Nutrition intervention system and method for tumor chemoradiotherapy patient

Abstract

The invention discloses a nutrition intervention system and method for tumor chemoradiotherapy patients. The system comprises a molecular nutrition fingerprint acquisition module, a space-time prediction modeling module, a drug-nutrition collaborative optimization module, an intestinal flora-host metabolism modeling module and an active nutrition intervention execution module. By integrating gene polymorphism, intestinal flora and metabonomics data, an individualized molecular nutrition fingerprint is constructed, a deep learning algorithm is adopted to predict the change trend of a nutrition state, a molecular docking technology is adopted to optimize a drug-nutrition synergistic effect, a flora-host metabolism model is established to predict nutrient absorption efficiency, and the nutrient absorption efficiency is improved. And precise active nutrition intervention is realized. The technical problems that in the prior art, nutrition intervention is low in accuracy, preventive intervention cannot be achieved, and medicine-nutrition interaction is neglected are solved, and individualized nutrition management based on the molecular level is achieved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical information technology, and particularly relates to a nutrition intervention system and method for tumor radiotherapy and chemotherapy patients. BACKGROUND

[0002] At present, the nutrition intervention for tumor radiotherapy and chemotherapy patients mainly adopts the following technical solutions:

[0003] Existing technology one: a nutrition assessment system based on the nutrition risk screening scale (such as NRS-2002), which determines the nutrition risk level through patient self-evaluation and medical staff evaluation, and formulates a standardized nutrition intervention plan. The problem of this technology is that it is only based on phenotype characteristics evaluation, and cannot reflect individual metabolic differences, so the precision of nutrition intervention is low.

[0004] Existing technology two: a nutrition status assessment system based on blood biochemical index monitoring, which determines the nutrition status changes by regularly detecting serum albumin, prealbumin and other indicators. The defect of this technology is that it is a lagging indicator, cannot achieve preventive intervention, and does not consider the influence of drug treatment on nutrition metabolism.

[0005] Existing technology three: a nutrition recommendation system based on artificial intelligence, which analyzes patient basic information and treatment plan through machine learning algorithm, and recommends nutrition intervention measures. The disadvantage of this technology is that it lacks individualized feature recognition at the molecular level, does not establish a time series prediction model of nutrition status, and cannot achieve active intervention.

[0006] The above existing technologies all have the following common technical problems:

[0007] Lack of individualized nutrition demand precision identification technology based on genetic polymorphism and metabolic characteristics;

[0008] Cannot realize multi-time scale prediction and preventive intervention of nutrition status changes;

[0009] Does not consider the influence of chemotherapy drug metabolism and nutrient interaction on treatment effect;

[0010] Neglects the regulatory role of intestinal flora on nutrition absorption efficiency, resulting in unstable intervention effect.

[0011] Therefore, it is urgent to develop a nutrition intervention system and method for tumor radiotherapy and chemotherapy patients to realize individualized precision prediction and active nutrition intervention. SUMMARY

[0012] To solve the above technical problems, on the one hand, the present application provides a nutrition intervention system for tumor radiotherapy and chemotherapy patients, comprising:

[0013] A molecular nutrition fingerprint collection module is used to detect patient genetic polymorphism, intestinal flora composition and blood metabolite concentration, and generate an individualized molecular nutrition fingerprint vector; preferably, the molecular nutrition fingerprint collection module detects single nucleotide polymorphism sites of CYP2D6, MTHFR and COMT nutrition metabolism related genes using whole genome association analysis technology, detects intestinal microbial composition using 16S rRNA gene sequencing technology, and detects amino acid, fatty acid and vitamin metabolite concentrations in serum using liquid chromatography-mass spectrometry; preferably, the molecular nutrition fingerprint collection module establishes an individual genotype-nutrition demand mapping table, calculates Shannon diversity index and Bray-Curtis distance to quantify flora structure characteristics, and constructs an individual metabolic profile; preferably, the molecular nutrition fingerprint collection module uses principal component analysis and clustering algorithm to reduce and fuse genotype, flora and metabolome data, and generates a 128-dimensional individualized molecular nutrition fingerprint vector;

