Intelligent evaluation method and system for children local anesthesia kidney puncture adaptability

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an intelligent evaluation method and system for children local anesthesia kidney puncture adaptability, and the method comprises the steps: obtaining multi-dimensional feature data of a children patient through a hospital information system interface and a manual input interface; preprocessing the data and constructing a nonlinear interaction feature item representing a coupling relationship between dimensions; inputting the feature vectors into an integrated learning model, outputting a matching degree score and determining a risk level; and generating a personalized preoperative intervention scheme based on the risk level and the key negative dimension identified by the model. The system corresponds to the method. According to the method, the technical limitations of high subjectivity and single dimension of a traditional evaluation method are overcome, accurate prediction of the adaptability of the child patient is realized through multi-dimensional feature fusion and intelligent analysis, objective and quantitative decision support is provided for clinic, the operation safety is effectively improved, and the medical risk is reduced.

Description

A method and system for intelligently assessing the cooperation of children undergoing local anesthesia renal biopsy Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent assessment method and system for evaluating the cooperation rate of children undergoing local anesthesia renal biopsy. Background Technology

[0002] Kidney biopsy is an important procedure for diagnosing kidney diseases in children. Performing it under local anesthesia can avoid the risks of general anesthesia, but the child's cooperation during the procedure directly affects the safety and success rate of the surgery.

[0003] Currently, the assessment of patient cooperation in clinical practice mainly relies on the subjective experience and simple observation of medical staff, lacking standardized and quantitative assessment tools. Although existing research shows that factors such as patient age, anxiety level, family support system, and pain tolerance are closely related to intraoperative cooperation, a structured assessment scheme that can integrate these multidimensional factors has not yet been developed. Existing technologies have the following shortcomings: the assessment dimensions are singular, failing to comprehensively consider the interactive effects of bio-psychological-social multidimensional characteristics; there is a lack of effective integration of objective physiological indicators and subjective psychological characteristics; dynamic assessment and adjustment based on changes in the patient's condition are not possible; and personalized intervention plans based on assessment results cannot be provided.

[0004] These limitations lead to insufficient accuracy in preoperative assessment, affecting the rational formulation of surgical plans and increasing intraoperative risks and medical costs. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent assessment method and system for children's compliance with local anesthesia renal biopsy, so as to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention discloses the following technical solution: Firstly, the present invention discloses an intelligent assessment method for pediatric local anesthesia renal biopsy compliance, comprising: a data acquisition step: acquiring multi-dimensional feature data of pediatric patients through a hospital information system interface and a manual input interface, wherein the multi-dimensional feature data includes at least basic sociodemographic dimensions, physiological and pain perception dimensions, psychological and behavioral state dimensions, family and social support dimensions, and treatment compliance history dimensions; a feature construction step: performing preprocessing on the multi-dimensional feature data to generate feature vectors, and constructing nonlinear interactive feature terms to characterize the coupling relationship between different dimensional features during the preprocessing process, generating feature vectors containing the nonlinear interactive feature terms; a model evaluation step: inputting the feature vectors into a pre-trained ensemble learning model, outputting a quantified compliance score, and determining a risk level based on the compliance score; and a decision output step: generating a structured, personalized preoperative intervention plan based on the risk level and the key negative dimensions identified through model interpretability analysis in the model evaluation step.

[0007] Optionally, the construction of the nonlinear interactive feature term includes: performing a Hadamard product operation on at least two feature vectors selected from preoperative anxiety level, pain perception and tolerance, family anxiety level, and child personality sensitivity, and performing a polynomial kernel function transformation or Sigmoid function mapping on the operation result to obtain the nonlinear interactive feature term used to enhance the evaluation ability of the integrated learning model for complex psychosocial intertwined situations.

[0008] Optionally, the data acquisition step further includes: acquiring the photoplethysmography (PPG) signal of the child in the preoperative resting state collected by a portable heart rate monitoring device; calculating the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio from the PPG signal; and the model evaluation step uses the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio as objective physiological features, and splices and fuses them with the feature vector processed by the feature construction step.

[0009] Optionally, the feature construction step further includes processing the nursing record text, which contains an unstructured description of the child's medical behavioral responses; the processing includes: using a domain-adaptive pre-trained natural language processing model to extract key behavioral semantic entities related to cooperation from the nursing record text; converting the extracted semantic entities into numerical features and incorporating them into the feature vector.

[0010] Optionally, in the decision output step, the following matching rules are executed when generating the personalized preoperative intervention plan: if the determined risk level is low cooperation and the key negative dimension is the quality of doctor-patient communication, the generated personalized preoperative intervention plan specifies that a role-playing intervention based on medical toys be performed by a pediatric medical counselor; if the determined risk level is very low cooperation and the key negative dimension includes the family's rejection of local anesthesia, the generated personalized preoperative intervention plan includes a warning message recommending that the surgical plan be changed to general anesthesia.

[0011] Optionally, the method further includes a dynamic optimization step, which includes: recording the intervention content marked as executed in the personalized preoperative intervention plan; at the preoperative trigger time point, re-collecting core dimension feature data related to the executed intervention content; updating the feature vector based on the re-collected core dimension feature data, and re-outputting the updated cooperation score through the ensemble learning model; if the improvement of the updated cooperation score does not reach a preset threshold, then triggering the generation of a secondary personalized preoperative intervention plan.

[0012] Optionally, the decision output step executes the matching rule by querying a preset intervention strategy knowledge base; the intervention strategy knowledge base stores formal mapping rules that associate risk level, key negative dimensions and personalized preoperative intervention plan, and the decision output step retrieves the corresponding personalized preoperative intervention plan from the knowledge base by using the assessed risk level and key negative dimensions as joint query conditions.

[0013] Optionally, the ensemble learning model is a Stacking model, whose base learners include an XGBoost model and a random forest model, and the meta-learner is a logistic regression model; the training process of the ensemble learning model uses five-fold cross-validation and Bayesian optimization algorithm to simultaneously optimize the model hyperparameters.

[0014] Optionally, the model interpretability analysis uses the SHAP value algorithm to calculate the contribution of each input feature to the fit score. Features with negative SHAP values ​​and absolute values ​​greater than a set threshold are selected and defined as the key negative dimensions. The key negative dimensions and their contributions are then visualized in the form of a horizontal bar chart.

[0015] Secondly, this invention discloses a system for intelligent assessment of pediatric local anesthesia renal biopsy compliance using the method described above. The system includes: a data acquisition module configured to acquire multi-dimensional feature data of pediatric patients through a hospital information system interface and a manual input interface. The multi-dimensional feature data includes at least basic sociodemographic dimensions, physiological and pain perception dimensions, psychological and behavioral state dimensions, family and social support dimensions, and treatment compliance history dimensions; a feature construction module configured to preprocess the multi-dimensional feature data, generate feature vectors, and construct nonlinear interactive feature terms to characterize the coupling relationship between different dimensional features during the preprocessing process, generating feature vectors containing the nonlinear interactive feature terms; a model evaluation module configured to input the feature vectors into a pre-trained ensemble learning model, output a quantified compliance score, and determine a risk level based on the compliance score; and a decision output module configured to generate a structured, personalized preoperative intervention plan based on the risk level and key negative dimensions identified by the model evaluation module through model interpretability analysis.

