System and method for evaluating treatment compliance of guardian for chronic disease treatment

Abstract

The invention relates to the technical field of intelligent medical assistance, and discloses a system and method for evaluating the therapeutic compliance of a guardian for chronic disease treatment, and the system comprises a data collection module which is used for collecting the multi-dimensional heterogeneous data of the guardian and a child patient from a plurality of information systems of a hospital; the feature engineering module is used for cleaning and standardizing the structured data, performing vectorization and emotion feature extraction on the unstructured text, and generating a uniform feature vector through feature screening; the medical compliance risk assessment model is used for calculating and outputting medical compliance risk levels and key influence factors of the guardians according to the feature vectors; and the dynamic optimization module is used for performing periodic re-evaluation and parameter adjustment on the model according to newly acquired data and medical care feedback during hospitalization. The method corresponds to the system. According to the invention, early, dynamic and explainable quantitative evaluation of medical compliance of the guardian is realized, and intelligent decision support is provided for clinical precise intervention.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical assistance technology, specifically a caregiver compliance assessment system and method for chronic disease treatment. Background Technology

[0002] Treatment adherence in children with chronic diseases is highly dependent on their guardians, and whether guardians follow medical advice is influenced by a complex interplay of social, psychological, economic, cognitive, and behavioral factors. Currently, clinical practice relies primarily on the limited experience of healthcare professionals and one-off communications for assessing guardian adherence, resulting in inherent limitations such as assessment delays, strong subjectivity, and a single-dimensional approach. This makes it difficult to identify high-risk guardians in the early stages of hospitalization, especially during the critical initial admission window; it also hinders dynamic tracking and early warning based on changes in guardian behavior, attitudes, and circumstances during subsequent hospitalizations. At the data level, relevant clinical, behavioral, and communication information is scattered across different systems, preventing effective integration and analysis, thus creating data silos. These issues collectively lead to passive and inefficient clinical interventions, hindering the achievement of precise preventative management. Summary of the Invention

[0003] The purpose of this invention is to provide a caregiver compliance assessment system and method for the treatment of chronic diseases, in order to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention discloses the following technical solutions: In a first aspect, the present invention discloses a guardian's treatment compliance assessment system for the treatment of chronic diseases, the system comprising: The data acquisition module is used to collect multi-dimensional heterogeneous data of guardians from multiple medical information systems in the hospital. The multi-dimensional heterogeneous data includes structured data and unstructured text data. The feature engineering module, connected to the data acquisition module, is used to perform fusion processing on the multi-dimensional heterogeneous data. Specifically, it cleans and standardizes structured data, vectorizes and extracts sentiment features from unstructured text data, and generates a unified feature vector by filtering all features based on their importance. The medical compliance risk assessment model, connected to the feature engineering module, is built based on machine learning algorithms. It is used to calculate the input feature vector and output a risk level representing the guardian's level of medical compliance and key influencing factors used to explain the risk level. The dynamic optimization module connects the data acquisition module and the medical compliance risk assessment model. During the child's hospitalization, it is used to periodically reassess and adjust the parameters of the medical compliance risk assessment model based on newly acquired data and feedback from medical staff on the assessment results.

[0005] Optionally, the feature engineering module performs vectorization and sentiment feature extraction on unstructured text data, including: The BERT pre-trained model is used to encode the doctor-patient communication record text to generate text semantic vectors; and an attention-based sentiment analysis model is used to analyze the doctor-patient communication record text to output quantitative sentiment scores representing positive, neutral or negative attitudes as independent sentiment features.

[0006] Optionally, the feature engineering module generates a feature vector by performing importance filtering on all features, including: The contribution of each feature to the preliminary risk prediction results generated by the medical compliance risk assessment model is calculated using the feature importance analysis method based on SHAP values. The top 15 to 30 features by contribution are selected to form the feature vector used for the final training and inference of the medical compliance risk assessment model.

[0007] Optionally, the construction and training process of the medical compliance risk assessment model includes: The candidate model constructed by the ensemble learning algorithm was trained and its hyperparameters were tuned using the five-fold cross-validation method, with AUC value and F1 score as the main evaluation metrics. The trained model is configured to output a risk probability value between 0 and 1, and a high-risk threshold and a low-risk threshold are set to divide the risk probability value into low-risk, medium-risk, and high-risk levels.

[0008] Optionally, the medical compliance risk assessment model may output key influencing factors in the following ways: The SHAP interpretability algorithm is used to compute the SHAP value corresponding to each feature in the feature vector in parallel when the model performs forward inference on the feature vector. The N features with the highest absolute values ​​of SHAP are output as the key influencing factors, where N is a preset positive integer.

[0009] Optionally, it also includes a report generation module connected to the medical compliance risk assessment model, the report generation module being used for: Receive the risk level, the risk probability value, and the key influencing factors; The key influencing factors are sorted and visualized according to their SHAP values ​​and their positive or negative orientation, generating a guardian's treatment compliance profile report with risk attribution analysis.

[0010] Optionally, the dynamic optimization module performs periodic reassessment and parameter adjustment on the medical compliance risk assessment model, including: Set a fixed assessment cycle. At the beginning of each assessment cycle, use the guardian's behavioral observation scores, communication text data and the child's latest electronic medical record data added in the previous assessment cycle to update the parameters of the medical compliance risk assessment model in an incremental learning manner. Establish a human feedback channel and add treatment interruption events marked by medical staff and subsequently confirmed as strong negative samples to the training set for the next round of model updates.

