculated as a quantitative matching degree value. If the matching degree value is lower than the system's preset adaptability threshold, it is determined that the plan is out of sync with the latest needs, and the optimization process is initiated. For the latest implicit demand feature set with insufficient matching degree, a hierarchical clustering algorithm is used for secondary analysis and clustering to identify its internal structural pattern; the adjusted feature clusters generated by the clustering are re-mapped and associated with each nursing component in the iterative plan version to build an optimized correspondence between features and nursing components; Based on the feature-nursing component optimization correspondence, components with low matching degree in the original planned version are covered and updated to generate candidate update plans. Then, the matching degree between the candidate update plans and the latest implicit requirement features is calculated to verify the optimization effect. When the verification is successful, the optimized correspondence update is obtained.

9. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 1, characterized in that: Also includes: S7. Through the optimized correspondence update, it is integrated into the enhanced indicator file mapping model to determine the final dynamic mapping structure, which is used to support the continuous adaptive adjustment of the care plan.

10. The method for automatically generating personalized care plans based on standardized nursing levels and health records according to claim 9, characterized in that: Step S7 specifically includes: The obtained feature-nursing component optimization correspondence is used as a key dynamic association rule and injected into the core knowledge base of the current enhanced indicator file mapping model to complete the incremental update of the model logic and the reconstruction of its internal structure. Using the updated model, a remapping calculation is performed on all patient records to generate a global view of the relationship between records, levels, key features, and nursing components. Based on the stability and consistency of the global view, a dynamic mapping structure that accurately depicts complex nursing relationships is determined and solidified. The solidified dynamic mapping structure is encapsulated as the core adaptive engine of the system. Based on real-time input of files and feature changes, it automatically drives the fine-tuning of mapping relationships and updates of nursing plans, realizing closed-loop and continuous adaptive adjustment of nursing plans.

Images