[0014] A spatiotemporal prediction modeling module is used to predict nutrition status change trends and risk classification based on patient historical nutrition status data; preferably, the spatiotemporal prediction modeling module uses wavelet transform technology to extract hour-level, day-level and week-level time-frequency features of nutrition status data, and constructs a bidirectional LSTM network containing 256 hidden units; preferably, the input layer of the spatiotemporal prediction modeling module receives historical 7-day nutrition status data, the output layer predicts future 72-hour nutrition index change trends, and the self-attention mechanism is used to identify key time nodes and feature variables affecting nutrition status changes; preferably, the spatiotemporal prediction modeling module dynamically adjusts prediction weights, calculates the probability of malnutrition occurrence based on prediction results, and sets low risk (<20%), medium risk (20%-60%) and high risk (>60%) three-level early warning thresholds;

[0015] A drug-nutrition synergy optimization module is used to analyze the interaction of chemotherapy drugs and nutrients, and generate a synergistically optimized nutrition ratio scheme; preferably, the drug-nutrition synergy optimization module stores the half-life, clearance rate and metabolic pathway pharmacokinetic parameters of common chemotherapy drugs, as well as the molecular structure information of drug metabolites; preferably, the drug-nutrition synergy optimization module uses AutoDock algorithm to calculate the binding free energy of nutrient small molecules and drug metabolites, and identifies interaction pairs with binding affinity >-5.0 kcal / mol; preferably, the drug-nutrition synergy optimization module establishes a quantitative evaluation model for drug efficacy enhancement and toxicity reduction, calculates the influence coefficient of nutrient ratio on therapeutic index, and uses genetic algorithm to optimize the nutrient ratio;

[0016] a gut microbiota-host metabolism modeling module for predicting the bioavailability of nutrients under different microbiota states and formulating a synergistic plan for microbiota regulation and nutritional supplementation; preferably, the gut microbiota-host metabolism modeling module constructs a correlation network of microbiota abundance and serum metabolite concentration based on Spearman correlation analysis, and identifies significant correlation pairs with |r|>0.6 and p<0.01; preferably, the gut microbiota-host metabolism modeling module establishes a random forest regression model to predict the bioavailability of various nutrients under different microbiota states, and the model contains 500 decision trees with a maximum depth of 10; preferably, the gut microbiota-host metabolism modeling module matches corresponding probiotic strains according to the type of microbiota imbalance, establishes a three-dimensional decision matrix of strain-function-dose, and simultaneously optimizes the nutrient ratio and probiotic formula using a particle swarm optimization algorithm;

[0017] an active nutritional intervention execution module for generating an optimal intervention plan by comprehensively outputting the results of each module, controlling the delivery of nutrients to implement precise intervention and real-time adjustment; preferably, the active nutritional intervention execution module takes the improvement of nutritional status, the enhancement of treatment effect, and cost control as the objective function, and generates a Pareto optimal solution set using the NSGA-II algorithm; preferably, the active nutritional intervention execution module formulates a dynamic intervention timing table including the timing, duration, and adjustment frequency of intervention according to the predicted nutritional risk time node; preferably, the active nutritional intervention execution module realizes the precise delivery of nutrients by a programmable infusion pump and an intelligent capsule system, with a delivery accuracy of ±2%, and adjusts the intervention plan in real time according to the changes in the patient's nutritional indicators, with an adjustment period of 6-12 hours.

[0018] In another aspect, the present application also provides a nutritional intervention method for tumor radiotherapy and chemotherapy patients, comprising the following steps:

[0019] Step S1: Perform molecular nutrition fingerprint collection, start gene polymorphism detection program to obtain CYP2D6, MTHFR, COMT genotype data, at the same time start 16S rRNA sequencing program to obtain intestinal flora composition data and calculate diversity index, start liquid chromatography-mass spectrometry detection program to obtain serum metabolite concentration data, input the above three types of data into the fusion algorithm to generate a 128-dimensional individualized molecular nutrition fingerprint vector and store it in the patient's file;