[0011] Optionally, the dynamic optimization module updates the parameters of the medical compliance risk assessment model using incremental learning, including: For each new sample data point, the predicted risk probability value is determined based on the medical compliance risk assessment model. And the time difference between the time the sample data was collected and the current time. Calculate their joint weight coefficients ,in, , An uncertainty-aware weighting function is used to assign higher weights to samples with high model prediction uncertainty. Let be the time decay function, which is A monotonically decreasing function is used to assign higher weights to more recent samples; The joint weight coefficients of all newly added sample data are normalized to obtain the final learning weight for each sample. Using the newly added sample set weighted according to the final learning weights, the parameters of the medical compliance risk assessment model are fine-tuned through an online learning algorithm.

[0012] Optionally, the structured data collected by the data acquisition module specifically includes: guardian sociodemographic data, disease cognition data obtained through standardized scales, the historical appointment and follow-up visit fulfillment rate of the child, and the guardian's behavioral cooperation score entered daily by the responsible nurse.

[0013] Secondly, the present invention provides a method for assessing the treatment compliance of guardians in the treatment of chronic diseases, comprising: Object tracking steps: Continuously collect facial image streams and voice information of the target patient, and perform time-series analysis on the facial image streams and voice information to obtain the emotional state curve and core complaint symptom set of the target patient; Intervention decision-making steps: Monitor the changing trend of the emotional state curve. When the emotional state curve is identified as exhibiting a preset negative deterioration pattern, generate a decision instruction to switch from a primary education strategy to a secondary reassurance strategy. The secondary reassurance strategy includes at least one in-depth disease guidance content generated based on the core chief complaint symptom set and one high-priority attention shifting content selected based on preset rules. Interactive execution steps: In response to the decision instruction, execute the secondary reassurance strategy, which includes at least: broadcasting the in-depth disease guidance content and playing the high-priority attention diversion content, and adaptively adjusting the tilt angle of the interactive output interface according to the height of the target patient. Link management steps: Record the medical path nodes of the target patient from pre-examination triage to discharge, and when the medical path node indicates discharge, generate and trigger a personalized post-discharge follow-up plan based on the core chief complaint symptom set.

[0014] Beneficial Effects: This invention provides a system and method for assessing patient compliance in chronic disease treatment. Through the systematic collaboration of data collection, feature engineering, a patient compliance risk assessment model, and dynamic optimization, it constructs an intelligent assessment closed loop that enables early identification, dynamic tracking, and interpretable feedback. Specifically, by integrating multi-dimensional heterogeneous data, it achieves preliminary risk assessment based on quantitative characteristics early in the child's hospitalization, solving the problem of early identification. Utilizing the patient compliance risk assessment model and a periodic incremental learning mechanism, it achieves dynamic updates and tracking of the guardian's risk status. By integrating interpretability algorithms, it outputs key influencing factors and visual profiles, making the risk assessment results clinically understandable and directly supporting personalized intervention decisions. The entire technical process forms a closed loop from data collection and intelligent analysis to result feedback and model self-optimization, overcoming the shortcomings of existing technologies such as lagging, one-sided, and static assessments, and providing core technical support for implementing forward-looking and precise clinical management. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural block diagram of a guardian's treatment compliance assessment system for chronic disease treatment provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for assessing the medical compliance of guardians in the treatment of chronic diseases, as provided in an embodiment of the present invention. Detailed Implementation

[0017] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0018] Treatment of children with chronic diseases (such as asthma, diabetes, and epilepsy) is a long-term, continuous process, and its success largely depends on the guardian's (usually parents) understanding, execution, and adherence to medical advice—that is, treatment compliance. Clinical practice shows that insufficient guardian compliance is one of the key factors leading to treatment interruption, disease recurrence, and even poor prognosis in children. Therefore, early identification of guardians with compliance risks during hospitalization and implementation of targeted interventions are of great significance for improving the long-term health outcomes of children with chronic diseases.

[0019] However, the current clinical setting for assessing caregiver compliance suffers from the following systemic deficiencies: First, the assessment methods are subjective and fragmented. Currently, they mainly rely on the personal experience and intuitive feelings of medical staff during limited contact, or are supplemented by simple questionnaires. This method lacks unified and quantitative standards, the assessment results are easily influenced by the subjective judgment of medical staff, and often only reflect one-sided information at a single point in time, failing to form objective and consistent assessment conclusions.

[0020] Secondly, the assessment dimensions are too simplistic, neglecting complexity. A guardian's compliance behavior is not an isolated phenomenon, but rather deeply influenced by a complex interplay of factors, including their social background (such as education level and economic status), psychological cognition (understanding of the disease and attitude towards treatment), behavioral patterns (communication style and level of cooperation), and past medical interactions. Existing methods cannot systematically collect and integrate this heterogeneous data, resulting in superficial assessments that fail to address the root causes affecting compliance and lead to low predictive accuracy.

[0021] Third, identification is delayed and lacks dynamic early warning capabilities. Traditional assessments often occur after problems have emerged (such as missed medication), failing to proactively identify high-risk guardians early in the child's hospitalization (e.g., within the critical 72 hours), thus missing the optimal window for early intervention. More importantly, throughout the hospitalization period, there is a lack of a mechanism for continuous, automated tracking and risk reassessment of guardians' status, making it impossible to dynamically respond to changes in their attitudes or behaviors, and resulting in a lack of early warning mechanisms.