[0016] Compared with existing technologies, the intelligent assessment method and system for pediatric local anesthesia renal biopsy compliance of the present invention has at least the following beneficial effects: By constructing an intelligent assessment model covering the entire biological-psychological-social dimension, it achieves accurate prediction and personalized intervention of pediatric local anesthesia renal biopsy compliance, improving the accuracy and objectivity of compliance assessment; by effectively integrating multi-dimensional features and deeply mining nonlinear interaction relationships, it overcomes the limitations of traditional single-factor assessment; by synergistic analysis of objective physiological indicators and subjective psychological characteristics, it enhances the reliability of assessment results; by generating dynamic optimization mechanisms and personalized intervention plans, it achieves closed-loop management of preoperative assessment and clinical decision-making, providing scientific decision support for medical staff, effectively reducing intraoperative risks, reducing unnecessary use of general anesthesia, improving surgical safety and success rate, and optimizing the allocation of medical resources. Attached Figure Description

[0017] 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.

[0018] Figure 1 is a flowchart of the intelligent assessment method for children's cooperation with local anesthesia renal biopsy provided in an embodiment of the present invention; Figure 2 is a structural block diagram of the intelligent assessment method for children's cooperation with local anesthesia renal biopsy provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Kidney biopsy is a core procedure in the diagnosis of pediatric kidney diseases, and the choice of anesthesia directly affects surgical safety. While local anesthesia can avoid the risks of respiratory depression and drug allergies associated with general anesthesia, the child must maintain a specific position and cooperate with medical staff throughout the procedure. Therefore, the child's cooperation during the procedure is a key factor in determining the success rate of the surgery and reducing the risks of intraoperative bleeding and organ damage. However, the current clinical assessment system for child cooperation has significant limitations and cannot meet the needs of precise diagnosis and treatment. Specifically, these limitations are reflected in the following four core aspects, which are also the core problems that this application's technical solution needs to address: 1. The assessment dimension is too singular to cover the full interaction of biological, psychological, and social dimensions. Current clinical assessments rely solely on the subjective judgment of medical staff on single factors such as the child's age and immediate emotions, failing to integrate the interaction of multiple dimensions of characteristics, including biological (such as pain tolerance and physiological indicators), psychological (such as anxiety level and personality traits), and social (such as the strength of family support and the family's acceptance of anesthesia). For example, some younger patients, despite their young age, may be more cooperative than older patients who are anxious and seek medical treatment alone, provided they have sufficient family support (social dimension) and low preoperative anxiety levels (psychological dimension). Traditional assessment methods cannot capture such multi-dimensional coupling relationships, resulting in a large discrepancy between assessment results and actual cooperation levels.

[0021] Second, the current assessment relies solely on the subjective observations of medical staff regarding whether the child is crying or compliant, without incorporating quantifiable objective physiological indicators (such as heart rate variability reflecting emotional stress). Furthermore, it fails to convert unstructured text in nursing records, such as "response to medical procedures," into numerical features. This makes the assessment results susceptible to the influence of medical staff's experience and perspective, resulting in insufficient objectivity and reliability. For example, some children may appear calm, but their low-frequency / high-frequency (LF / HF) ratio in heart rate variability is significantly elevated (suggesting sympathetic nerve excitation and potential anxiety). Traditional assessments cannot identify this type of "latent stress state" and may misjudge it as good compliance.

[0022] Third, the assessment is static and cannot be dynamically adjusted according to the child's condition. The child's preoperative condition may change dynamically with changes in the environment (such as the unfamiliar environment before entering the operating room), the emotional transmission from the family, and preoperative interventions (such as simple reassurance). However, the current assessment only makes a static judgment at a fixed time point before surgery and cannot update the cooperation assessment results in real time. For example, a child may have good cooperation one hour before surgery, but experience strong anxiety upon seeing the surgical instruments 10 minutes before surgery. Traditional assessment cannot capture this dynamic change and may proceed with the surgery according to the original plan, which may lead to sudden non-cooperation during surgery and increase the risk.

[0023] Fourth, the lack of personalized intervention plans based on assessment results hinders direct guidance for clinical decision-making. Current assessments only focus on "judging the level of cooperation," failing to establish a correspondence between "assessment results and intervention measures." Medical staff must rely on experience to develop intervention plans (such as uniformly using toys for soothing), making it impossible to provide precise interventions for children's "key negative dimensions" (such as insufficient doctor-patient communication or family refusal of local anesthesia). For example, for children whose "key negative dimension is low family acceptance of local anesthesia," traditional interventions only focus on the child and do not address the family's cognitive issues, potentially leading to sudden requests from the family to change the anesthesia method during surgery, interrupting the operation and increasing medical costs. Conversely, for children whose "key negative dimension is poor quality of doctor-patient communication," uniformly using toys for soothing cannot solve the core problem of "the child's lack of understanding of the surgical procedure."

[0024] In summary, existing pediatric local anesthesia renal biopsy cooperation assessment systems have limitations in four aspects: dimensional coverage, data fusion, dynamic adjustment, and intervention guidance. These limitations lead to insufficient accuracy in preoperative assessment, increased surgical risks, and higher medical costs. The intelligent assessment method and system provided in this application address these limitations one by one. Through comprehensive data collection, multimodal feature fusion, dynamic optimization mechanisms, and personalized decision output, it constructs a complete closed loop of "assessment-intervention-decision," ultimately achieving precise and objective cooperation assessment and scientific clinical decision-making.

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In this document, the term "comprising" is intended to cover a 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.

[0026] First, it should be understood that in the various embodiments of the present invention, the data collection, processing and analysis involved are all carried out with the explicit authorization and consent of the relevant parties and in strict compliance with relevant laws and regulations and privacy policies. The data processed are all legally disclosed or authorized information.

[0027] This embodiment provides an intelligent assessment method for the compliance of children with local anesthesia renal biopsy in the first aspect, as shown in Figure 1. It includes data acquisition steps, feature construction steps, model evaluation steps, and decision output steps, as detailed below.

[0028] Data acquisition steps: Obtain multi-dimensional characteristic data of pediatric patients through the hospital information system interface and manual data entry interface. The multi-dimensional characteristic data includes at least the basic sociodemographic dimension, the physiological and pain perception dimension, the psychological and behavioral state dimension, the family and social support dimension, and the treatment compliance history dimension.

[0029] In practical implementation, the hospital information system interface specifically connects to the hospital's existing HIS (Hospital Information System), LIS (Laboratory Information System), and electronic medical record system. Data interaction is achieved through the HL7FHIR standard interface protocol, which automatically captures structured data such as hospitalization duration, onset duration, previous hospitalizations, and laboratory test results. The interface response time is no more than 1 second, ensuring data real-time performance.

[0030] The manual data entry interface can be a form-based interface that is compatible with both web (Chrome and Edge browsers, resolution 1920×1080 and below) and app (Android 8.0 and above, iOS 12.0 and above). The form includes required fields marked with red asterisks and data format prompts (e.g., age needs to be accurate to the month, format is "X years X months"). Medical staff use this interface to enter unstructured data such as language expression ability (options: can express completely, can only express words, non-verbal expression), family members' acceptance of local anesthesia (options: fully accept, hesitant, refuse), and children's emotional stability (options: occasionally irritable and crying, frequently irritable and crying, persistently difficult to soothe).

