Enteral nutrition information management method and system based on artificial intelligence

Abstract

The invention discloses an enteral nutrition information management method and system based on artificial intelligence, and relates to the technical field of medical artificial intelligence, and the method comprises the steps: carrying out the forward reasoning of an intestinal function-clinical decision fusion feature vector through an artificial intelligence decision model, obtaining a decision probability distribution vector, and carrying out the optimization and integration in combination with a clinical constraint condition, generating a nutrition intervention parameter set; performing decision basis analysis on the nutrition intervention parameter set, calculating each feature contribution degree, mapping each feature contribution degree into a natural language description fragment, and generating an enteral nutrition AI decision traceability report; and performing clinical research and judgment on the nutrition intervention parameter set and the enteral nutrition AI decision traceability report to generate an enteral nutrition clinical decision execution scheme. According to the method, the contribution degree of each feature is calculated and mapped into the natural language description fragment, the enteral nutrition AI decision traceability report is generated, transparency and interpretability of an artificial intelligence decision process are realized, and trust and adoption efficiency of an intelligent result are enhanced.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a method and system for managing enteral nutrition information based on artificial intelligence. Background Technology

[0002] Enteral nutrition support is a crucial component of basic clinical treatment, and its management effectiveness directly impacts patient recovery. Traditional enteral nutrition management primarily relies on the experience and judgment of clinicians, combining static laboratory test indicators (such as albumin and prealbumin) and intermittent vital sign monitoring data to formulate nutritional plans. With the development of medical informatization, electronic medical records have made the recording and retrieval of patient data more convenient, and some decision support tools have begun to employ basic statistical models or rule engines to perform preliminary calculations of nutritional supply.

[0003] However, existing methods still have room for improvement. There is a lack of effective means of collecting and analyzing physiological signals such as bowel sounds that can reflect the state of intestinal function in real time and continuously, making it difficult to adjust nutritional plans quickly based on immediate changes in intestinal tolerance. Furthermore, the decision-making process lacks sufficient transparency; the correlation between the output of artificial intelligence decision-making models and clinical characteristics is not clearly presented, affecting clinicians' trust in and adoption of intelligent results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based enteral nutrition information management method to address the problems of lagging dynamic adjustments to nutrition protocols and insufficient clinical trust.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an artificial intelligence-based enteral nutrition information management method, comprising: collecting and preprocessing patient clinical multimodal data and raw bowel sound audio signals; extracting features from the preprocessed raw bowel sound audio signals using a deep learning acoustic model to generate an intestinal acoustic feature vector; weighting and scoring the intestinal acoustic feature vector to obtain an intestinal peristalsis activity index, and fusing it with the preprocessed patient clinical multimodal data to generate an intestinal function-clinical decision fusion feature vector; performing forward reasoning on the intestinal function-clinical decision fusion feature vector using an artificial intelligence decision model to obtain a decision probability distribution vector, and optimizing and integrating it with clinical constraints to generate a nutritional intervention parameter set; analyzing the decision basis of the nutritional intervention parameter set, calculating the contribution of each feature, and mapping each feature contribution to a natural language description fragment to generate an enteral nutrition AI decision traceability report; and conducting clinical evaluation of the nutritional intervention parameter set and the enteral nutrition AI decision traceability report to generate an enteral nutrition clinical decision execution plan.

[0008] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the patient's clinical multimodal data includes vital signs, laboratory indicators, current nutritional infusion parameters, active drug records, and historical nutritional support data.

[0009] The preprocessing includes noise reduction filtering, missing value imputation, outlier removal, unit unification, and time sequence alignment.

[0010] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for obtaining the intestinal motility activity index are as follows:

[0011] The preprocessed raw audio signal of bowel sounds is input into a deep learning acoustic model to extract frequency domain, time domain and nonlinear features, and generate intestinal acoustic feature vectors.

[0012] Based on preset clinical guideline judgment rules, logical operations and weighted scoring are performed on the intestinal acoustic feature vector to obtain the intestinal peristalsis activity index.

[0013] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for generating the intestinal function-clinical decision fusion feature vector are as follows:

[0014] The intestinal peristalsis activity index and pre-processed patient clinical multimodal data were aligned and stitched together according to time windows to form multi-dimensional time-series data blocks;

[0015] Feature filtering and normalization are performed on multi-dimensional time-series data blocks to generate a fusion feature vector of intestinal function and clinical decision-making.

[0016] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for obtaining the decision probability distribution vector are as follows:

[0017] The intestinal function-clinical decision fusion feature vector is input into the artificial intelligence decision model for feature transformation and abstract representation learning to generate a high-level abstract feature vector.

[0018] The probability value of each preset nutritional intervention operation is calculated based on the high-level abstract feature vector, and a decision probability distribution vector is generated.

[0019] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for generating the nutritional intervention parameter set are as follows:

[0020] Load preprocessed patient clinical multimodal data, and optimize and filter the decision probability distribution vector using a constraint satisfaction algorithm to obtain a subset of candidate nutritional intervention operations;

[0021] A multi-objective weighted scoring algorithm is used to determine the optimal nutritional intervention operation from a subset of candidate nutritional intervention operations, and then dynamically integrates it with the parameters of the currently executed nutritional plan to generate a set of nutritional intervention parameters.

[0022] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for calculating the contribution of each feature are as follows:

[0023] The nutritional intervention parameter set and the gut function-clinical decision fusion feature vector were time-aligned and standardized to obtain a standardized feature-parameter dataset.

[0024] Based on the standardized feature-parameter dataset, the marginal contribution of each feature to the nutritional intervention parameter set in the gut function-clinical decision fusion feature vector is calculated, and the contribution of each feature is obtained.

[0025] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for generating the enteral nutrition AI decision traceability report are as follows:

[0026] Features whose contribution exceeds a preset contribution threshold and their corresponding nutritional intervention parameters are selected, and a list of high-contribution features and parameters is generated.