[0022] Fourth, the problem of data silos is severe, hindering intelligent analysis. Information related to guardians is scattered across various hospital information systems (such as Electronic Medical Records (EMR), Hospital Information System (HIS), and nursing record systems), and includes a large amount of unstructured text records (such as communication minutes). This fragmented data cannot be effectively correlated, integrated, or used for in-depth analysis, making it impossible to build a foundation for data-driven intelligent assessment models.

[0023] Finally, the assessment results are opaque and difficult to guide clinical action. Even when simple statistical models are introduced, their predictions often lack interpretability. Clinicians cannot understand why the model makes a certain risk judgment, nor are they clear about the key risk factors, leading to low trust in the assessment results and making it difficult to develop precise and personalized intervention strategies based on them.

[0024] In summary, existing technologies fail to provide a systematic solution that can be initiated early in hospitalization, integrate multi-dimensional data, achieve dynamic tracking, and provide interpretable results. The embodiments of this application are proposed precisely to fundamentally overcome these deficiencies.

[0025] The technical solutions of the embodiments of the present invention described below are all committed to solving one or more of the above-mentioned technical problems in a collaborative manner from different dimensions. Their core concept is to realize full-cycle, personalized service support for service recipients from arrival and waiting to departure through a technical system that integrates dynamic perception, intelligent decision-making, multimodal interaction and full-link management.

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. Secondly, in this document, the term "comprising" is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0027] Firstly, this embodiment provides a caregiver compliance assessment system for chronic disease treatment, such as... Figure 1 As shown, the system includes: The data acquisition module is used to collect multi-dimensional heterogeneous data of guardians from multiple medical information systems in the hospital. The multi-dimensional heterogeneous data includes structured data and unstructured text data. The feature engineering module, connected to the data acquisition module, is used to fuse multi-dimensional heterogeneous data. It cleans and standardizes structured data, vectorizes and extracts sentiment features from unstructured text data, and generates a unified feature vector by filtering all features based on their importance. The medical compliance risk assessment model, connected to the feature engineering module, is built based on machine learning algorithms. It is used to calculate the input feature vector and output the risk level representing the guardian's level of medical compliance and the key influencing factors used to explain the risk level. The dynamic optimization module connects the data acquisition module and the medical compliance risk assessment model. During the child's hospitalization, it is used to periodically reassess and adjust the parameters of the medical compliance risk assessment model based on newly acquired data and feedback from medical staff on the assessment results.

[0028] The data acquisition module interfaces with the hospital's information system, electronic medical record system, nursing information system, and laboratory information system through predefined application programming interfaces (APIs). In actual deployment, data exchange can be achieved through an intermediate database, the HL7 protocol, or the FHIR standard. This module extracts data from these systems periodically (e.g., hourly) or triggered on a case-by-case basis. Structured data includes, but is not limited to, patient diagnoses and hospitalization counts from electronic medical records, basic guardian information from the HIS, and daily behavioral score records from the nursing system. Unstructured text data mainly refers to summaries of doctor-patient communication and relevant descriptions in medical records filled out by medical staff in electronic medical records or dedicated communication record systems. The data acquisition module associates the extracted raw data with a unified patient ID and guardian ID and stores it in the system's dedicated database, forming an initial data pool.

[0029] The feature engineering module processes the raw data from the data acquisition module. For structured data, cleaning operations include handling missing values, correcting obvious errors (such as negative age), and standardizing units of measurement. For example, missing values ​​in the household annual income field are filled using the median of the same residential area and similar occupational groups. Standardization operations use the Z-score standardization method to transform continuous variables (such as age and income) into a distribution with a mean of 0 and a standard deviation of 1. The feature engineering module ultimately outputs a feature vector with fixed dimensions and normalized values ​​for use by the model.

[0030] The medical compliance risk assessment model is a classification model built on machine learning algorithms. In one embodiment, the model is implemented using a gradient boosting decision tree algorithm, such as XGBoost or LightGBM. The model receives feature vectors generated by the feature engineering module as input. Internally, it contains multiple base learners (decision trees), and through ensemble learning, it combines the prediction results of all base learners to finally output a risk probability value between 0 and 1, which represents the likelihood that the guardian belongs to the "low medical compliance" group.

[0031] The dynamic optimization module is responsible for maintaining the timeliness and accuracy of the model. This module sets a fixed evaluation and update cycle (e.g., 24 hours). At the beginning of each cycle, the module checks if the data acquisition module has collected new data. If so, it triggers the model update process. The update not only utilizes the newly added data but also incorporates feedback from healthcare professionals. For example, if the system previously assessed a caregiver as low-risk, but subsequently the caregiver experienced a confirmed treatment interruption event, healthcare professionals can mark this event on the system interface. The dynamic optimization module will add this event and its corresponding historical feature data as a strong negative sample labeled "high-risk" to the model's retraining dataset.

[0032] Based on the above, the system works as follows: Upon admission of a child, the data acquisition module immediately begins collecting data from various hospital systems. The feature engineering module quickly processes this data and generates feature vectors. The compliance risk assessment model calculates based on these feature vectors, outputting the initial risk assessment results within a short period (e.g., within 24-72 hours of admission), achieving early identification. Subsequently, the dynamic optimization module drives the system into a loop: new data and human feedback continuously flow in, the model is updated regularly, and the risk status is dynamically tracked. The assessment results (risk level and key influencing factors) are visualized through the report generation module, providing interpretable decision support for medical staff. Through the collaborative operation of each module, the system forms an intelligent closed loop of "data acquisition-analysis-output-feedback-optimization," effectively overcoming the shortcomings of traditional methods in terms of timeliness, objectivity, dynamism, and interpretability.