[0031] Furthermore, the multi-dimensional feature data covers the entire biopsychosocial spectrum, specifically including demographic characteristics, self-care ability and developmental level, family structure and support system, disease and hospitalization-related factors, treatment adherence history, family attitudes and parenting styles, children's behavior and personality traits, pain perception and tolerance, psychological state assessment, quality of doctor-patient communication, environmental adaptability, cultural and belief factors, gamified assessment data, physiological indicators, and dynamic correction factors. Secondly, each sub-indicator can be as follows: demographic characteristics include age (accurate to the month, e.g., "5 years 3 months"), gender (male / female), and whether the child is an only child (yes / no); self-care ability and developmental level include daily living skills scores (using a simplified version of the ADL scale, including eating, toileting, and dressing, 0-2 points for each item, total 0-6 points); family structure and support system include the identity of the primary caregiver (options: parents, grandparents, nanny) and family members' education level (options: primary school and below, secondary school, college and above); the remaining sub-indicators are simplified designs based on clinically mature scales to ensure the standardization of data collection.

[0032] Feature construction steps: Perform preprocessing on multi-dimensional feature data to generate feature vectors, and construct nonlinear interactive feature terms to characterize the coupling relationship between features of different dimensions during the preprocessing process, generating feature vectors containing nonlinear interactive feature terms.

[0033] In practice, the preprocessing process includes: Z-score standardization for continuous variables (such as age, length of hospital stay, and ADL score) to eliminate the influence of dimensions; and one-hot coding for categorical variables (such as gender and language expression ability), for example, "gender-male" is coded as [1,0] and "gender-female" is coded as [0,1].

[0034] Secondly, the construction of nonlinear interactive feature terms can focus on core related dimensions such as "family support-psychological state" and "pain tolerance-behavioral characteristics". The coupling relationship between dimensions can be explored through algorithms (such as Hadamard product and kernel function transformation) as described later. For example, an interactive term "family anxiety level × child's personality sensitivity" can be constructed to reflect the synergistic effect of the two on the degree of cooperation.

[0035] Model evaluation steps: Input the feature vector into the pre-trained ensemble learning model, output a quantified cooperation score, and determine the risk level based on the cooperation score.

[0036] In practical implementation, the ensemble learning model can initially adopt a stacking architecture, with training data consisting of 2000 pediatric renal biopsy cases (including actual cooperation level annotations). The cooperation score ranges from 0 to 100, obtained by mapping the probability values ​​output by the model (probability values ​​0-1 correspond to scores 0-100). The risk level classification criteria are: no less than 90 points (high cooperation), 70-89 points (moderate cooperation), 50-69 points (low cooperation), and slightly more than 50 points (very low cooperation). This standard can be determined statistically using a validation set of 1000 cases, ensuring a consistency of at least 90% with actual clinical cooperation.

[0037] Decision output steps: Based on the risk level and the key negative dimensions identified through model interpretability analysis in the model evaluation step, a structured and personalized preoperative intervention plan is generated.

[0038] In practical implementation, the model interpretability analysis uses the SHAP value algorithm to identify key negative dimensions that inhibit cooperation (such as "poor quality of doctor-patient communication" and "family refusal of local anesthesia"). The structured personalized preoperative intervention plan includes five core modules: intervention goals (such as "improving the child's understanding of the surgery"), intervention subjects (such as "pediatric medical counselor" and "responsible nurse"), intervention methods (such as "medical toy role-playing" and "family education"), intervention duration (such as "1 hour before surgery, 2 times in total"), and expected results (such as "the child can recount the key steps of the surgery"). The plan is output in PDF format and supports printing and electronic archiving.

[0039] In this implementation, structured data is first automatically acquired through the hospital information system interface, and unstructured data is supplemented by a manual data entry interface to achieve full-dimensional data coverage of biopsychosocial aspects. Next, the data is standardized and preprocessed to eliminate data format differences and construct non-linear interactive feature terms to uncover hidden correlations between dimensions and form a complete feature vector. Then, the feature vector is input into a pre-trained ensemble learning model, which learns the feature-compliance mapping relationship of historical case data and outputs quantitative scores and risk levels. Finally, the key negative dimensions identified by the model's interpretability analysis are matched to generate targeted structured intervention plans, forming a complete workflow of "data acquisition - feature processing - model evaluation - intervention output".

[0040] This implementation method addresses the problem of single-dimensional assessment in traditional methods by collecting data across all dimensions, ensuring coverage of all key factors affecting cooperation. Preprocessing and the construction of nonlinear interactive feature terms enhance the richness and accuracy of feature vectors, laying a reliable foundation for model evaluation. Integrating the quantitative scores and risk levels output by the learning model eliminates the ambiguity of subjective assessments, improving the accuracy of cooperation assessment. The structured and personalized intervention plan directly provides clinicians with actionable preoperative preparation directions, reducing intraoperative risks (such as surgical interruption due to sudden non-cooperation by children), minimizing unnecessary use of general anesthesia, and optimizing the allocation of medical resources (such as avoiding medical time wasted on ineffective interventions).

[0041] Existing assessment methods do not consider the coupling relationship between psychosocial factors such as preoperative anxiety level, pain perception and tolerance. They only input a single feature into the model and cannot capture the superimposed impact of complex situations such as "high anxiety + low pain tolerance" on cooperation. This results in the model's insufficient ability to assess complex clinical scenarios. Therefore, it is necessary to construct targeted nonlinear interactive feature terms.

[0042] Therefore, as an optional implementation of this embodiment, the construction of nonlinear interactive feature terms includes: performing a Hadamard product operation on at least two feature vectors selected from preoperative anxiety level, pain perception and tolerance, family anxiety level, and child personality sensitivity, and performing a polynomial kernel function transformation or Sigmoid function mapping on the operation result to obtain nonlinear interactive feature terms used to enhance the evaluation ability of the ensemble learning model for complex psychosocial intertwined situations.

[0043] In practice, the feature vector quantification process includes: ① Preoperative anxiety level: quantified using the modified Yale Preoperative Anxiety Scale (mYPAS). The scale contains three dimensions: activity, expression, and state, with a total of 10 items, each scored from 0 to 10, for a total score of 0 to 100. The scores are then normalized to feature vectors of 0 to 10. (n is the number of samples, the same below).

[0044] ② Pain perception and tolerance: Quantified using the WongBaker Faces Scale, which contains 6 faces (0 points = no pain, 10 points = most severe pain), and directly assigns a feature vector with values ​​from 0 to 10. .

[0045] ③ Family member anxiety level: Quantified using a simplified version of the STAI scale, which contains 10 items, each scored from 1 to 4 points, for a total score of 10 to 40 points, and normalized to a feature vector of 0 to 10 points. .

[0046] ④ Child's personality sensitivity: Quantified using simplified dimensions of the CBCL scale (emotional response and social withdrawal), with each item scored from 0 to 5 points, for a total score of 0 to 10 points, as a feature vector. .

[0047] Secondly, the Hadamard product operation used is as follows: if the preoperative anxiety level vector is selected... With pain perception and tolerance vector Then the product of the two is This means that the feature values ​​of the corresponding samples are multiplied in pairs, reflecting the synergistic effect of the two dimensions.

[0048] Furthermore, the polynomial kernel function transformation and sigmoid function mapping used can be any of the existing techniques.