[0027] Based on a pre-defined clinical terminology mapping table, the list of high-contribution feature-parameter associations is converted into natural language description fragments.

[0028] By spatiotemporally aligning and integrating natural language description fragments with nutritional intervention parameter sets, an AI-driven enteral nutrition decision-making traceability report is generated.

[0029] As a preferred embodiment of the artificial intelligence-based enteral nutrition information management method of the present invention, the steps for generating an enteral nutrition clinical decision-making execution plan are as follows:

[0030] The pre-set clinical pathway rule set is invoked to perform indication verification and contraindication screening on the nutritional intervention parameter set, and a nutritional intervention verification parameter set is generated.

[0031] Based on a predefined safe dose range rule base, the nutritional intervention verification parameter set is checked for safe dose range, and a safety level score is output.

[0032] Based on the safety level score and the preset clinical safety operation rule library, the constraint boundary of the nutritional intervention parameter set adjustment is dynamically determined, and a safety constraint condition set is generated.

[0033] Under a set of safety constraints, a multi-objective optimization algorithm is used to dynamically adjust the set of nutritional intervention verification parameters and generate a set of alternative nutritional intervention parameters.

[0034] The set of alternative parameters for nutritional intervention is packaged into a standardized structured medical order document, and a digital signature is added to generate a clinical decision-making and execution plan for enteral nutrition.

[0035] Secondly, the present invention provides an artificial intelligence-based enteral nutrition information management system, including a data acquisition module, a feature extraction and fusion module, an intervention parameter prediction module, a contribution calculation module, and a clinical judgment module.

[0036] The data acquisition module is used to collect and preprocess the patient's clinical multimodal data and raw bowel sound audio signals.

[0037] The feature extraction and fusion module is used to extract features from the preprocessed raw audio signal of bowel sounds using a deep learning acoustic model to generate an intestinal acoustic feature vector; the intestinal acoustic feature vector is weighted and scored to obtain the intestinal peristalsis activity index, and then fused with the preprocessed patient clinical multimodal data to generate an intestinal function-clinical decision fusion feature vector.

[0038] The intervention parameter prediction module is used to perform forward reasoning on the intestinal function-clinical decision fusion feature vector through an artificial intelligence decision model, obtain the decision probability distribution vector, and optimize and integrate it with clinical constraints to generate a set of nutritional intervention parameters.

[0039] The contribution calculation module is used to analyze the decision basis of the nutritional intervention parameter set, calculate the contribution of each feature, map the contribution of each feature into natural language description fragments, and generate an enteral nutrition AI decision traceability report.

[0040] The clinical assessment module is used to conduct clinical assessments of the nutritional intervention parameter set and the enteral nutrition AI decision traceability report, and generate an enteral nutrition clinical decision execution plan.

[0041] The beneficial effects of this invention are as follows: By utilizing a deep learning acoustic model for feature extraction, an intestinal peristalsis activity index is obtained, and a fusion feature vector of intestinal function and clinical decision-making is generated, realizing real-time and objective quantification of intestinal function status. This provides a dynamic and personalized data foundation for nutritional decision-making, improving the accuracy and response speed of the assessment. By calculating the contribution of each feature and mapping it to natural language description fragments, an AI decision-making traceability report for enteral nutrition is generated, realizing the transparency and interpretability of the artificial intelligence decision-making process, and enhancing the trust in and adoption efficiency of intelligent results. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0043] Figure 1 This is a flowchart of an artificial intelligence-based enteral nutrition information management method.

[0044] Figure 2 This is a schematic diagram of an artificial intelligence-based enteral nutrition information management system.

[0045] Figure 3 A flowchart for nutritional intervention decision-making and generation.

[0046] Figure 4 A flowchart for decision tracing and report generation. Detailed Implementation

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0050] Reference Figures 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based enteral nutrition information management method, comprising the following steps:

[0051] S1. Collect patient clinical multimodal data and raw bowel sound audio signals and perform preprocessing;

[0052] Patient clinical multimodal data includes vital signs, laboratory indicators, current nutritional infusion parameters, active drug records, and historical nutritional support data;

[0053] It should be noted that vital signs refer to heart rate, blood pressure, blood oxygen saturation, and respiratory rate, which are continuously measured by monitoring equipment connected to the patient's body; laboratory indicators are the albumin and glucose levels obtained from the analysis of the patient's blood sample; current nutritional infusion parameters are the formula type, infusion rate, and volume already infused from the operating nutritional infusion equipment; active drug records are the names and dosages of sedative and vasopressor drugs currently in use, entered into the medical record by medical staff; and historical nutritional support data are records of previously used formula types, infusion rates, and physical responses found in the patient's past medical records.

[0054] It should be noted that the original audio signal of bowel sounds is a digital waveform data obtained by analog-to-digital conversion of a continuous analog electrical signal containing the sound of intestinal peristalsis, which is collected by an acoustic sensor attached to the patient's abdomen.

[0055] Preprocessing includes noise reduction filtering, missing value imputation, outlier removal, unit unification, and time sequence alignment;

[0056] It should be noted that noise reduction filtering refers to using digital filters to eliminate environmental noise, heart sounds, and respiratory sounds from the raw bowel sound audio signal; missing value imputation refers to using interpolation methods to fill in blank fields in the patient's clinical multimodal data; outlier removal refers to identifying and removing values ​​in the patient's clinical multimodal data that significantly exceed the physiologically reasonable range; dimensional unification refers to standardizing and converting indicators with different units in the patient's clinical multimodal data to make the data comparable; and temporal alignment refers to matching the raw bowel sound audio signal and the patient's clinical multimodal data along a unified time axis to ensure data synchronization.

[0057] S2. Feature extraction is performed on the preprocessed raw audio signal of bowel sounds using a deep learning acoustic model to generate an intestinal acoustic feature vector; the intestinal acoustic feature vector is weighted and scored to obtain the intestinal peristalsis activity index, and then fused with the preprocessed patient clinical multimodal data to generate an intestinal function-clinical decision fusion feature vector.