[0033] Furthermore, it should be noted that the guardian's treatment compliance assessment system for chronic disease treatment described in this application strictly adheres to national laws and regulations on personal information protection and medical data security management standards in its entire data collection, processing, and usage process, ensuring that it is conducted under the premise of legal authorization and privacy protection. The specific compliance design of the system is reflected in the following key aspects: 1. Informed Consent and Authorization for Data Collection: Upon admission of the child, the hospital will provide the guardian with a specific informed consent form, clearly explaining that to improve medical quality and safety, an anonymous compliance risk assessment study will be conducted based on their relevant information, and obtaining their written informed consent. The data collection module may only retrieve data related to the guardian and the child from the hospital information system after obtaining this valid authorization. The use of all data is strictly limited to the compliance risk assessment and clinical decision support described in this system.

[0034] 2. Data De-identification Processing: After extracting data from the source system, the data acquisition module immediately performs de-identification. Specifically, the system automatically replaces direct identifiers (such as name, ID number, contact information, and detailed address) with irreversible, unique internal system codes (such as GUIDs). Date information retains only the relative time or undergoes offset processing. After this process, the data processed by the subsequent feature engineering module and model is de-identified data and cannot be directly associated with a specific individual.

[0035] 3. Minimal Necessity and Secure Storage: The system adheres to the principle of minimum necessity, collecting only data dimensions directly related to medical compliance risk assessment. All data, including raw data and intermediate features, is stored in an encrypted state on a controlled secure server within the hospital's intranet or in a certified private cloud environment, with strict role management and log auditing for access permissions.

[0036] 4. Full Lifecycle Compliance Management: The system has a data retention and destruction policy. When the assessment objective is achieved (e.g., the child is discharged and follow-up is completed) or the guardian withdraws consent, the associated data will automatically trigger a secure deletion process according to the prescribed period. Model training and optimization are all conducted on de-identified data to ensure that no personal privacy is leaked.

[0037] Through the aforementioned technical and management measures embedded in each layer of the system architecture, this system, while realizing intelligent assessment functions, ensures the legality, legitimacy, and necessity of data processing activities, fully protects the personal information rights of guardians and sick children, and complies with the ethical and legal requirements for the application of medical and health data.

[0038] To address the problem that unstructured text data (such as communication records) is difficult to quantify and utilize using traditional assessment methods, resulting in the loss of a large amount of key information reflecting the guardian's attitude and cognition, as an optional implementation method in this embodiment, the feature engineering module performs vectorization and sentiment feature extraction on the unstructured text data, including: The BERT pre-trained model was used to encode the doctor-patient communication records and generate text semantic vectors. Then, an attention-based sentiment analysis model was used to analyze the doctor-patient communication records and output quantitative sentiment scores representing positive, neutral or negative attitudes as independent sentiment features.

[0039] In practical implementation, for a doctor-patient communication record text, preprocessing is first performed, including removing irrelevant characters and word segmentation. Then, it is input into the BERT pre-trained model. This embodiment can use the BERT-base-chinese model, which, through its deep Transformer encoder, can understand the specific meaning of Chinese words in context. The output vector corresponding to the [CLS] marker of the last layer of the BERT model is taken as the semantic vector representation of the entire text. This vector is a 768-dimensional floating-point array containing deep semantic information of the text. In practical applications, other pre-trained language models, such as RoBERTa and ERNIE, can also be selected; this embodiment does not limit this choice.

[0040] Simultaneously, the same text was fed into an attention-based sentiment analysis model. This model can be a sentiment classification model specifically fine-tuned on medical communication corpora. The model ultimately outputs a sentiment score between -1 and 1, where a score greater than 0.1 is considered a positive attitude, less than -0.1 is considered a negative attitude, and a score in between is considered a neutral attitude. This sentiment score, as an independent scalar feature, participates in subsequent feature fusion along with the semantic vector obtained from the BERT model. In this way, the system not only captures what the guardian "said" (semantics) but also quantifies the sentiment tendency of "how they said it," providing richer dimensions for risk assessment.

[0041] Based on the above, this processing method enables the system to transform subjective, qualitative communication text into computable quantitative features, solving the problem of utilizing unstructured data. Semantic vectors preserve complex contextual information, while sentiment scores provide intuitive attitude indicators. The combination of the two enhances the ability to characterize the psychological and cognitive state of guardians, thereby improving the predictive accuracy and interpretability of the overall assessment model.

[0042] To address the issue that the initial feature set generated from multi-dimensional fusion has high dimensionality and may contain redundant or noisy features, directly affecting model efficiency, generalization ability, and interpretability, as an optional implementation method in this embodiment, the feature engineering module generates feature vectors by filtering the importance of all features, including: The contribution of each feature to the preliminary risk prediction results generated by the medical compliance risk assessment model was calculated using the feature importance analysis method based on SHAP values. The top 15 to 30 features by contribution were selected to form the feature vector for the final training and inference of the medical compliance risk assessment model.