[0020] Step S2: Perform nutritional status prediction, retrieve the molecular nutrition fingerprint vector from the patient's file, obtain the patient's historical 7-day nutritional status data from the monitoring system, start the wavelet transform program to extract multi-time scale features, input the feature data into a 256-unit bidirectional LSTM network for forward calculation, run the attention mechanism to assign prediction weights, and output the future 72-hour nutritional index prediction curve and risk probability classification results;

[0021] Step S3: Perform drug synergy optimization, obtain current chemotherapy drug information of the patient from the hospital information system, match corresponding metabolic parameters from the pharmacokinetics database, start the AutoDock program to calculate the binding energy of nutrients and drug metabolites, screen out interaction pairs with binding affinity greater than-5.0 kcal / mol, run a genetic algorithm to optimize the nutrient ratio, and output the synergistically optimized nutrient formula scheme;

[0022] Step S4: Perform microbial metabolism modeling, retrieve patient microbiota data and serum metabolite data, run Spearman correlation analysis to construct a correlation network, screen out significant correlation pairs with a correlation coefficient greater than 0.6 and a p value less than 0.01, start a random forest model of 500 decision trees to predict nutrient bioavailability, match probiotic strains according to the microbiota imbalance pattern, and run a particle swarm algorithm to optimize the synergistic configuration scheme of nutrients and probiotics;

[0023] Step S5: Perform intervention scheme implementation, receive the output data of the above four modules, run an NSGA-II multi-objective optimization algorithm to generate a Pareto solution set, develop an intervention timing plan according to the predicted risk time node, send control instructions to a programmable infusion pump and an intelligent capsule system, execute nutrient delivery according to a ±2% accuracy standard, collect patient nutrition index data every 6-12 hours, and feed the monitoring results back to step S2 to restart the prediction program, adjust the intervention parameters according to the new prediction results and repeat step S5.

[0024] The steps S2 to S4 are executed in parallel based on the molecular nutrition fingerprint vector generated in step S1, and step S5 forms a closed-loop feedback optimization mechanism with step S2 until the patient's nutritional status reaches the expected target range.

[0025] The beneficial effects of the present application are:

[0026] 1. Solve the problem of individualized nutrition demand recognition: the present application constructs a molecular nutrition fingerprint by integrating gene polymorphism, intestinal flora and metabolomics three-dimensional data, breaks through the limitation of relying only on phenotypic characteristics evaluation in the prior art, realizes precise recognition of individualized nutrition demand based on molecular level, and establishes a unique nutritional identity for each patient;

[0027] 2. Realize predictive management of nutritional risk: the present application uses a deep time series network and an attention mechanism to construct a multi-time scale prediction model, which can early warn of nutritional risk 24-72 hours in advance, changes the passive response mode of nutritional problems in the prior art, and realizes a fundamental change from therapeutic intervention to preventive intervention;

[0028] 3. Optimize drug treatment synergistic effect: the present application accurately calculates the interaction of nutrients and chemotherapy drug metabolites by molecular docking technology, generates synergistic formula with synergistic effect and attenuation, solves the problem of unstable treatment effect caused by ignoring drug-nutrient interaction in the prior art, and improves the safety and effectiveness of chemotherapy;

[0029] 4. Establish a nutrition absorption optimization mechanism regulated by flora: the present application constructs a flora-host metabolic correlation model, predicts the bioavailability of nutrients in different microecological environments, and makes up for the defects of the prior art in considering the influence of intestinal flora on nutrient absorption efficiency, thereby improving the controllability and stability of nutritional intervention;

[0030] 5. Construct an active intelligent intervention system: the present application establishes a closed-loop management mechanism of prediction-optimization-execution-feedback, realizes automation and intelligentization of nutritional intervention through multi-objective optimization algorithm and precise delivery technology, reduces human operation error, and reduces the work intensity of medical staff;

[0031] 6. Provide precise nutrition delivery guarantee: the present application uses programmable infusion pump and intelligent capsule system to realize timing and quantitative delivery with ±2% precision, avoids the problems of time delay and dose deviation existing in traditional manual drug delivery, and ensures accurate implementation and effect consistency of the intervention scheme. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The figure shows the architecture schematic diagram of the system described in the present application;

[0033] Figure 2 The figure shows the step flow chart of the method described in the present application DETAILED DESCRIPTION

[0034] In order for those skilled in the art to better understand the technical solutions of the present application, the present application will be described in detail below with reference to the drawings.