[0058] The preprocessed raw audio signal of bowel sounds is input into a deep learning acoustic model to extract frequency domain, time domain and nonlinear features, and generate intestinal acoustic feature vectors.

[0059] Furthermore, the preprocessed raw audio signal of bowel sounds is input into a deep learning acoustic model. A convolutional neural network layer performs local feature scanning on the preprocessed raw audio signal, extracting the sound pressure level amplitude, which characterizes short-term sound changes, as a time-domain feature. A short-time Fourier transform is applied to the time-domain features to convert the signal to the frequency domain, extracting the dominant frequency and frequency bandwidth, which characterize the frequency distribution of the sound, as frequency-domain features. Nonlinear dynamic analysis is performed on the activation values ​​of the intermediate layers of the deep learning acoustic model to calculate the sample entropy and the Hearst exponent, capturing the nonlinear features inherent in the bowel sound signal. The extracted time-domain features, frequency-domain features, and nonlinear features are then concatenated and standardized to generate an intestinal acoustic feature vector.

[0060] The expression for calculating sample entropy is:

[0061] ;

[0062] in, Indicates tolerance Below, the length is The number of matching templates between data point sequences; Indicates tolerance Below, the length is The number of matching templates between data point sequences; It is the pattern length, which is the number of consecutive data points used for comparison; It is a similarity tolerance; It is the total length of the data points of the preprocessed original audio signal of bowel sounds; It is the sample entropy. The lower the sample entropy, the stronger the regularity and the more predictable the preprocessed bowel sound original audio signal. The higher the sample entropy, the more complex and irregular the preprocessed bowel sound original audio signal.

[0063] The expression for calculating the Hearst exponent is:

[0064] ;

[0065] Taking the logarithm of both sides, we get:

[0066] ;

[0067] in, It is the range, indicating a length of Within a sub-interval, the maximum range of cumulative deviation of the preprocessed original bowel sound audio signal from its mean. It is a length of The standard deviation of the preprocessed original audio signal of bowel sounds within the sub-interval; It is the length of the subinterval; It is the Hearst index; It is a proportionality constant related to the data distribution and is used as the intercept term for linear fitting in the calculation of the Hearst exponent.

[0068] It should be noted that, in training the deep learning acoustic model, the preprocessed raw bowel sound audio signal and objective quantitative indicators directly related to intestinal function (such as gastric retention volume and clinical score of bowel sound auscultation) extracted from preprocessed patient clinical multimodal data are divided into training and validation sets according to a certain ratio. On the training set, the waveform amplitude of the raw bowel sound audio signal is normalized and scaled to the range of [-1,1] to generate standardized training samples. The standardized training samples are input into the deep learning acoustic model, and the difference between the output of the deep learning acoustic model and the objective quantitative indicators is calculated using the Adam optimizer and the cross-entropy loss function. The network parameters are updated through backpropagation to obtain the optimized deep learning acoustic model parameters. On the validation set, forward inference is performed on the optimized deep learning acoustic model parameters to calculate the validation loss. When the validation loss no longer decreases, training is stopped, and the trained deep learning acoustic model is output.

[0069] It should be noted that gastric retention volume is the volume of residual fluid in the stomach measured by gastric tube aspiration and ultrasound examination, used to objectively assess gastric emptying function and nutritional tolerance; the clinical score of bowel sound auscultation is the result of converting the description of bowel sound characteristics that meet medical standards in clinical auscultation records (such as frequency "X times per minute", intensity "hyperactive / decreased") into a numerical score through predefined scoring mapping rules (e.g., "hyperactive" corresponds to 3 points, "decreased" corresponds to 1 point), used to indirectly reflect the state of intestinal peristalsis; the scoring mapping rules are standardized conversion clauses set based on authoritative bowel sound auscultation grading standards (such as Fordran grading), including the correspondence between different bowel sound characteristics (such as frequency, intensity) and specific scores (such as 0-3 points).

[0070] Based on the preset clinical guideline judgment rules, logical operations and weighted scoring are performed on the intestinal acoustic feature vector to obtain the intestinal peristalsis activity index;

[0071] Furthermore, the system invokes preset clinical guideline judgment rules to compare the dominant frequency, frequency bandwidth, and sound pressure level amplitude in the intestinal acoustic feature vector with the corresponding dominant frequency threshold, frequency bandwidth threshold, and sound pressure level amplitude threshold in the clinical guideline judgment rules: if the dominant frequency is within the dominant frequency threshold, it is judged as normal rhythm; if the dominant frequency is below the dominant frequency threshold, it is judged as slow peristalsis; if the dominant frequency is above the dominant frequency threshold, it is judged as fast peristalsis; if the frequency bandwidth is within the frequency bandwidth threshold, it is judged as spectral concentration; if the frequency bandwidth is above the frequency bandwidth threshold, it is judged as spectral dispersion; if the frequency bandwidth is below the frequency bandwidth threshold, it is judged as... If the sound pressure level amplitude is too wide, it is judged as having too narrow a spectrum; if the sound pressure level amplitude is within the sound pressure level amplitude threshold, it is judged as having normal intensity; if the sound pressure level amplitude is below the sound pressure level amplitude threshold, it is judged as having too weak intensity; if the sound pressure level amplitude is above the sound pressure level amplitude threshold, it is judged as having too strong intensity. Each judgment result is assigned a corresponding quantitative score according to the weights defined in the clinical guideline judgment rules, and the quantitative scores obtained from the main frequency, frequency bandwidth and sound pressure level amplitude are weighted and summed to obtain a preliminary total score. The preliminary total score is normalized and mapped to an integer in a fixed range (such as 0 to 100) to obtain the intestinal peristalsis activity index.