[0043] In practice, the model is first trained on a benchmark medical compliance risk assessment model (such as a random forest model) using all initial features (potentially more than 100, including structured features, principal components of text semantic vectors, sentiment scores, etc.) to obtain the model's predictions for a batch of validation samples. Then, the SHAP interpretation framework is used to analyze this benchmark model. For each feature, its SHAP value is calculated, which quantitatively describes the degree to which the presence of that feature deviates the model's predicted output from the baseline (the average of all features). The global importance score for each feature is obtained by summing the absolute values ​​of all feature SHAP values ​​for all samples.

[0044] After ranking by importance, the system selects the top 15 to 30 features based on their importance scores. This range is chosen to balance model performance and complexity: fewer than 15 features may result in the loss of crucial information, while more than 30 features may introduce excessive noise and increase computational burden. These selected core features constitute the final feature vector. For example, the final feature vector might include the most predictive indicators such as "total score on the disease awareness scale," "follow-up visit fulfillment rate over the past year," "average communication and emotional score over the past three days," and "family out-of-pocket payment ratio."

[0045] Based on the above, this SHAP-based screening method not only achieves feature dimensionality reduction and improves model training and inference efficiency, but more importantly, it filters features according to their true contribution to the prediction results, ensuring that each item in the final feature vector is highly correlated with medical compliance risk. This lays a reliable foundation for high-precision predictions in subsequent models and the generation of interpretable key impact factor reports.

[0046] To ensure the reliability and practicality of the medical compliance risk assessment model, avoid overfitting, and intuitively link its output to clinical intervention actions, as an optional implementation method in this embodiment, the construction and training process of the medical compliance risk assessment model includes: The candidate model constructed by the ensemble learning algorithm was trained and its hyperparameters were tuned using the five-fold cross-validation method, with AUC value and F1 score as the main evaluation metrics. The trained model is configured to output a risk probability value between 0 and 1, and high-risk and low-risk thresholds are set to divide the risk probability value into low-risk, medium-risk, and high-risk levels.

[0047] In the specific implementation, historical data processed by the feature engineering module is used as the training set. A five-fold cross-validation strategy is adopted: the data is randomly divided into 5 parts, with 4 parts used as the training subset and the remaining part as the validation subset, repeated 5 times. In each training iteration, the selected ensemble learning algorithm (such as LightGBM) is subjected to hyperparameter tuning. The parameters for grid search may include the maximum tree depth, learning rate, minimum number of samples per leaf node, etc. Model performance is evaluated using the AUC value and F1 score on the validation subset as the core indicators. The AUC value measures the overall ranking ability of the model, while the F1 score combines the precision and recall of the model for the "high-risk" category. Finally, the model parameter combination with the highest average AUC value and F1 score in the 5 cross-validations is selected, and the model is retrained on all training data using these parameters to obtain the final medical compliance risk assessment model.

[0048] When deploying the model, configure it to output a continuous risk probability value. ,in To convert continuous probabilities into discrete risk levels that are easy to implement clinically, two thresholds are set: a low-risk threshold and a low-risk threshold. (e.g., 0.3) and high-risk threshold (For example, 0.7). The division rules are as follows: If If it is, it is judged as a low-risk level; if If so, it is classified as a medium-risk level; if If the threshold is too high, it is classified as a high-risk level. The threshold can be adjusted according to actual clinical needs; for example, it can be increased when medical resources are strained. To focus on the cases with the highest risk.

[0049] Based on the above, this training and configuration method ensures the model's generalization ability and stability through cross-validation and robust evaluation metrics. The three-level risk classification mechanism simplifies the complex model output into clear and actionable clinical guidelines, enabling healthcare professionals to quickly understand the risk level and take corresponding strategies such as routine management, enhanced education, or targeted interventions, achieving seamless integration between intelligent assessment and clinical practice.

[0050] To address the issue that machine learning models, being black boxes, have unknowable decision-making processes, leading to low trust in assessment results among clinical medical staff and an inability to provide targeted interventions, as an optional implementation method in this embodiment, the medical compliance risk assessment model outputs key influencing factors in the following ways: The SHAP interpretability algorithm is used to compute the SHAP value of each feature in the feature vector in parallel when the model performs forward inference on the feature vector. The N features with the highest absolute values ​​of SHAP are output as key influence factors, where N is a preset positive integer.

[0051] In practical implementation, when the model is given an input feature vector Perform forward propagation and calculate the risk probability. Simultaneously, the interpretability module is activated. This module computes the current sample based on either KernelExplainer in SHAP or TreeExplainer suitable for tree models. Each feature SHAP value SHAP value Representation of features The probability of the final prediction relative to the baseline value (such as the average value of the entire dataset) for this feature. The contribution of the feature. A positive value indicates that the feature increases the probability of risk, while a negative value indicates that it decreases the probability of risk.

[0052] After the calculation is complete, the system takes the absolute value of all feature SHAP values. Then sort them in descending order. Select the N features with the largest absolute values, where N is a preset value, for example, 5. These N features and their corresponding SHAP values ​​are then analyzed. (Including symbols) are collectively output as key influencing factors. For example, for a risk assessment, the output might be: Key Influencing Factor 1: {Feature: "Average Communication Emotional Score over the Past Three Days", SHAP value: +0.15}; Key Influencing Factor 2: {Feature: "Percentage of Out-of-Household Expenses", SHAP value: +0.12}; Key Influencing Factor 3: {Feature: "Total Score of Disease Cognition Scale", SHAP value: -0.08}.