[0035] As shown in the figure, Figure 1 The nutritional intervention system for tumor radiotherapy and chemotherapy patients of the present application includes a molecular nutrition fingerprint collection module, a space-time prediction modeling module, a drug-nutrition synergistic optimization module, an intestinal flora-host metabolism modeling module and an active nutritional intervention execution module. The individualized molecular nutrition fingerprint vector generated by the molecular nutrition fingerprint collection module serves as the basic data, the space-time prediction modeling module, the drug-nutrition synergistic optimization module and the intestinal flora-host metabolism modeling module perform parallel analysis and calculation based on the molecular nutrition fingerprint vector, and the active nutritional intervention execution module generates the optimal intervention scheme and controls the execution based on the output results of the above three modules.

[0036] In one aspect, the system of the present application, when implemented, the molecular nutrition fingerprint collection module establishes a patient's unique biological characteristic file through an integrated multi-omics detection platform. The molecular nutrition fingerprint collection module is configured with an Illumina NovaSeq 6000 gene sequencer, a special 16S rRNA sequencing device, and a Thermo Fisher Q Exactive HF mass spectrometer to form a detection capability covering the genome, microbiome, and metabolome. Specifically, the molecular nutrition fingerprint collection module uses Qiagen DNeasy Blood & Tissue Kit to extract genomic DNA, and focuses on detecting single nucleotide polymorphism sites of CYP2D6, MTHFR, and COMT nutrition metabolism-related genes. The genotype data is processed using a numerical coding method, with wild type assigned a value of 0, heterozygous mutation assigned a value of 1, and homozygous mutation assigned a value of 2, thereby constructing a gene feature matrix wherein represents the number of patients, and 20 represents the total number of key gene sites.

[0037] Further, the molecular nutrition fingerprint collection module uses 16S rRNA gene V3-V4 variable region sequencing technology to detect the intestinal microecological characteristics of the patient. In order to quantify the diversity of the flora, the molecular nutrition fingerprint collection module calculates the Shannon diversity index:

[0038] wherein represents the Shannon diversity index value, represents the total number of detected bacterial species, represents the relative abundance of the th bacterial species, is the bacterial species number from 1 to , and a higher index value indicates a richer flora diversity. At the same time, the molecular nutrition fingerprint collection module uses Bray-Curtis distance to evaluate the similarity of the flora structure between samples:

[0039] wherein represents the Bray-Curtis distance between sample and sample , and respectively represent the abundance of the bacterial species in sample and sample , is the total number of bacterial species, is the bacterial species number from 1 to , and a smaller distance value indicates a more similar flora structure between the two samples. After CLR (center logarithmic ratio) transformation of the flora data, a feature matrix is formed , where 54 represents the extracted microbial community feature dimension.

[0040] Furthermore, the molecular nutrient fingerprinting module employs UPLC-MS / MS technology to analyze nutrient metabolites in patient serum. The detection range covers 54 key nutrients and their metabolites, including 26 amino acids, 13 vitamins, and 15 fatty acids. Metabolite data are processed using log10 transformation and Z-score normalization to form a feature matrix. , where 54 represents the metabolite feature dimension. To achieve the fusion of multi-omics data, the molecular nutrition fingerprinting module uses principal component analysis (PCA) for dimensionality reduction and integration, calculating the covariance matrix of the multi-omics data:

[0041] in Represents the covariance matrix. This is the complete feature matrix concatenated column by column. Representation matrix transpose, These are adjustment coefficients for the degrees of freedom. Through eigenvalue decomposition:

[0042] in For the first 1 eigenvalue, For the corresponding feature vector, This indicates the feature value number. The molecular nutrient fingerprint acquisition module selects the top 128 principal components with a cumulative contribution rate of 85% to form a 128-dimensional individualized molecular nutrient fingerprint vector. .