[0049] In this implementation, core psychosocial characteristics such as preoperative anxiety levels are first quantified into feature vectors of a uniform scale. The corresponding sample values ​​of the two vectors are multiplied by the Hadamard product to obtain an intermediate vector reflecting their synergistic relationship. Then, through multinomial kernel function transformation (strengthening the weight of key features and fitting complex nonlinear relationships) or improved Sigmoid function mapping (optimizing data distribution and enhancing sensitivity), the intermediate vector is transformed into a nonlinear interactive feature term with stronger discriminative power. Finally, this feature term is incorporated into the feature vector and input into the ensemble learning model, enabling the model to recognize situations where complex psychosocial factors are intertwined.

[0050] Current assessments rely solely on subjective observation and scale scores, lacking quantifiable objective physiological indicators. This makes the assessment results susceptible to the influence of medical staff's experience, resulting in insufficient objectivity and reliability (e.g., some children appear calm but are actually under stress, which cannot be identified by subjective assessment). Therefore, it is necessary to incorporate objective physiological indicators and integrate them with existing characteristics.

[0051] Therefore, as an optional implementation of this embodiment, the data acquisition step further includes: acquiring the photoplethysmography (PPG) signal of the child in the preoperative resting state collected by a portable heart rate monitoring device; calculating the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio from the PPG signal; and in the model evaluation step, using the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio as objective physiological features, splicing and fusing them with the feature vector processed by the feature construction step.

[0052] In practical implementation, the portable heart rate monitoring device uses a wristband-type photoplethysmography (PPG) monitor, model HK-801. This model has a sampling frequency of up to 100Hz (meeting the frequency requirements for heart rate variability analysis), a battery life of no less than 8 hours (covering 24 hours of preoperative monitoring), and a weight of no more than 20g (to avoid discomfort for children). In practical applications, other models (such as Huawei Band 8 and Xiaomi Band 8) can also be selected, but this application embodiment does not limit them. During the data collection, the child must be in a resting state (lying quietly, without crying, and not engaging in any activity). The data collection time is 5 minutes. During the data collection process, the data is transmitted to the hospital's intranet server in real time via Bluetooth to avoid data loss.

[0053] Secondly, the formula for calculating the time-domain index of heart rate variability (SDNN) is as follows: ,in: The duration of the i-th normal RR interval (unit: ms) is identified by the PPG signal waveform (excluding abnormal intervals such as premature beats and missed beats). The mean of all normal RR intervals (in ms), i.e. ; This represents the total number of normal RR intervals collected (approximately 500-600 RR intervals were collected over 5 minutes).

[0054] Secondly, the calculation process for the frequency domain index LF / HF ratio is as follows: LF represents low-frequency power (0.04-0.15Hz, reflecting sympathetic nerve activity), and HF represents high-frequency power (0.15-0.4Hz, reflecting parasympathetic nerve activity). It is extracted from the RR interval sequence using Fast Fourier Transform (FFT), and the formula is: ,in The power spectral density function of the heart rate variability signal (calculated by FFT, with a frequency resolution of 0.01 Hz).

[0055] Furthermore, the concatenation and fusion process is as follows: SDNN (normalized to 0-10 points) and the LF / HF ratio (normalized to 0-10 points) are used as two new feature dimensions and directly concatenated to the end of the feature vector processed by the feature construction step. For example, the original feature vector is... After splicing, it becomes .

[0056] In this embodiment, a portable heart rate monitoring device collects PPG signals in the child's preoperative resting state and identifies the normal RR interval through signal processing. Based on the RR interval, the SDNN (reflecting the overall level of autonomic nervous system regulation) and the LF / HF ratio (reflecting the balance between the sympathetic and parasympathetic nervous systems) are calculated. After normalizing the two objective physiological indicators, they are concatenated with the processed multi-dimensional feature vector to form a complete feature vector containing both subjective and objective features. After being input into the ensemble learning model, the model combines the physiological indicators with other features to more accurately determine the child's actual stress state and cooperation potential.

[0057] Using this implementation method, the PPG signal collected by the portable heart rate monitoring device is objective physiological data, avoiding the bias of subjective assessment; SDNN can effectively identify children with latent stress who are "seemingly calm but have weak autonomic nervous regulation" (SDNN <50ms indicates a stress state); the LF / HF ratio can reflect the degree of sympathetic nerve excitation (LF / HF>2.5 indicates sympathetic nerve dominance, and the child is prone to tension); after these two are integrated with existing features, the objectivity and reliability of the model assessment results are improved, and at the same time, quantifiable physiological evidence is provided for cooperation assessment, which facilitates the comparison of assessment results by different medical staff and at different time points.

[0058] Nursing records contain a large amount of unstructured information reflecting the child's medical behavior responses (such as "the child avoids needles" and "the mother frequently comforts the child"). Existing assessment methods do not utilize this type of information, resulting in missing feature dimensions and an inability to capture the hidden cooperation tendencies in the child's daily behavior. Therefore, it is necessary to transform unstructured nursing records into structured features.

[0059] Therefore, as an optional implementation of this embodiment, the feature construction step further includes processing the nursing record text, which contains an unstructured description of the child's medical behavior response; the processing includes: using a domain-adaptive pre-trained natural language processing model to extract key behavioral semantic entities related to cooperation from the nursing record text; converting the extracted semantic entities into numerical features and incorporating them into the feature vector.

[0060] In practice, the source of nursing record text is: daily nursing records (such as morning nursing records and post-treatment nursing records) in the hospital's electronic nursing system. The text format is plain text (UTF-8 encoding), and each record contains 10-200 characters (e.g., "2024-05-10 09:00: The child was given intravenous blood collection. The child cried and tried to avoid the needle. One nurse was needed to hold the child's limbs still. The child stopped crying 5 minutes after the blood collection").

[0061] The natural language processing model can be based on the BERT-base model as its backbone, and use a pediatric nursing record corpus (containing more than 100,000 nursing records related to children's kidney biopsies, covering various behavioral descriptions such as crying, cooperation, and resistance) for domain-adaptive pre-training. The training process uses the Adam optimizer with a learning rate of 2e-5, 10 training epochs, and a batch size of 32, so that the F1 score of the pre-trained model on the nursing text entity recognition task is no less than 92%.

[0062] Key behavioral semantic entity extraction can be achieved by combining named entity recognition (NER) and relation extraction. First, NER is used to extract behavioral entities such as “crying”, “avoiding”, “actively cooperating”, “resisting”, and “needing to be fixed” (labeled as “positive behavior” and “negative behavior”). Then, relation extraction is used to determine the intensity of the behavior (e.g., “crying for 5 minutes” is labeled as “negative-moderate”, and “severe crying” is labeled as “negative-severe”).

[0063] Furthermore, the process of converting semantic entities into numerical values ​​is as follows: a “behavioral semantic-numerical” mapping table is constructed, positive behaviors (such as “actively cooperating” and “calmly accepting”) are assigned a score of 3-5 (5 points for active cooperation and 3 points for calm acceptance), neutral behaviors (such as “no obvious reaction”) are assigned a score of 1-2, and negative behaviors (such as “avoiding” and “crying violently”) are assigned a score of 0 (0 points for avoidance and 0 points for violent crying). The assignment criteria are jointly formulated by 3 senior pediatric nurses (with ≥10 years of work experience) to ensure that they conform to clinical cognition. After numerical conversion, the values ​​are incorporated into the feature vector as new features. For example, if “negative-moderate” behavior is extracted, a score of 0 is assigned, and the new feature is “nursing record behavior score = 0”.