[0053] Based on the above, this technical solution includes an explanation with each risk assessment. By identifying key influencing factors, healthcare professionals can intuitively understand which specific factors (such as negative communication emotions or economic stress) primarily led to a high-risk assessment, or which protective factors (such as a high level of disease awareness) supported a low-risk assessment. This enhances the transparency and credibility of the assessment results and provides a direct and clear focus for developing personalized communication or intervention plans.

[0054] To transform the generated machine-readable data, such as key impact factors, into decision support tools that are easy for clinical teams to understand and use, as a further optional implementation of this embodiment, the system also includes a report generation module connected to the medical compliance risk assessment model. The report generation module is used for: Receive risk level, risk probability value and key influencing factors; The key influencing factors are sorted and visualized according to their SHAP values ​​and their positive or negative orientation, generating a guardian's treatment compliance profile report with risk attribution analysis.

[0055] In its implementation, the report generation module receives an initial assessment output from the compliance risk assessment model: risk level (e.g., "high risk"), risk probability value (e.g., 0.82), and a list of key influencing factors. This module has a built-in report template and visualization engine. First, it generates a concise report header displaying the child's basic information, assessment time, and final risk level. Then, the core part is the risk attribution analysis (SHAP) graph. For example, it generates a horizontal bar chart where each bar represents a key influencing factor, the bar's length corresponds to its absolute SHAP value, and the bar's direction (left for negative / blue, right for positive / red) and color correspond to the sign of the SHAP value. Each bar is labeled with its feature name and specific SHAP value.

[0056] In addition to visual charts, the report also provides summaries of the charts in natural language, such as: "This assessment is high-risk. The main risk drivers are: the guardian's recent negative communication attitude (contribution +0.15) and high out-of-pocket expenses from the family (contribution +0.12). The protective factor is: the guardian's relatively good understanding of the disease (contribution -0.08)." The final output is a graphic-rich "Guardian's Treatment Compliance Profile Report," which can be presented in PDF format or directly on the system interface.

[0057] Based on the above, the report generation module transforms the abstract model output into an intuitive and professional clinical report. By ranking and visualizing key influencing factors, it achieves clear attribution of risk root causes, making complex multifactorial assessment results readily understandable. This profile report becomes a core basis for communication between medical staff and caregivers, for multidisciplinary team discussions, and for developing precise intervention plans, demonstrating the value of data intelligence in assisting clinical decision-making.

[0058] To address the issue of static models failing to adapt to dynamic changes in the guardian's state, as an optional implementation method in this embodiment, a dynamic optimization module periodically re-evaluates and adjusts the parameters of the medical compliance risk assessment model, including: Set a fixed assessment cycle. At the beginning of each assessment cycle, use the newly added guardian behavior observation scores, communication text data and the latest electronic medical record data of the child in the previous assessment cycle to update the parameters of the medical compliance risk assessment model in an incremental learning manner. Establish a human feedback channel and add treatment interruption events marked by medical staff and subsequently confirmed as strong negative samples to the training set for the next round of model updates.

[0059] In its implementation, the dynamic optimization module incorporates a timer that fixes the evaluation cycle at 24 hours. The module automatically starts every morning at midnight. It first queries the data acquisition module for newly generated or updated data from the past 24 hours; this data constitutes the dataset for the current round of incremental updates. Simultaneously, the module checks whether it has received a "feedback confirmation" signal submitted by medical staff through the system interface. For example, if a guardian is assessed as low-risk by the system but actually fails to take medication as prescribed, the nurse can mark "actual treatment interruption" in the guardian's file and submit it. The module will extract the historical feature data corresponding to these marked events and assign them a "high-risk" label, forming a strong negative sample set. .

[0060] The module then performs a model update. This will add the new dataset. and strong negative sample set The merged data will serve as training data for this round of incremental learning. Once updated, the new version of the model will be used for risk assessment in the next 24-hour period.

[0061] Based on the above, this approach allows the system to transcend the role of a static analysis tool. Regular incremental learning ensures that the model's knowledge base is continuously updated with the accumulation of new data, adapting to potential trends in group behavior. More importantly, the human feedback channel enables a reinforcement learning cycle of "practice verification - model correction." The system can quickly learn from its own prediction errors (such as misjudging actual non-compliance individuals as low-risk), thereby continuously optimizing its prediction accuracy and achieving increasingly precise dynamic tracking of the risk status of guardians.

[0062] Building upon the aforementioned periodic reassessment and parameter adjustment, to address the uneven value of newly added samples (some containing abundant information, while others may be noise) and the problem of how to efficiently utilize limited new data to optimize the model, as a further optional implementation method in this embodiment, the dynamic optimization module updates the parameters of the medical compliance risk assessment model through incremental learning, including: For each new sample data point, the predicted risk probability value is determined based on the medical compliance risk assessment model. And the time difference between the time the sample data was collected and the current time. Calculate their joint weight coefficients ,in, , For uncertainty-aware weighting functions, its... When the maximum value is obtained, The value decreases as it approaches 0 or 1, and is used to assign higher weights to samples with high uncertainty in model prediction. Let be the time decay function, which is A monotonically decreasing function is used to assign higher weights to more recent samples; The joint weight coefficients of all newly added sample data are normalized to obtain the final learning weight for each sample. Using the newly added sample set weighted according to the final learning weights, the parameters of the medical compliance risk assessment model are fine-tuned through an online learning algorithm.