[0072] It should be noted that the clinical guideline judgment rules are based on a digital rule base established according to the validated quantitative correspondence between acoustic characteristics and intestinal functional status. This includes a normal threshold table for acoustic parameters, judgment logic for different abnormal states, and corresponding weighting tables. The dominant frequency threshold is set based on large-scale statistical studies of bowel sounds in healthy individuals, with an exemplary range of 2-4 Hz, used to define normal peristaltic rhythms. Values ​​below 2 Hz indicate intestinal paralysis, while values ​​above 4 Hz suggest intestinal spasm or obstruction. The frequency bandwidth threshold is set based on the normal fluctuation range of intestinal motility, with an exemplary range of 1.5-3 kHz, used to distinguish between rhythmic peristalsis and disordered states. Values ​​above 3 kHz indicate regulatory dysregulation, while values ​​below 1.5 kHz indicate motility rigidity. The sound pressure level amplitude threshold is set based on the clinical correlation between bowel sound intensity and effective peristalsis, with an exemplary range of 40-60 dB, used to assess contraction strength. Values ​​below 40 dB indicate weak contractions, while values ​​above 60 dB indicate hyperactivity or a painful response.

[0073] The intestinal peristalsis activity index and pre-processed patient clinical multimodal data were aligned and stitched together according to time windows to form multi-dimensional time-series data blocks;

[0074] Furthermore, the intestinal motility activity index is aligned with the preprocessed patient clinical multimodal data using a time window consisting of a unified start time and duration. Within the time window, the numerical sequence of the intestinal motility activity index and the numerical sequence in the preprocessed patient clinical multimodal data are precisely matched according to the same sampling timestamp. For each matched time point, the value of the intestinal motility activity index and all the values ​​in the corresponding preprocessed patient clinical multimodal data are concatenated into a complete data row according to the field order. The data rows generated at all time points within the time window are arranged in chronological order to form a multidimensional time-series data block.

[0075] Feature filtering and normalization are performed on multi-dimensional time-series data blocks to generate a fusion feature vector of intestinal function and clinical decision-making.

[0076] Furthermore, the variance of each feature column in the multi-dimensional time-series data block is calculated, feature columns with variance close to zero are removed, and the maximum-minimum normalization method is used to linearly scale each retained feature column to the [0,1] interval to generate a standardized feature matrix. Principal component analysis is performed on the standardized feature matrix. By constructing the covariance matrix of the standardized feature matrix and performing eigenvalue decomposition, the characteristic polynomial of the covariance matrix is ​​solved to obtain all eigenvalues. At the same time, the non-zero solution of the corresponding homogeneous linear equation system for each eigenvalue is solved as the eigenvector. Principal components with a cumulative variance contribution rate exceeding the preset cumulative variance contribution rate threshold are selected in descending order of eigenvalues. The values ​​of the selected principal components are arranged in order of the corresponding eigenvalues ​​and concatenated to form a gut function-clinical decision fusion feature vector.

[0077] It should be noted that the cumulative variance contribution rate is set based on the statistical standard of information retention in principal component analysis, and is a fixed value of 95%. This is to achieve a balance between dimensionality reduction efficiency and information loss, and to ensure that the retained principal components can represent the vast majority of the original data features.

[0078] S3. By using an artificial intelligence decision-making model to perform forward reasoning on the intestinal function-clinical decision fusion feature vector, obtain the decision probability distribution vector, and combine it with clinical constraints for optimization and integration to generate a set of nutritional intervention parameters.

[0079] The intestinal function-clinical decision fusion feature vector is input into the artificial intelligence decision model for feature transformation and abstract representation learning to generate a high-level abstract feature vector.

[0080] Furthermore, the intestinal function-clinical decision fusion feature vector is Z-score standardized (standard deviation standardized) to transform the values ​​of each dimension into a distribution with a mean of 0 and a standard deviation of 1 before being input into the input layer of the artificial intelligence decision model. The artificial intelligence decision model uses a fully connected network and activation function in the hidden layer to perform nonlinear weighted summation and transformation on the standardized and scaled intestinal function-clinical decision fusion feature vector, extracting and combining features layer by layer. The output of each hidden layer is used as the input of the next hidden layer for deeper feature abstraction. After the step-by-step transformation of all hidden layers of the artificial intelligence decision model, a high-level abstract feature vector is output in the final hidden layer.

[0081] It should be noted that, in training the artificial intelligence decision-making model, the gut function-clinical decision fusion feature vector and the corresponding nutritional intervention parameter set are divided into a training set and a validation set according to a certain ratio. On the training set, the gut function-clinical decision fusion feature vector is Z-score standardized to generate standardized training samples, which are then input into the artificial intelligence decision-making model. The Adam optimizer (adaptive moment estimator) and the smoothed L1 loss function (Huber loss function) are used to calculate the regression difference between the output of the artificial intelligence decision-making model and the final nutritional intervention parameter set approved by clinical experts. The network parameters are then updated through backpropagation to obtain the optimized artificial intelligence decision-making model parameters. On the validation set, forward inference is performed on the optimized artificial intelligence decision-making model parameters to calculate the validation loss. When the validation loss no longer decreases, training is stopped, and the trained artificial intelligence decision-making model is output.

[0082] The probability value of each preset nutritional intervention operation is calculated based on the high-level abstract feature vector, and a decision probability distribution vector is generated.

[0083] Furthermore, the high-level abstract feature vector is input into the output layer of the artificial intelligence decision-making model. The output layer of the artificial intelligence decision-making model is a fully connected layer with the same number of neurons as the number of preset nutritional intervention operations. The high-level abstract feature vector and the weight matrix of the output layer are linearly calculated to obtain the original score of each preset nutritional intervention operation. The original score of each preset nutritional intervention operation is transformed by the Softmax function to obtain a probability value between a fixed interval (such as 0-1), and it is ensured that the sum of the probability values ​​of all preset nutritional intervention operations is 1. All probability values ​​are arranged in a fixed order of preset nutritional intervention operations to generate a decision probability distribution vector.