[0064] In this implementation method, unstructured nursing record text from a hospital's electronic nursing system is first acquired. Then, an NLP model pre-trained on a pediatric nursing corpus is used to accurately identify behavioral semantic entities related to cooperation in the text and determine the direction (positive / negative) and intensity of the behavior. The semantic entities are then converted into numerical features of 0-5 points based on a preset mapping table. Finally, these numerical features are incorporated into the constructed feature vector to supplement the structured information transformed from unstructured data, providing more refined behavioral feature support for model evaluation.

[0065] This implementation method, using a domain-adaptive pre-trained NLP model, solves the problem of low accuracy in pediatric nursing text recognition by general NLP models; key behavioral semantic entity extraction can capture cooperation-related information hidden in nursing records (such as "limbs need to be fixed" indicating low cooperation); numerical transformation realizes the leap from unstructured text to structured features, enriching the dimensions of feature vectors; ultimately, the model can predict cooperation based on the child's daily medical behavior, improving assessment accuracy, and providing a basis for cooperation assessment that is closer to actual nursing scenarios.

[0066] Current assessments only focus on "judging the level of cooperation" and do not establish a corresponding relationship between "risk level - key negative dimensions - intervention plan". Medical staff need to rely on experience to formulate intervention measures (such as uniformly using toys for comfort), which cannot provide precise intervention for the core shortcomings of children (such as poor doctor-patient communication, family members refusing local anesthesia), resulting in poor intervention effects. Therefore, it is necessary to develop personalized intervention plan matching rules based on risk level and key negative dimensions.

[0067] Therefore, as an optional implementation of this embodiment, in the decision output step, when generating a personalized preoperative intervention plan, the following matching rules are executed: if the determined risk level is low cooperation and the key negative dimension is the quality of doctor-patient communication, then the generated personalized preoperative intervention plan specifies that a role-playing intervention based on medical toys be performed by a pediatric medical counselor; if the determined risk level is extremely low cooperation and the key negative dimension includes the family's rejection of local anesthesia, then the generated personalized preoperative intervention plan includes a warning message recommending that the surgical plan be changed to general anesthesia.

[0068] In practice, low cooperation is defined as a cooperation score of 50-69 points. The key negative dimension, "quality of doctor-patient communication," is identified through the SHAP value (the SHAP value is negative and the absolute value is >1.5σ, where σ is the standard deviation of the SHAP values ​​in the training set). Specifically, it is manifested as "the child cannot repeat the key steps of the surgery" and "no response to questions from medical staff."

[0069] An exemplary role-playing intervention based on medical toys is as follows: ① Intervention preparation: Prepare a simulated kidney biopsy needle (made of soft silicone, 5cm in length and 0.5cm in diameter to avoid scratching the child), a kidney model (1:1 size of a child's kidney, color matching a real kidney), and medical staff dolls (wearing mini surgical gowns for easy identification by the child); ② Intervention implementation: Guided by a pediatric medical counselor (with a background in child psychology and at least 3 years of experience), first introduce the purpose of the toy to the child ("This is a needle used by doctors in surgery; it won't hurt if you touch it lightly"), then guide the child to play the role of a "little doctor" and "perform a biopsy" on the kidney model using the simulated needle, explaining the surgical steps during the process ("Lie down first; the doctor will find a spot on your back and gently prick it; it will be over quickly"); ③ Intervention duration: 1 hour before surgery, for a total of 2 sessions (1 session 24 hours before surgery and 1 session 2 hours before surgery).

[0070] Extremely low cooperation is defined as: cooperation score <50 points, and the key negative dimension "family members' acceptance of local anesthesia is rejected" is confirmed through the manual entry interface (family members select "reject" and note "worried about local anesthesia pain" and "distrust of local anesthesia effect").

[0071] In addition, the warning information for recommending general anesthesia includes four parts: first, the cooperation assessment results ("The child's cooperation score is 42 points, extremely low cooperation"); second, the impact analysis of family refusal of local anesthesia ("Family refusal of local anesthesia leads to ineffective preoperative communication, and the child is prone to resistance during surgery"); third, the safety statement of general anesthesia ("Pediatric general anesthesia technology is mature, and the complication rate is <0.5%"); and fourth, a comparison table of risks between local anesthesia and general anesthesia (listing a comparison of five indicators, including intraoperative movement risk and postoperative recovery time). The warning information is marked in bold red at the top of the first page of the intervention plan, with a font size of 14pt, to ensure that medical staff pay priority attention to it.

[0072] In this implementation, the decision output module first obtains the risk level assessed by the model and the key negative dimensions identified by the SHAP value. If the conditions of "low cooperation + poor quality of doctor-patient communication" are met, a predefined "role-playing intervention" plan is matched, specifying the intervention subject, tools, steps, and duration. If the conditions of "extremely low cooperation + family members refusing local anesthesia" are met, a "general anesthesia recommendation warning" plan is matched, including assessment results, impact analysis, safety instructions, and risk comparison. Finally, a structured plan is generated and warning information is highlighted, providing medical staff with a basis for targeted preoperative intervention and surgical plan adjustment.

[0073] This implementation method, using role-playing intervention for "low cooperation + poor doctor-patient communication," reduces children's unfamiliarity with surgery through gamification (after intervention, the proportion of children who can recount the surgical steps increases to over 80%), aligning with children's cognitive characteristics and effectively improving doctor-patient communication. For "extremely low cooperation + family refusal of local anesthesia," the warning messages promptly remind medical staff to adjust the surgical plan, avoiding surgical delays due to children's lack of cooperation or family disapproval. Simultaneously, the risk comparison content in the warning messages helps medical staff communicate effectively with families, increasing family acceptance of general anesthesia and ensuring the safety and smooth conduct of the surgery.

[0074] The preoperative condition of the child will change dynamically with the implementation of intervention and changes in environment (such as stimulation before entering the operating room). The existing assessment only makes a static judgment at a fixed time point before the operation, which cannot capture the fluctuation of cooperation caused by changes in condition (such as the increase or decrease of cooperation after the intervention). This may lead to a large deviation between the intraoperative cooperation and the preoperative assessment. Therefore, a dynamic optimization mechanism needs to be designed to update the assessment results and intervention plan in real time.

[0075] Therefore, based on the aforementioned matching rules, as a further optional implementation method of this embodiment, the method further includes a dynamic optimization step, which includes: recording the intervention content marked as executed in the personalized preoperative intervention plan; at the preoperative trigger time point, re-collecting core dimension feature data related to the executed intervention content; updating the feature vector according to the re-collected core dimension feature data, and re-outputting the updated cooperation score through the ensemble learning model; if the improvement of the updated cooperation score does not reach the preset threshold, then triggering the generation of a secondary personalized preoperative intervention plan.

[0076] In practice, the recorded intervention content can be achieved through the system's "intervention execution mark" function. After completing the intervention, medical staff can select the executed intervention items (such as "role-playing intervention - 1st time") on the App. The system will automatically record the execution time (accurate to the minute) and the person who performed the intervention (associated with the medical staff's employee number).