[0063] In the specific implementation, for the first unprocessed step in this round... For each new sample, the risk probability is first calculated using the current model. Define the uncertainty-aware weighting function. for: , This function is a parabola that opens downwards. When it reaches its maximum value of 1, When the probability approaches 0 or 1, the function value approaches 0. This indicates that for samples where the model is very confident in its prediction (probability close to 0 or 1), the information content may be small, and the weight should be reduced; while for samples where the model is not very confident in its prediction (probability close to 0.5), they may represent fuzzy cases or new patterns on the model's decision boundary, containing more information, and therefore should be given higher weight.

[0064] At the same time, define the time decay function. for: , in, Specifically, it refers to the number of hours from the time the sample was collected to the current time. It is the attenuation rate hyperparameter (e.g.) This function ensures that more recent behavioral data contributes more to reflecting the guardian's current state and has a higher weight.

[0065] Calculate joint weights Then, for all newly added samples Normalize: , making This yields the final learning weights for each sample. .

[0066] During the model fine-tuning phase, an online gradient descent algorithm is employed. When calculating the loss function (such as log loss), the loss term for each sample is multiplied by its weight. This means that high-uncertainty, recent samples will have a greater gradient impact in this parameter update, causing the model to adjust in a direction that better fits these "valuable" new samples.

[0067] Based on the above, this weighted strategy guides the model to prioritize learning data that most challenges its existing knowledge (high uncertainty) and best represents the latest dynamics (time decay), improving the efficiency and relevance of incremental learning. This enables the system to more sensitively capture subtle changes and early signals of guardians' medical compliance risks, achieving more accurate and adaptive dynamic risk assessment.

[0068] To clarify the core data dimensions upon which the system relies and to ensure the comprehensiveness and operability of the assessment, as an optional implementation method in this embodiment, the structured data collected by the data acquisition module specifically includes: guardian sociodemographic data, disease cognition data obtained through standardized scales, the historical appointment and follow-up visit fulfillment rate of the child, and the guardian's behavioral cooperation score entered daily by the responsible nurse.

[0069] In the specific implementation: (1) Social demographic data of guardians: obtained through the hospital information system, including the relationship between the guardian and the child (parents / grandparents, etc.), age, education level, occupation, type of permanent residence, family annual income range, medical insurance type, etc. This information is recorded objectively and combined with other dimensions of data to construct a comprehensive profile of the guardian. The acquisition and processing of this information follow privacy protection standards (such as de-identification).

[0070] (2) Disease cognition data obtained through standardized scales: In the early stages of a child's hospitalization, nurses guide guardians to complete standardized scales using electronic devices provided by the hospital. The scales used are mature tools that have undergone reliability and validity testing, such as knowledge questionnaires for specific chronic diseases or general health literacy scales. The system automatically receives and calculates the scale scores, using them as a key quantitative feature input.

[0071] (3) Historical appointment and follow-up visit fulfillment rate of children: The past appointment and visit records of children are extracted from the follow-up management module of the electronic medical record system. By calculating the ratio of the number of appointments for follow-up visits within a preset time window (such as the past 12 months) to the number of follow-up visits actually completed, an objective quantitative indicator of historical behavioral compliance is obtained.

[0072] (4) Daily entry of guardians' behavior cooperation scores by the responsible nurse: In daily nursing work, the responsible nurse uses the nursing terminal integrated with this system to score each guardian's behavior cooperation on the same day based on a unified and clear behavior observation checklist (such as whether they assist in administering medication on time, whether they actively cooperate with examination procedures, whether they proactively ask questions and provide feedback, etc.). This score provides a continuous and direct record of behavior observation.

[0073] Based on the above, these structured data items are information objectively generated from the hospital's existing information systems and clinical workflows, and are closely related to the caregiver's behavior. Together with unstructured communication text data, they serve as the raw input for the feature engineering module, jointly contributing to subsequent fusion analysis and risk assessment, aiming to provide quantifiable decision-making basis for the model from multiple perspectives.

[0074] Secondly, this embodiment provides a method for assessing the patient's compliance with treatment in a guardianship system for chronic disease treatment, as described above (i.e., a method for assessing patient compliance with treatment in a guardianship system for chronic disease treatment), such as... Figure 2 As shown, the method includes the following steps in sequence: S1 - Collects multi-dimensional heterogeneous data of guardians from multiple medical information systems in the hospital. The multi-dimensional heterogeneous data includes structured data and unstructured text data. S2- Fusion processing of multi-dimensional heterogeneous data, including cleaning and standardizing structured data, vectorizing and extracting sentiment features from unstructured text data, and generating a unified feature vector by filtering all features based on their importance. S3 - Through a medical compliance risk assessment model built on machine learning algorithms, the feature vector is calculated, and the risk level representing the guardian's level of medical compliance and the key influencing factors used to explain the risk level are output. S4 - During the child's hospitalization, the compliance risk assessment model is periodically reassessed and its parameters are adjusted based on newly collected data and feedback from medical staff on the assessment results.

[0075] It should be noted that this method corresponds to the aforementioned guardian treatment compliance assessment system for chronic disease treatment. Therefore, the parts of this method that are not specifically described (including but not limited to specific technical means and effects) can be referred to the relevant descriptions in the aforementioned guardian treatment compliance assessment system for chronic disease treatment, and will not be elaborated here.