[0084] It should be noted that the preset nutritional interventions are a set of discrete actions based on clinical enteral nutrition support guidelines and common medical procedures. Each nutritional intervention is clearly defined and coded with a unique identifier.

[0085] Load preprocessed patient clinical multimodal data, and optimize and filter the decision probability distribution vector using a constraint satisfaction algorithm to obtain a subset of candidate nutritional intervention operations;

[0086] Furthermore, based on clinical guidelines, key medical logic (such as "potassium supplementation is contraindicated in patients with hyperkalemia") is formalized into initial rules. Specific numerical values ​​from preprocessed patient clinical multimodal data are used to set specific quantitative judgment criteria for the conditions in the initial rules (e.g., quantifying "hyperkalemia" as "serum potassium concentration > 5.5 mmol / L"), forming a constraint rule base. All preset nutritional intervention operations and their corresponding probability values ​​are read from the decision probability distribution vector. A constraint satisfaction algorithm iterates through each preset nutritional intervention operation in the decision probability distribution vector, verifying whether it satisfies all relevant constraints in the constraint rule base. Preset nutritional intervention operations that violate any constraint are removed from the decision probability distribution vector, obtaining a subset of candidate nutritional intervention operations.

[0087] It should be noted that the constraint rule base is a set of executable logical rules obtained by converting the text clauses of clinical guidelines into "if [quantified condition] then [execute / prohibit operation]", including threshold conditions set for vital signs, laboratory indicators and medication status data and corresponding operation permissions.

[0088] A multi-objective weighted scoring algorithm is used to determine the optimal nutritional intervention operation from the subset of candidate nutritional intervention operations, and dynamic integration is performed with the parameters of the currently executed nutritional plan to generate a set of nutritional intervention parameters;

[0089] Furthermore, a multi-objective weighted scoring algorithm is used to establish a scorecard for each candidate nutritional intervention operation in the subset of candidate nutritional intervention operations. The scorecard scores three objectives: clinical benefit, implementation risk, and execution cost. The clinical benefit score is based on the probability value of the candidate nutritional intervention operation in the decision probability distribution vector; the higher the probability value, the higher the clinical benefit score. The implementation risk score is based on the matching degree between the candidate nutritional intervention operation and the vital signs and laboratory indicators in the pre-processed patient clinical multimodal data; the lower the matching degree, the lower the implementation risk score. The execution cost score is based on the resources required for the candidate nutritional intervention operation and the current allocation. The greater the difference in settings, the lower the execution cost score. Based on preset weighting coefficients, the scores of each target are weighted and fused to obtain the comprehensive weighted total score of each candidate nutritional intervention operation. The candidate nutritional intervention operation with the highest comprehensive weighted total score is selected as the optimal nutritional intervention operation. The adjustment instruction represented by the optimal nutritional intervention operation (such as "increase the infusion rate by 10 ml / h") is dynamically integrated with the parameters of the currently executed nutritional protocol (such as the current infusion rate of 50 ml / h) to obtain the new target parameter value. The type of the optimal nutritional intervention operation and the new target parameter value are combined and encapsulated to generate a nutritional intervention parameter set.

[0090] It should be noted that the weighting coefficients are fixed values ​​calculated using mathematical modeling methods based on the explicit definitions of priority for each objective in clinical guidelines, large-scale clinical research data, and standardized operating procedures of medical institutions.

[0091] It should be noted that the nutritional intervention parameter set refers to the set of key parameters automatically generated by the artificial intelligence decision-making model based on the patient's clinical multimodal data and intestinal function status, used to guide enteral nutrition infusion. It includes core variables such as nutritional formula type, energy supply, protein and trace element ratio, infusion rate, infusion volume and duration, which are used to describe the nutritional supply structure of the patient at a specific stage. It reflects the comprehensive decision-making results on "what nutrients to input, at what rate, and within what time range" in enteral nutrition therapy, and in essence constitutes the parameter basis for intelligent nutritional intervention and refined nutritional management.

[0092] S4. Analyze the decision basis of the nutritional intervention parameter set, calculate the contribution of each feature, map the contribution of each feature into natural language description fragments, and generate an enteral nutrition AI decision traceability report.

[0093] The nutritional intervention parameter set and the gut function-clinical decision fusion feature vector were time-aligned and standardized to obtain a standardized feature-parameter dataset.

[0094] Furthermore, the nutritional intervention parameter set and the gut function-clinical decision fusion feature vector are aligned based on the timestamp of the generation time of the nutritional intervention parameter set. The dimensions of the gut function-clinical decision fusion feature vector and the data structure of the nutritional intervention parameter set are verified to confirm the consistency in data format and dimension definition. Z-score standardization is performed on the gut function-clinical decision fusion feature vector and the nutritional intervention parameter set respectively, transforming the values ​​of each dimension in the gut function-clinical decision fusion feature vector and the values ​​of each parameter in the nutritional intervention parameter set into a distribution with a mean of 0 and a standard deviation of 1, and then concatenating them to generate a standardized feature-parameter dataset.

[0095] Based on the standardized feature-parameter dataset, calculate the marginal contribution of each feature in the gut function-clinical decision fusion feature vector to the nutritional intervention parameter set, and obtain the contribution of each feature;

[0096] Furthermore, the variance of each parameter in the nutritional intervention parameter set portion of the standardized feature-parameter dataset is calculated. For each feature in the gut function-clinical decision fusion feature vector, the Pearson correlation coefficient between it and each nutritional intervention parameter in the nutritional intervention parameter set portion is calculated. The absolute values ​​of the correlation coefficients between each feature and all nutritional intervention parameters are weighted and averaged, with the weight being the variance of the corresponding nutritional intervention parameter, to obtain the comprehensive weighted correlation coefficient of each feature. The absolute value of the comprehensive weighted correlation coefficient is used as the linear association strength index of each feature with respect to the nutritional intervention parameter set. The linear association strength indices of all features are normalized to obtain the contribution of each feature.