[0043] Accordingly, the spatiotemporal prediction modeling module constructs a nutritional status prediction model. After receiving the molecular nutritional fingerprint vector, the module combines it with nutritional status data collected in real time from the patient monitoring system to form a time-feature prediction framework. Specifically, the spatiotemporal prediction modeling module uses wavelet transform technology for time-frequency analysis.

[0044] in Represents the wavelet transform coefficients. The scaling parameter controls the degree of wavelet scaling. The translation parameters control the position of the wavelet on the time axis. For Daubechies wavelet basis functions, The input is a time series of nutrient states. Represents a time variable. This is achieved by setting the number of decomposition levels. The spatiotemporal prediction modeling module extracts features at three different time scales: hourly, daily, and weekly, forming a 96-dimensional time-frequency feature vector. where the hourly features 32 capture short-term fluctuations, the daily features 32 reflect circadian rhythms, and the weekly features 32 embody long-term trends.

[0045] Then, the spatio-temporal prediction modeling module constructs a bidirectional LSTM neural network to realize time series prediction. The bidirectional LSTM network contains two processing branches, forward and backward. The state update of the forward LSTM unit follows the following equation set:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052] where represents the sigmoid activation function, represents the Hadamard product, i.e., element-wise multiplication, represents the current time step, is the activation value of the forget gate at time t, is the activation value of the input gate at time t, is the activation value of the output gate at time t, is the candidate cell state at time t, is the current cell state at time t, is the cell state at time t, is the current hidden state at time t, is the hidden state at time t, is the input vector at time t. The weight matrix where 256 represents the number of hidden units, 288 is the input dimension including 256 previous time step hidden state dimensions and 32 current time step nutrition indicator dimensions, and the bias vector corresponds to the bias parameters of each gate unit, respectively.

[0053] Further, the spatio-temporal prediction modeling module integrates a self-attention mechanism. The self-attention calculation formula is:

[0054] wherein is a query matrix, is a key-value matrix, is a numerical matrix, all of which belong to , denotes the number of historical time steps, i.e. 7 days multiplied by 24 hours, denotes the dimension of the key-value vector, denotes the transpose of the key-value matrix, is a scaling factor used to prevent gradient vanishing, is a normalization function. The multi-head attention mechanism adopts attention heads in parallel calculation, and the dimension of each head is , satisfying .

[0055] Based on the prediction results of the LSTM network and the attention mechanism, the spatio-temporal prediction modeling module calculates the risk probability of malnutrition occurrence:

[0056] wherein denotes the risk probability of malnutrition occurrence, is a learned weight vector, denotes the transpose of the weight vector, is a scalar bias term, is the final hidden state vector of the LSTM network, denotes the exponential function. The spatio-temporal prediction modeling module sets the risk classification threshold: is low risk, is medium risk, is high risk.

[0057] Meanwhile, the drug-nutrition synergistic optimization module analyzes the interaction between chemotherapy drugs and nutrients. After receiving the patient's molecular nutrition fingerprint and chemotherapy regimen information, the drug-nutrition synergistic optimization module retrieves the corresponding molecular structure data and pharmacokinetic parameters from the pre-constructed drug knowledge base. The drug knowledge base covers 50 commonly used chemotherapy drugs, including platinum, taxanes, anthracyclines, and other major drug families.

[0058] Specifically, the drug-nutrition synergistic optimization module uses the AutoDock Vina algorithm for molecular docking calculation to evaluate the binding affinity of nutrient small molecules and drug metabolites. The molecular docking calculation combines the free energy:

[0059] wherein denotes the binding free energy, denotes the total number of energy terms, is the weight coefficient of the term, corresponding to the energy contribution of the energy term numbered from 1 to The formula for calculating the van der Waals interaction term is:

[0060] wherein represents the van der Waals interaction energy, represents the distance between atom and atom , is the repulsion term coefficient in the Lennard-Jones force field parameter, is the attraction term coefficient, and represent the atom numbers involved in the interaction, respectively. The hydrogen bond interaction term is:

[0061] wherein represents the hydrogen bond interaction energy, is a hydrogen bond geometry factor depending on the hydrogen bond angle, and are hydrogen bond parameters, represents the geometry configuration parameter of the hydrogen bond. The electrostatic interaction term adopts the Coulomb formula:

[0062] wherein represents the electrostatic interaction energy, and are the partial charges of atom and atom , respectively, is the dielectric constant, is the constant pi. The drug-nutrient synergy optimization module sets the binding affinity threshold value to be kcal / mol, wherein kcal / mol is the energy unit kilocalorie per mole.