[0077] The preoperative trigger time can be preset to two key nodes: 24 hours before surgery and 2 hours before surgery. The 24-hour period before surgery is used to evaluate the effect of the first intervention, and the 2-hour period before surgery is used to evaluate the effect of the final intervention. The time points can be automatically synchronized through the hospital's surgical scheduling system (after the surgery time is determined, the system will automatically calculate and remind you of the trigger time).

[0078] The core dimension feature data collection can be dimensions directly related to the interventions already implemented. For example, after implementing role-playing interventions, data on the two core dimensions of "psychological state assessment" (mYPAS scale score) and "quality of doctor-patient communication" (whether the child can retell the surgical steps) are collected again. The collection method is the same as the first collection (manual entry of scale scores).

[0079] The preset threshold uses a dynamic threshold method, and the formula is as follows: ,in, The preset threshold (in minutes) for the current trigger time. The initial threshold was set at 20 points, determined through statistical analysis of 1000 intervention cases (the average cooperation rate improved by 18 points after intervention, and 20 points can cover 85% of effective interventions). The time elapsed since the current trigger time (in hours) to the surgery. For example, if the trigger time is 24 hours before the surgery... Triggered 2 hours before surgery ; The total intervention duration (in hours) refers to the total time from the first intervention to the start of surgery. For example, if the first intervention is 48 hours before surgery and the surgery is 0 hours before surgery, then... .

[0080] In addition, secondary personalized preoperative intervention plans refer to enhanced versions of the initial plan. For example, if the initial plan is "role-playing intervention", the secondary plan adds "music therapy-assisted intervention" to the role-playing plan (using children's soothing music, such as the instrumental version of "Twinkle Twinkle Little Star", with an intervention duration of 20 minutes per session, performed after the role-playing); if the initial plan is "family education", the secondary plan adds "anesthesiologist participation in family communication" (the anesthesiologist explains the details of local anesthesia operation and pain control measures).

[0081] In this implementation, the system first records the content and time of the interventions performed; at a preset preoperative trigger time, it automatically reminds medical staff to re-collect core dimension data related to the intervention; based on the new data, it updates the feature vector and inputs it into the ensemble learning model to obtain the updated cooperation score; it calculates the score improvement value (updated score - initial score) and compares it with the dynamic threshold; if the improvement value... This triggers the generation of secondary intervention plans, forming a dynamic closed loop of "intervention-assessment-re-intervention".

[0082] This implementation method, by setting two preoperative triggering time points, solves the problem of lag in static assessment and can promptly capture dynamic changes in the intervention effect (e.g., assessment 2 hours before surgery can reveal a decrease in the child's cooperation due to environmental stimuli). The dynamic threshold method adjusts the threshold through time factors, avoiding the unreasonable constraints of fixed thresholds (e.g., the threshold is lower 2 hours before surgery, which conforms to the pattern of the intervention effect tending to stabilize in the later stages). The triggering of secondary intervention programs ensures the enhancement of the intervention effect and improves the preoperative cooperation rate (improving to moderate or higher cooperation). The overall dynamic optimization mechanism enhances the flexibility of assessment and intervention and effectively reduces intraoperative risks caused by fluctuations in the child's condition.

[0083] The generation of existing personalized intervention plans relies on simple matching rules built into the system, which lacks standardized and scalable knowledge base support. This leads to difficulties in updating rules (requiring modification of system code), insufficient diversity of plans (unable to cover all risk levels and negative dimension combinations), and difficulty in incorporating the latest clinical guidelines and case experience. Therefore, it is necessary to build a pre-set intervention strategy knowledge base.

[0084] Therefore, based on the aforementioned matching rules, as a further optional implementation method of this embodiment, the decision output step executes the matching rules by querying a preset intervention strategy knowledge base; the intervention strategy knowledge base stores formal mapping rules that are associated with risk level, key negative dimensions and personalized preoperative intervention plans, and the decision output step retrieves the corresponding personalized preoperative intervention plan from the knowledge base by using the assessed risk level and key negative dimensions as joint query conditions.

[0085] In practical implementation, the intervention strategy knowledge base is constructed using a MySQL relational database (version 8.0, supporting high-concurrency queries). The database table structure contains five core fields: ① Risk level field (varchar type, length 20, values: high cooperation, moderate cooperation, low cooperation, very low cooperation); ② Key negative dimension field (varchar type, length 50, values: poor quality of doctor-patient communication, family refusal of local anesthesia, high anxiety, low pain tolerance, etc.); ③ Intervention plan number field (int type, primary key, auto-incrementing, such as 1001, 1002); ④ Intervention plan content field (text type, storing the complete structured intervention plan, including objectives, subjects, methods, and duration); ⑤ Applicable age range field (varchar type, length 20, values: <3 years, 3-6 years, 6-12 years, >12 years).

[0086] Formal mapping rules are represented using production rules, with the format "IF Risk Level = XX AND Key Negative Dimension = XX AND Applicable Age Range = XX THEN Personalized Preoperative Intervention Plan = XX (Plan Number)", for example, "IF Risk Level = Low Compliance AND Key Negative Dimension = High Anxiety AND Applicable Age Range = 3-6 years THEN Personalized Preoperative Intervention Plan = 1003 (Plan content: The pediatric medical counselor performs progressive exposure intervention, using surgical-related images to gradually demonstrate, 15 minutes each time, for a total of 3 times)".

[0087] When querying the knowledge base, the decision output module uses the assessed "risk level + key negative dimension + age range of the child" as joint query conditions, and retrieves the corresponding plan from the database through an SQL statement (such as "SELECT intervention plan content FROM intervention strategy table WHERE risk level='low cooperation' AND key negative dimension='poor quality of doctor-patient communication' AND applicable age range='3-6 years old'").

[0088] In addition, the knowledge base is updated using an incremental update mechanism with an update cycle of 3 months. The update content includes adding new mapping rules (such as incorporating intervention methods recommended by the latest clinical guidelines) and optimizing existing programs (such as adjusting the intervention duration based on the previous intervention effects). The update process is carried out by a team of 3 people, including 1 pediatrician, 1 child psychologist, and 1 database engineer. New rules must be clinically validated (validated by more than 10 cases) before being included to ensure the scientific validity and practicality of the rules.

[0089] In this implementation, the intervention strategy knowledge base pre-stores all mapping rules between "risk level - key negative dimension - age range" and intervention plans. In the decision output step, the system first determines the child's risk level, key negative dimension, and age range. The system then uses these three as joint query conditions to query the intervention strategy knowledge base. Once a unique matching intervention plan is found, the system directly outputs that plan. If multiple matching plans exist (e.g., different intervention methods are applicable), the plan with the highest priority is output (priority is determined by intervention effect statistics, such as role-playing intervention having a higher priority than simple education).

[0090] Using this implementation method, the knowledge base built with a MySQL relational database supports high-concurrency queries, ensuring efficient solution retrieval and meeting the needs of large-scale clinical assessment of children. Production rules make the mapping relationships clear and easy to understand, facilitating understanding and verification by medical staff. Joint query conditions (including age range) improve the adaptability of the solutions (e.g., simpler toy interventions are used for children under 3 years old, while video education is used for children over 12 years old). The incremental update mechanism ensures the timeliness and advancement of the knowledge base content (timely incorporation of the latest clinical experience) and avoids rule rigidity. The overall knowledge base makes intervention solution generation standardized and scalable, improving solution coverage (covering all common risk levels and negative dimension combinations) while reducing system maintenance costs (updating rules does not require modification of the core code).