[0097] The expression for calculating the linear correlation strength index is:

[0098] ;

[0099] in, It is the first The linear correlation strength index of each feature with respect to the set of nutritional intervention parameters; It is the first The first feature and the second Pearson correlation coefficients among nutritional intervention parameters; It is the first Variance of each nutritional intervention parameter; This represents the total number of nutritional intervention parameters;

[0100] Features whose contribution exceeds a preset contribution threshold and their corresponding nutritional intervention parameters are selected, and a list of high-contribution features and parameters is generated.

[0101] Furthermore, the contribution of each feature is compared with a preset contribution threshold. Features whose contribution exceeds the contribution threshold are identified as high-contribution features. The Pearson correlation coefficients corresponding to the high-contribution features are retrieved and paired to form "feature-parameter" association pairs. All "feature-parameter" association pairs are arranged in descending order of feature contribution to generate a high-contribution feature-parameter association list.

[0102] It should be noted that the contribution threshold is set based on the statistical distribution of the contribution of each feature (such as the median or upper quartile). An exemplary value range is 0.05-0.15. A value higher than 0.15 indicates that the feature has a key impact on the decision and needs to be explained in detail, while a value lower than 0.05 indicates that the feature has a weak impact and can be ignored.

[0103] Based on a pre-defined clinical terminology mapping table, the list of high-contribution feature-parameter associations is converted into natural language description fragments.

[0104] Furthermore, based on the feature and nutritional intervention parameter names of each "feature-parameter" association pair, the corresponding clinical description phrases and clinical operation descriptions are searched in a preset clinical terminology mapping table (e.g., mapping "dominant frequency" to "intestinal peristalsis rhythm" and "infusion rate" to "adjusting nutrient solution infusion rate"). According to a preset Pearson correlation coefficient threshold, the Pearson correlation coefficient in each "feature-parameter" association pair is converted into a correlation description: the direction of correlation is determined by the sign of the Pearson correlation coefficient; a positive Pearson correlation coefficient indicates a positive correlation, and a negative Pearson correlation coefficient indicates a negative correlation. Simultaneously, the strength of correlation is determined by comparing the absolute value of the Pearson correlation coefficient with the Pearson correlation coefficient threshold; a significant correlation is described if the absolute value of the Pearson correlation coefficient is greater than or equal to the threshold, and a weak correlation is described if the absolute value is less than the threshold. The mapped clinical description phrases, clinical operation descriptions, and correlation descriptions are then combined according to a fixed sentence template of "feature description and parameter description are correlated" to generate a natural language description fragment.

[0105] It should be noted that the Pearson correlation coefficient threshold is set based on the general criteria for judging strong correlation in statistics. The exemplary value range is -0.7 to 0.7. Values ​​above 0.7 may lead to oversensitivity to the linear relationship between the feature and the nutritional intervention parameter, while values ​​below -0.7 may lead to the neglect of significant negative association risks.

[0106] By spatiotemporally aligning and integrating natural language description fragments with nutritional intervention parameter sets, an AI decision-making traceability report for enteral nutrition is generated.

[0107] Furthermore, the natural language description fragments are time-aligned with the nutritional intervention parameter set based on the generation timestamp of the nutritional intervention parameter set. The natural language description fragments are used as explanatory text fields and mapped and bound at the field level to the structured numerical parameters (such as infusion rate and formula type) in the nutritional intervention parameter set. The natural language description fragments are used as the "decision basis" part and the nutritional intervention parameter set is used as the "recommended plan" part. They are encapsulated together in a standard document structure and a decision timestamp is added to generate an enteral nutrition AI decision traceability report.

[0108] S5. Conduct clinical analysis of the nutritional intervention parameter set and the enteral nutrition AI decision traceability report to generate an enteral nutrition clinical decision execution plan.

[0109] The pre-set clinical pathway rule set is invoked to perform indication verification and contraindication screening on the nutritional intervention parameter set, and a nutritional intervention verification parameter set is generated.

[0110] Furthermore, the pre-set clinical pathway rule set is invoked to match and verify the protocol features (such as formula type and infusion rate) described in the nutritional intervention parameter set with the indication standards, and screen out the permissible operations that conform to the current clinical scenario; the nutritional intervention parameter set is compared and screened with the contraindication clauses to identify and remove nutritional intervention parameter settings that conflict with the patient's current contraindications (such as using a high-potassium formula for a patient with hyperkalemia); the compliance status of the nutritional intervention parameters that pass the verification and screening is marked, and a nutritional intervention verification parameter set is generated.

[0111] It should be noted that the clinical pathway rule set is a set of executable logical rules based on disease diagnosis and treatment logic and indications / contraindications. It includes indication judgment conditions, contraindication screening criteria, and corresponding operational permissions defined for specific diseases, physiological states, and complications.

[0112] Based on a predefined safe dose range rule base, the nutritional intervention verification parameter set is checked for safe dose range, and a safety level score is output.

[0113] Furthermore, each target parameter (such as infusion rate and formula type) in the nutritional intervention verification parameter set is paired with the corresponding real-time physiological indicators (such as blood glucose and blood potassium concentration) in the pre-processed patient clinical multimodal data. For each pair of paired data, the corresponding safety threshold is searched in the predefined safe dose range rule base according to the value of the real-time physiological indicator. The target parameter is compared with the corresponding safety threshold. If the target parameter exceeds the corresponding safety threshold, it is recorded as a violation. The severity of each violation is obtained by calculating the percentage of the target parameter deviating from the safety threshold. The severity of all violations is accumulated and converted to a safety level score within a fixed range (such as 0 to 100) through a linear mapping function.

[0114] It should be noted that the safe dose range rule library is based on pharmacokinetic and human physiological tolerance limits to nutrients. It precisely quantifies the safe boundary threshold of each nutritional intervention parameter under different physiological states by establishing dose-response curves between physiological indicators and the maximum safe infusion rate.