[0076] In the embodiments provided by this invention, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any suitable combination thereof. For hardware implementation, the processor can be implemented in one or more of the following: application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to implement the functions described herein, or combinations thereof. For software implementation, some or all of the processes of the embodiments can be performed by a computer program instructing the associated hardware. During implementation, the program can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. Computer-readable storage media can include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer.

[0077] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A guardian's treatment compliance assessment system for chronic disease treatment, characterized in that, The system includes: The data acquisition module is used to collect multi-dimensional heterogeneous data of guardians from multiple medical information systems in the hospital. The multi-dimensional heterogeneous data includes structured data and unstructured text data. The feature engineering module, connected to the data acquisition module, is used to perform fusion processing on the multi-dimensional heterogeneous data. Specifically, it cleans and standardizes structured data, vectorizes and extracts sentiment features from unstructured text data, and generates a unified feature vector by filtering all features based on their importance. The medical compliance risk assessment model, connected to the feature engineering module, is built based on machine learning algorithms. It is used to calculate the input feature vector and output a risk level representing the guardian's level of medical compliance and key influencing factors used to explain the risk level. The dynamic optimization module connects the data acquisition module and the medical compliance risk assessment model. During the child's hospitalization, it is used to periodically reassess and adjust the parameters of the medical compliance risk assessment model based on newly acquired data and feedback from medical staff on the assessment results.

2. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The feature engineering module performs vectorization and sentiment feature extraction on unstructured text data, including: The BERT pre-trained model is used to encode the doctor-patient communication record text to generate text semantic vectors; and an attention-based sentiment analysis model is used to analyze the doctor-patient communication record text to output quantitative sentiment scores representing positive, neutral or negative attitudes as independent sentiment features.

3. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The feature engineering module generates feature vectors by filtering all features based on their importance, including: The contribution of each feature to the preliminary risk prediction results generated by the medical compliance risk assessment model is calculated using a feature importance analysis method based on SHAP values. The top 15 to 30 features by contribution are selected to form the feature vector used for the final training and inference of the medical compliance risk assessment model.

4. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The construction and training process of the medical compliance risk assessment model includes: The candidate model constructed by the ensemble learning algorithm was trained and its hyperparameters were tuned using the five-fold cross-validation method, with AUC value and F1 score as the main evaluation metrics. The trained model is configured to output a risk probability value between 0 and 1, and a high-risk threshold and a low-risk threshold are set to divide the risk probability value into low-risk, medium-risk, and high-risk levels.

5. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The medical compliance risk assessment model outputs key influencing factors in the following ways: The SHAP interpretability algorithm is used to compute the SHAP value corresponding to each feature in the feature vector in parallel when the model performs forward inference on the feature vector. The N features with the highest absolute values ​​of SHAP are output as the key influencing factors, where N is a preset positive integer.

6. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 5, characterized in that, It also includes a report generation module that connects to the medical compliance risk assessment model, the report generation module being used for: Receive the risk level, the risk probability value, and the key influencing factors; The key influencing factors are sorted and visualized according to their SHAP values ​​and their positive or negative orientation, generating a guardian's treatment compliance profile report with risk attribution analysis.

7. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The dynamic optimization module performs periodic reassessment and parameter adjustment on the medical compliance risk assessment model, including: Set a fixed assessment cycle. At the beginning of each assessment cycle, use the guardian's behavioral observation scores, communication text data and the child's latest electronic medical record data added in the previous assessment cycle to update the parameters of the medical compliance risk assessment model in an incremental learning manner. Establish a human feedback channel and add treatment interruption events marked by medical staff and subsequently confirmed as strong negative samples to the training set for the next round of model updates.

8. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 7, characterized in that, The dynamic optimization module updates the parameters of the medical compliance risk assessment model using incremental learning, including: For each new sample data point, the predicted risk probability value is determined based on the medical compliance risk assessment model. And the time difference between the time the sample data was collected and the current time. Calculate their joint weight coefficients ,in, , Weighting function for uncertainty perception; Let be the time decay function, which is A monotonically decreasing function is used to assign higher weights to more recent samples; The joint weight coefficients of all newly added sample data are normalized to obtain the final learning weight for each sample. Using the newly added sample set weighted according to the final learning weights, the parameters of the medical compliance risk assessment model are fine-tuned through an online learning algorithm.

9. The guardian's treatment compliance assessment system for chronic disease treatment according to claim 1, characterized in that, The structured data collected by the data acquisition module specifically includes: guardian sociodemographic data, disease cognition data obtained through standardized scales, the historical appointment and follow-up visit fulfillment rate of the child, and the guardian's behavioral cooperation score entered daily by the responsible nurse.

10. A method for assessing the treatment compliance of guardians in the treatment of chronic diseases, characterized in that, The method includes the following steps: Multidimensional heterogeneous data of guardians are collected from multiple medical information systems of the hospital. The multidimensional heterogeneous data includes structured data and unstructured text data. The multi-dimensional heterogeneous data is fused, wherein structured data is cleaned and standardized, unstructured text data is vectorized and sentiment features are extracted, and a unified feature vector is generated by filtering all features based on their importance. The feature vector is calculated using a medical compliance risk assessment model built based on machine learning algorithms, and the risk level representing the guardian's level of medical compliance and the key influencing factors used to explain the risk level are output. During the child's hospitalization, the compliance risk assessment model is periodically reassessed and its parameters are adjusted based on newly collected data and feedback from medical staff on the assessment results.

Images