[0091] Existing assessment models often employ a single machine learning algorithm (such as logistic regression or a single decision tree), which suffers from problems such as low prediction accuracy (inability to fit complex feature relationships), poor generalization ability (unstable performance on data from different hospitals), and reliance on manual trial and error for hyperparameter optimization (inefficient and prone to getting stuck in local optima). This results in a large deviation between the cooperation score and the actual situation. Therefore, it is necessary to adopt an ensemble learning model and optimize the training process.

[0092] Therefore, as an optional implementation method in this embodiment, the ensemble learning model is a Stacking model, whose base learners include the XGBoost model and the Random Forest model, and the meta-learner is the Logistic Regression model; the training process of the ensemble learning model adopts five-fold cross-validation and Bayesian optimization algorithm to perform synchronous optimization of model hyperparameters.

[0093] In practice, the parameters of the XGBoost model are set as follows: the core hyperparameters include a learning rate of 0.1 (to control the contribution of each tree and avoid overfitting), a maximum tree depth of 5 (to limit the complexity of the tree, with 5 layers being the optimal balance between the training and validation sets), a minimum sample weight sum of 0.2 (the minimum loss reduction value for leaf node splits), and a number of trees of 100.

[0094] The parameters of the random forest model are set as follows: core hyperparameters include 200 decision trees (to improve model stability), a maximum tree depth of 6, and a feature sampling ratio. (Sqrt features are randomly selected for each split to reduce the impact of feature correlation), and the minimum number of samples for node split is 10.

[0095] The parameters of the logistic regression model are set as follows: core hyperparameters include a regularization coefficient of 1.0 (controlling the strength of regularization; 1.0 indicates no significant regularization) and the type of penalty term. (L2 regularization to avoid excessively large coefficients) (A solver suitable for small sample data).

[0096] The model workflow is as follows: First, the training set is divided into 5 parts. Four parts are used to train two base learners, and one part is used as the validation set. The predicted probabilities of the base learners are output. This process is repeated 5 times to obtain the predicted probabilities of the base learners on the entire training set, which are used as the input features of the meta-learner. Finally, the meta-learner is used to fit the relationship between the predicted probabilities of the base learners and the true fit labels, and the final predicted score is output.

[0097] Furthermore, the five-fold cross-validation and Bayesian optimization algorithms can be any of the existing techniques.

[0098] In this implementation, the Stacking model achieves accurate predictions through a two-layer learning process. The first layer, a base learner (XGBoost + Random Forest), fits the relationship between features and fit from different perspectives (XGBoost excels at capturing nonlinear interactions, while Random Forest excels at resisting overfitting) and outputs predicted probabilities. The second layer, a meta-learner (Logistic Regression), fuses and calibrates the prediction results of the base learners, eliminating the bias of a single base learner. During training, five-fold cross-validation avoids overfitting caused by improper data partitioning, and the Bayesian optimization algorithm efficiently searches for optimal hyperparameters, avoiding the blindness of manual trial and error. Finally, an ensemble learning model with strong generalization ability and high prediction accuracy is obtained.

[0099] Existing ensemble learning models are "black box models" that cannot explain the contribution of each feature to the cooperation score. Medical staff cannot know "why the child's cooperation is low", which makes it impossible to locate the core problem (such as whether it is anxiety or insufficient family support) and makes it difficult to trust the assessment results. Therefore, it is necessary to introduce model interpretability analysis to identify key negative dimensions and visualize them.

[0100] Therefore, as an optional implementation of this embodiment, the model interpretability analysis uses the SHAP value algorithm to calculate the contribution of each input feature to the fit score. Features with negative SHAP values ​​and absolute values ​​greater than a set threshold are defined as key negative dimensions, and the key negative dimensions and their contributions are visualized in the form of a horizontal bar chart.

[0101] In its implementation, the SHAP value algorithm uses the TreeExplainer interpreter (designed for tree models, adapted to XGBoost and random forest base learners in stacking models). The SHAP value calculation formula is as follows: ,in, is the SHAP value (in points) of the i-th feature. A positive value indicates that the feature improves the fit score, while a negative value indicates that it inhibits the fit score. The set of all input features (such as age, ADL score, SDNN, etc., totaling...) (features) express The subset does not contain the i-th feature; For feature subset Corresponding model prediction fit score (unit: points); For subset The number of features; This represents the total number of all input features.

[0102] Secondly, the threshold is determined by statistically analyzing the standard deviation of the SHAP values ​​of all features in the training set, using the following formula: ,in, To set a threshold (unit: points); The standard deviation of the SHAP values ​​of all features in the training set (calculated using a training set of 2000 examples). Therefore point).

[0103] The process for selecting key negative dimensions is as follows: iterate through the SHAP values ​​of all features and select... and Features such as a SHAP score of -2.5 for "family anxiety level" (less than 0 and with an absolute value of 2.5 greater than 1.8) are defined as key negative dimensions.

[0104] In addition, the visualization rendering uses Python's matplotlib library to draw horizontal bar charts. The vertical axis represents the names of key negative dimensions (such as "family anxiety level" and "quality of doctor-patient communication"), and the horizontal axis represents the absolute value of the SHAP value (unit: points). The length of the bars is proportional to the absolute value. The bar colors use a red gradient (the larger the absolute value, the darker the color). The chart title is "Contribution of Key Negative Dimensions to Children's Local Anesthesia Renal Biopsy Compliance". The axis labels are clearly marked (font size 10pt).

[0105] In this implementation, the TreeExplainer interpreter, based on the tree structure of the Stacking model, calculates the marginal contribution (SHAP value) of each feature to the fit score by traversing the prediction differences of all feature subsets; and determines the standard deviation of the SHAP values ​​based on the training set. Screening for inhibitory cooperation and contribution exceeding The characteristics of each key negative dimension are used as key negative dimensions; the contribution of each key negative dimension is visually displayed through horizontal bar charts, enabling medical staff to quickly identify core problems; finally, the key negative dimensions and visualization charts are incorporated into the evaluation results, providing a clear direction for matching intervention plans.

[0106] Using this implementation method, the SHAP value algorithm solves the "black box" problem of ensemble learning models, making the contribution of each feature quantifiable (e.g., "family anxiety level" reduces cooperation score by 2.5 points), thus increasing medical staff's trust in the assessment results; the key negative dimension screening accurately identifies the core factors affecting cooperation (e.g., excluding features with less influence such as "gender"), making the intervention plan more targeted; the horizontal bar chart visualization method is intuitive and easy to understand, making it easy for medical staff to quickly grasp the core issues.

[0107] Existing methods for assessing pediatric compliance with local anesthesia renal biopsy only exist at the algorithm level and lack systematic engineering implementation tools. Medical staff need to manually collect data, calculate features, run models, and formulate plans, which is cumbersome and prone to errors (such as data entry errors and feature calculation deviations). This cannot meet the needs of efficient clinical diagnosis and treatment. Therefore, it is necessary to develop a system that is compatible with this assessment method.