[0115] Based on the safety level score and the preset clinical safety operation rule library, the constraint boundary of the nutritional intervention parameter set adjustment is dynamically determined, and a safety constraint condition set is generated.

[0116] Furthermore, based on the safety level scoring threshold in the preset clinical safety operation rule base, the safety level score is divided into three scoring intervals. According to the scoring intervals, the corresponding constraint rule clauses are extracted from the preset clinical safety operation rule base: when the safety level score is ≥80 points (low risk interval), the maximum adjustment range of the nutritional intervention parameter is 30%; when the safety level score is between 60 and 79 points (medium risk interval), the maximum adjustment range of the nutritional intervention parameter is 20%; when the safety level score is <60 points (high risk interval), the maximum adjustment range of the nutritional intervention parameter is 5%. The constraint rule clauses corresponding to the scoring intervals are combined to generate a set of safety constraint conditions.

[0117] It should be noted that the clinical safety operation rule base is a quantitative operation guideline based on the principle of medical risk classification and control. It includes the upper limit of the adjustment range of nutritional intervention parameters and mandatory operation restriction clauses corresponding to different safety level score ranges, as detailed below.

[0118] High-risk zone (safety score < 60 points): the upper limit for adjusting the corresponding nutritional intervention parameters is no more than 5% of the baseline value;

[0119] Medium-risk range (safety score 60-79): the upper limit for adjusting the corresponding nutritional intervention parameters is no more than 20% of the baseline value;

[0120] Low-risk range (safety score ≥ 80): the upper limit for adjusting the corresponding nutritional intervention parameters is no more than 30% of the baseline value;

[0121] The high-risk threshold is set based on the consensus of the critical point of patient safety status in the clinical risk grading standard. An exemplary value range is 60-80 points. A score below 60 points indicates that multiple safety indicators have exceeded the acceptable range, and forcibly adjusting nutritional parameters is very likely to cause adverse clinical events. A score above 80 points indicates that the patient's physiological state is stable and can withstand greater adjustments to the nutritional plan. Setting it too high will excessively limit the personalized optimization for stable patients.

[0122] Under a set of safety constraints, a multi-objective optimization algorithm is used to dynamically adjust the set of nutritional intervention verification parameters and generate a set of alternative nutritional intervention parameters.

[0123] Furthermore, using the nutritional intervention validation parameter set as the initial solution, a multi-objective optimization algorithm is used to establish an objective function with the optimization objectives of maximizing clinical benefits (i.e., meeting the patient's energy needs as much as possible) and minimizing implementation risks (i.e., avoiding blood glucose fluctuations or gastrointestinal discomfort). The adjustment range of nutritional intervention parameters and mandatory operational restrictions defined in the safety constraint set are used as hard constraints to construct the algorithm's search space. The multi-objective optimization algorithm iteratively searches within the search space to select the Pareto optimal solution that achieves a balance between maximizing clinical benefits and minimizing implementation risks while satisfying all hard constraints. The Pareto optimal solution is then integrated to form a set of nutritional intervention candidate parameters containing multiple alternative options.

[0124] The set of alternative parameters for nutritional intervention is packaged into a standardized structured medical order document, and a digital signature is added to generate a clinical decision-making and execution plan for enteral nutrition.

[0125] Furthermore, parameter types (such as infusion rate) and corresponding parameter values ​​are extracted from the set of nutritional intervention candidate parameters. Based on the clinical medical order document specifications, the parameter types and parameter values ​​are filled into the corresponding fields in the standardized medical order document template to form a preliminary structured document. The preliminary structured document is digitally signed using an authorized private key and bound to the preliminary structured document to generate an enteral nutrition clinical decision execution plan.

[0126] This embodiment also provides an artificial intelligence-based enteral nutrition information management system, including: a data acquisition module, a feature extraction and fusion module, an intervention parameter prediction module, a contribution calculation module, and a clinical assessment module; the data acquisition module is used to collect and preprocess patient clinical multimodal data and raw bowel sound audio signals; the feature extraction and fusion module is used to extract features from the preprocessed raw bowel sound audio signals using a deep learning acoustic model to generate an intestinal acoustic feature vector; the intestinal acoustic feature vector is weighted and scored to obtain an intestinal peristalsis activity index, and then fused with the preprocessed patient clinical multimodal data to generate intestinal function. - Clinical decision fusion feature vector; Intervention parameter prediction module, used to perform forward reasoning on the gut function-clinical decision fusion feature vector through an artificial intelligence decision model, obtain the decision probability distribution vector, and optimize and integrate it with clinical constraints to generate a set of nutritional intervention parameters; Contribution calculation module, used to analyze the decision basis of the nutritional intervention parameter set, calculate the contribution of each feature, and map each feature contribution to a natural language description fragment to generate an enteral nutrition AI decision traceability report; Clinical judgment module, used to perform clinical judgment on the nutritional intervention parameter set and the enteral nutrition AI decision traceability report to generate an enteral nutrition clinical decision execution plan.

[0127] In summary, this invention achieves real-time and objective quantification of intestinal function status by: utilizing a deep learning acoustic model for feature extraction to obtain the intestinal peristalsis activity index and fusing it to generate an intestinal function-clinical decision fusion feature vector. This provides a dynamic and personalized data foundation for nutritional decision-making, improving the accuracy and response speed of the assessment. Furthermore, by calculating the contribution of each feature and mapping it to natural language description fragments, an AI-based enteral nutrition decision traceability report is generated, achieving transparency and interpretability in the artificial intelligence decision-making process and enhancing trust in and adoption efficiency of intelligent results.