[0063] Next, the drug-nutrient synergy optimization module optimizes the nutrient ratio by using a genetic algorithm. The fitness function of the genetic algorithm is:

[0064] wherein represents the individual fitness value, represents the efficacy enhancement index, represents the safety evaluation index, represents the toxicity risk index, , , are weight coefficients and satisfy . The genetic algorithm parameters include population size represents the number of individuals per generation, the crossover probability represents the probability of individual crossover, the mutation probability represents the probability of gene mutation, the maximum number of evolutionary generations represents the maximum number of iterations for algorithm termination.

[0065] In addition, the gut flora-host metabolic modeling module constructs a correlation model between the flora and nutritional metabolism. The gut flora-host metabolic modeling module uses the flora composition information in the molecular nutritional fingerprint, combines serum metabolite data, and uses Spearman correlation analysis to construct a flora-metabolite interaction network. The calculation formula of the Spearman correlation coefficient is:

[0066] wherein represents the Spearman correlation coefficient, represents the rank difference of the observation data, is the total number of samples, is the sample number from 1 to , , is a standardization factor. The significance of the correlation is evaluated by t-test:

[0067] wherein is the t-statistic, is the degree of freedom of the t-test, is the square of the correlation coefficient, is the residual variance term. The gut flora-host metabolic modeling module screens the correlation pairs that meet the conditions of and , wherein represents the absolute value of the correlation coefficient, represents the significance level.

[0068] Further, the gut flora-host metabolic modeling module establishes a random forest regression model to predict the bioavailability of nutrients under different flora states. The random forest adopts an ensemble learning strategy:

[0069] wherein represents the prediction result of the random forest, represents the total number of decision trees, is the prediction result of the th decision tree to the input , is the decision tree number from 1 to , represents the input feature vector. The maximum depth of each decision tree is limited to , and the minimum sample number for node splitting Minimum number of samples for leaf nodes The node splitting criterion uses Gini impurity:

[0070] in This indicates the impurity value of the Gini. For the number of categories, Category in the node The sample proportion Category numbers from 1 to .

[0071] Simultaneously, the gut microbiota-host metabolism modeling module employs a particle swarm optimization algorithm to synchronously optimize nutrient ratios and probiotic formulations. The particle velocity and position update formulas are as follows:

[0072]

[0073] in Indicates the first The particle in the first The velocity vector of the next iteration Indicates the first The particle in the first The velocity vector of the next iteration Indicates the first The particle in the first The position vector of the next iteration. Indicates the first The particle in the first The position vector of the next iteration. For inertial weights, As a learning factor, For particles The best historical position The optimal position for the entire group. and for Random numbers between, Indicates the particle number, This indicates the number of iterations. Algorithm parameter settings include the number of particles. Maximum number of iterations .

[0074] Finally, the proactive nutritional intervention execution module integrates the multidimensional analysis results, generates the optimal intervention plan, and implements precise control. The proactive nutritional intervention execution module uses the NSGA-II algorithm for multi-objective optimization, with the objective function being:

[0075] in Represents the objective function vector. is the vector of decision variables, , , are three optimization objectives, respectively, and superscript denotes vector transposition. Specifically, the nutrition status improvement objective function is:

[0076] where denotes the nutrition status improvement objective, denotes the improvement degree of the th nutrition index, is the weight of the th nutrition index, is the total number of nutrition indexes, is the nutrition index number from 1 to . The treatment effect enhancement objective function is:

[0077] where denotes the treatment effect enhancement objective, denotes the enhancement effect of the th treatment effect index, is the weight coefficient of the th treatment effect index, is the number of treatment effect indexes, is the treatment effect index number from 1 to . The cost control objective function is:

[0078] where denotes the cost control objective, denotes the use cost of the th nutrient, is the cost weight of the th nutrient, is the number of nutrient types, is the nutrient number from 1 to .