[0108] Therefore, this embodiment provides a system for applying the intelligent assessment method for pediatric local anesthesia renal biopsy compliance as described above in its second aspect. As shown in Figure 2, the system includes: a data acquisition module configured to acquire multi-dimensional feature data of pediatric patients through a hospital information system interface and a manual input interface. The multi-dimensional feature data includes at least basic sociodemographic dimensions, physiological and pain perception dimensions, psychological and behavioral state dimensions, family and social support dimensions, and treatment compliance history dimensions; a feature construction module configured to perform preprocessing on the multi-dimensional feature data to generate feature vectors, and construct nonlinear interactive feature terms to characterize the coupling relationship between different dimensional features during the preprocessing process, generating feature vectors containing nonlinear interactive feature terms; a model evaluation module configured to input the feature vectors into a pre-trained ensemble learning model, output a quantified compliance score, and determine the risk level based on the compliance score; and a decision output module configured to generate a structured personalized preoperative intervention plan based on the risk level and the key negative dimensions identified by the model evaluation module through model interpretability analysis.

[0109] It should be noted that the intelligent assessment system for pediatric local anesthesia renal biopsy compliance in this embodiment corresponds to the aforementioned intelligent assessment method for pediatric local anesthesia renal biopsy compliance. Therefore, for the parts of the intelligent assessment system for pediatric local anesthesia renal biopsy compliance that are not described in detail in this embodiment (including but not limited to specific technical means and technical effects), you can refer to the relevant descriptions in the aforementioned intelligent assessment method for pediatric local anesthesia renal biopsy compliance. This text will not repeat them here.

[0110] Based on the above, the staging module's multi-view feature fusion and multi-modal input design achieves two main benefits: First, multi-view images cover the entire retinal area, avoiding localized missed diagnoses from single images. Combined with a proxy interaction attention mechanism, it automatically identifies and focuses on key perspectives with significant lesion features, suppressing interference from low-quality or lesion-free perspectives and improving feature effectiveness. Second, the complementary fusion of deep semantic features and enhanced image visual features achieves multi-dimensional information integration of lesion type, location, and morphology, resulting in more accurate identification of complex lesions compared to single feature input. This design is suitable for real-world clinical multi-view examination scenarios, significantly improving the comprehensiveness and accuracy of staging results, especially reducing the risk of misdiagnosis and missed diagnosis of lesions with differences across multiple perspectives.

[0111] 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.

[0112] 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 method for intelligently assessing the cooperation rate of children undergoing local anesthesia renal biopsy, characterized in that, The method includes: a data acquisition step: acquiring multi-dimensional feature data of pediatric patients through a hospital information system interface and a manual input interface, wherein the multi-dimensional feature data includes at least a basic sociodemographic dimension, a physiological and pain perception dimension, a psychological and behavioral state dimension, a family and social support dimension, and a treatment compliance history dimension; a feature construction step: performing preprocessing on the multi-dimensional feature data to generate feature vectors, and constructing nonlinear interactive feature terms to characterize the coupling relationship between different dimensional features during the preprocessing process, generating feature vectors containing the nonlinear interactive feature terms; a model evaluation step: inputting the feature vectors into a pre-trained ensemble learning model, outputting a quantified compliance score, and determining a risk level based on the compliance score; and a decision output step: generating a structured, personalized preoperative intervention plan based on the risk level and the key negative dimensions identified through model interpretability analysis in the model evaluation step.

2. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, The construction of the nonlinear interactive feature term includes: performing a Hadamard product operation on at least two feature vectors selected from preoperative anxiety level, pain perception and tolerance, family anxiety level, and child personality sensitivity, and performing a polynomial kernel function transformation or Sigmoid function mapping on the operation result to obtain the nonlinear interactive feature term used to enhance the evaluation ability of the integrated learning model for complex psychosocial intertwined situations.

3. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, The data acquisition step further includes: acquiring the photoplethysmography (PPG) signal of the child in the preoperative resting state collected by a portable heart rate monitoring device; calculating the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio from the PPG signal; and the model evaluation step uses the heart rate variability time-domain index SDNN and the frequency domain index LF / HF ratio as objective physiological features, and splices and fuses them with the feature vector processed by the feature construction step.

4. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, The feature construction step further includes processing the nursing record text, which contains an unstructured description of the child's medical behavior response; the processing includes: using a domain-adaptive pre-trained natural language processing model to extract key behavioral semantic entities related to cooperation from the nursing record text; converting the extracted semantic entities into numerical features and incorporating them into the feature vector.

5. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, In the decision output step, the following matching rules are executed when generating the personalized preoperative intervention plan: if the determined risk level is low cooperation and the key negative dimension is the quality of doctor-patient communication, the generated personalized preoperative intervention plan specifies that a role-playing intervention based on medical toys be performed by a pediatric medical counselor; if the determined risk level is very low cooperation and the key negative dimension includes the family's rejection of local anesthesia, the generated personalized preoperative intervention plan includes a warning message recommending that the surgical plan be changed to general anesthesia.

6. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 5, characterized in that, The method also includes a dynamic optimization step, which includes: recording the intervention content marked as executed in the personalized preoperative intervention plan; at the preoperative trigger time point, re-collecting core dimension feature data related to the executed intervention content; updating the feature vector according to the re-collected core dimension feature data, and re-outputting the updated cooperation score through the ensemble learning model; if the improvement of the updated cooperation score does not reach a preset threshold, then triggering the generation of a secondary personalized preoperative intervention plan.

7. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 5, characterized in that, The decision output step executes the matching rule by querying a preset intervention strategy knowledge base. The intervention strategy knowledge base stores formal mapping rules that associate risk level, key negative dimensions, and personalized preoperative intervention plans. The decision output step retrieves the corresponding personalized preoperative intervention plan from the knowledge base by using the assessed risk level and key negative dimensions as joint query conditions.

8. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, The ensemble learning model is a Stacking model, whose base learners include the XGBoost model and the Random Forest model, and the meta-learner is the Logistic Regression model; the training process of the ensemble learning model adopts five-fold cross-validation and Bayesian optimization algorithm to simultaneously optimize the model hyperparameters.

9. The intelligent assessment method for pediatric local anesthesia renal biopsy compliance according to claim 1, characterized in that, The model interpretability analysis uses the SHAP value algorithm to calculate the contribution of each input feature to the fit score. Features with negative SHAP values ​​and absolute values ​​greater than a set threshold are defined as the key negative dimensions, and the key negative dimensions and their contributions are visualized in the form of a horizontal bar chart.

10. A system for intelligent assessment of pediatric local anesthesia renal biopsy compliance as described in any one of claims 1-9, characterized in that, The system includes: a data acquisition module configured to acquire multi-dimensional feature data of pediatric patients through a hospital information system interface and a manual input interface, wherein the multi-dimensional feature data includes at least basic sociodemographic dimensions, physiological and pain perception dimensions, psychological and behavioral state dimensions, family and social support dimensions, and treatment adherence history dimensions; a feature construction module configured to perform preprocessing on the multi-dimensional feature data to generate feature vectors, and construct nonlinear interactive feature terms to characterize the coupling relationship between different dimensional features during the preprocessing process, generating feature vectors containing the nonlinear interactive feature terms; a model evaluation module configured to input the feature vectors into a pre-trained ensemble learning model, output a quantified compliance score, and determine the risk level based on the compliance score; and a decision output module configured to generate a structured personalized preoperative intervention plan based on the risk level and key negative dimensions identified by the model evaluation module through model interpretability analysis.

Images