[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for managing enteral nutrition information based on artificial intelligence, characterized in that: include, Collect and preprocess the patient's clinical multimodal data and raw bowel sound audio signals; Feature extraction is performed on the preprocessed raw audio signal of bowel sounds using a deep learning acoustic model to generate an intestinal acoustic feature vector. The intestinal acoustic feature vector is weighted and scored to obtain the intestinal peristalsis activity index, and then fused with the preprocessed patient clinical multimodal data to generate an intestinal function-clinical decision fusion feature vector. By using an artificial intelligence decision-making model to perform forward reasoning on the feature vector of gut function-clinical decision fusion, the decision probability distribution vector is obtained, and then optimized and integrated in combination with clinical constraints to generate a set of nutritional intervention parameters. Analyze the decision-making basis of the nutritional intervention parameter set, calculate the contribution of each feature, map the contribution of each feature to natural language description fragments, and generate an enteral nutrition AI decision-making traceability report; Clinically evaluate the set of nutritional intervention parameters and the AI ​​decision-making traceability report for enteral nutrition to generate a clinical decision-making implementation plan for enteral nutrition.

2. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The patient's clinical multimodal data includes vital signs, laboratory indicators, current nutritional infusion parameters, active drug records, and historical nutritional support data; The preprocessing includes noise reduction filtering, missing value imputation, outlier removal, unit unification, and time sequence alignment.

3. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for obtaining the intestinal peristalsis activity index are as follows: The preprocessed raw audio signal of bowel sounds is input into a deep learning acoustic model to extract frequency domain, time domain and nonlinear features, and generate intestinal acoustic feature vectors. Based on preset clinical guideline judgment rules, logical operations and weighted scoring are performed on the intestinal acoustic feature vector to obtain the intestinal peristalsis activity index.

4. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for generating the gut function-clinical decision fusion feature vector are as follows: The intestinal peristalsis activity index and pre-processed patient clinical multimodal data were aligned and stitched together according to time windows to form multi-dimensional time-series data blocks; Feature filtering and normalization are performed on multi-dimensional time-series data blocks to generate a fusion feature vector of intestinal function and clinical decision-making.

5. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for obtaining the decision probability distribution vector are as follows: The intestinal function-clinical decision fusion feature vector is input into the artificial intelligence decision model for feature transformation and abstract representation learning to generate a high-level abstract feature vector. The probability value of each preset nutritional intervention operation is calculated based on the high-level abstract feature vector, and a decision probability distribution vector is generated.

6. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for generating the nutritional intervention parameter set are as follows: Load preprocessed patient clinical multimodal data, and optimize and filter the decision probability distribution vector using a constraint satisfaction algorithm to obtain a subset of candidate nutritional intervention operations; A multi-objective weighted scoring algorithm is used to determine the optimal nutritional intervention operation from a subset of candidate nutritional intervention operations, and then dynamically integrates it with the parameters of the currently executed nutritional plan to generate a set of nutritional intervention parameters.

7. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for calculating the contribution of each feature are as follows: The nutritional intervention parameter set and the gut function-clinical decision fusion feature vector were time-aligned and standardized to obtain a standardized feature-parameter dataset. Based on the standardized feature-parameter dataset, the marginal contribution of each feature to the nutritional intervention parameter set in the gut function-clinical decision fusion feature vector is calculated, and the contribution of each feature is obtained.

8. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for generating the AI-driven enteral nutrition decision-making traceability report are as follows: Features whose contribution exceeds a preset contribution threshold and their corresponding nutritional intervention parameters are selected, and a list of high-contribution features and parameters is generated. Based on a pre-defined clinical terminology mapping table, the list of high-contribution feature-parameter associations is converted into natural language description fragments. By spatiotemporally aligning and integrating natural language description fragments with nutritional intervention parameter sets, an AI-driven enteral nutrition decision-making traceability report is generated.

9. The artificial intelligence-based enteral nutrition information management method as described in claim 1, characterized in that: The steps for generating a clinical decision-making and implementation plan for enteral nutrition are as follows. The pre-set clinical pathway rule set is invoked to perform indication verification and contraindication screening on the nutritional intervention parameter set, and a nutritional intervention verification parameter set is generated. Based on a predefined safe dose range rule base, the nutritional intervention verification parameter set is checked for safe dose range, and a safety level score is output. Based on the safety level score and the preset clinical safety operation rule library, the constraint boundary of the nutritional intervention parameter set adjustment is dynamically determined, and a safety constraint condition set is generated. Under a set of safety constraints, a multi-objective optimization algorithm is used to dynamically adjust the set of nutritional intervention verification parameters and generate a set of alternative nutritional intervention parameters. The set of alternative parameters for nutritional intervention is packaged into a standardized structured medical order document, and a digital signature is added to generate a clinical decision-making and execution plan for enteral nutrition.

10. An artificial intelligence-based enteral nutrition information management system, based on the artificial intelligence-based enteral nutrition information management method according to any one of claims 1 to 9, characterized in that: It includes a data acquisition module, a feature extraction and fusion module, an intervention parameter prediction module, a contribution calculation module, and a clinical assessment module; The data acquisition module is used to collect and preprocess the patient's clinical multimodal data and raw bowel sound audio signals. The feature extraction and fusion module is used to extract features from the preprocessed raw audio signal of bowel sounds using a deep learning acoustic model to generate an intestinal acoustic feature vector; the intestinal acoustic feature vector is weighted and scored to obtain the intestinal peristalsis activity index, and then fused with the preprocessed patient clinical multimodal data to generate an intestinal function-clinical decision fusion feature vector. The intervention parameter prediction module is used to perform forward reasoning on the intestinal function-clinical decision fusion feature vector through an artificial intelligence decision model, obtain the decision probability distribution vector, and optimize and integrate it with clinical constraints to generate a set of nutritional intervention parameters. The contribution calculation module is used to analyze the decision basis of the nutritional intervention parameter set, calculate the contribution of each feature, map the contribution of each feature into natural language description fragments, and generate an enteral nutrition AI decision traceability report. The clinical assessment module is used to conduct clinical assessments of the nutritional intervention parameter set and the enteral nutrition AI decision traceability report, and generate an enteral nutrition clinical decision execution plan.

Images