[0079] The safety constraint of daily nutrient intake needs to be met during the optimization process:

[0080] where denotes the intake amount of the th nutrient, is the minimum safe intake amount of the th nutrient, is the maximum safe intake amount of the th nutrient. The NSGA-II algorithm balances the convergence and diversity of solutions through non-dominated sorting and crowding distance calculation, and the crowding distance calculation formula is:

[0081] in Representing the solution Crowded distance, The number of objective functions, Representing the solution In the The function value of the next neighboring solution on each objective. Representing the solution In the The function value of the previous neighboring solution on each objective. For the first The maximum value of each target. For the first The minimum value of each objective. Number the objective functions from 1 to Algorithm parameter settings include population size. Maximum number of generations Crossover probability Probability of mutation .

[0082] Accordingly, the active nutritional intervention execution module establishes a real-time control system to execute the nutritional intervention plan. The control system uses a PID controller to achieve closed-loop feedback regulation.

[0083] in express The controller output at any given time. express The deviation between the set time value and the actual measured value for The target nutritional status setting at any given time. for The actual measured nutritional index values ​​at any given time. This is the proportionality coefficient. The integral coefficient is... The differential coefficients are... Indicates from 0 to The integral of the time deviation, This represents the derivative of the deviation with respect to time. This is the integral variable. Controller output. This is used to adjust the flow rate of the programmable infusion pump and the timing of drug delivery in the smart capsule system, thereby achieving delivery accuracy. Precise control.

[0084] like Figure 2As shown, on the other hand, the implementation process of the method of the present application embodies a multi-step synergistic, data-driven intelligent nutritional intervention. Step S1 as the data basis link of the whole method, through the standardized multi-omics detection process to establish the individual biological archives of the patient. Step S1 adopts a parallel detection strategy, simultaneously starting three detection procedures of gene polymorphism detection, intestinal flora analysis and metabolomics detection. Gene detection focuses on CYP2D6*1 / *2 / *4 variant sites, MTHFR C677T / A1298C polymorphic sites, COMT Val158Met variant sites, sequencing coverage requirement ≥ 30x, accuracy ≥ 99.9%. Intestinal flora analysis uses 16S rRNA gene V3-V4 variable region sequencing, sequencing depth is not less than 30,000 effective sequences, species annotation is based on Silva database version 138. Metabolomics detection covers 26 kinds of amino acids, 13 kinds of vitamins, 15 kinds of fatty acids, a total of 54 key metabolites, detection limit reaches ng / ml level.

[0085] Based on the 128-dimensional molecular nutrition fingerprint vector generated in step S1, steps S2 to S4 form a synergistic processing architecture of parallel analysis. Step S2 focuses on the time series prediction of nutritional status, step S2 retrieves the molecular nutrition fingerprint vector from the patient archives, combined with the historical 7-day nutritional status data of 168 time points obtained from the monitoring system. Data preprocessing includes outlier detection (using 3σ criterion), missing value filling (forward filling method) and standardization processing (Min-Max normalization). Wavelet transform uses db4 wavelet basis for 5-layer decomposition, bidirectional LSTM network training uses Adam optimizer, learning rate , batch size , training rounds , the loss function is the weighted combination of MSE and cross-entropy , wherein represents the total loss function, represents the mean square error loss, represents the cross-entropy loss, and 0.7 and 0.3 are the corresponding weight coefficients.

[0086] At the same time, step S3 focuses on the accurate analysis of drug-nutrition interaction. After step S3 obtains the patient's chemotherapy scheme from the hospital HIS system, it retrieves the corresponding molecular structure and pharmacokinetic parameters in the drug database. Molecular docking uses AutoDockVina algorithm, search space is set as 20x20x20Å cube, wherein Å represents angstrom unit, i.e. 10^{-10} meter, search accuracy represents the search thoroughness parameter, and the number of docking times Indicates the number of conformation modes generated. The nutrient molecule library contains the optimized three-dimensional structures of 47 compounds, and the conformation optimization is performed using the MMFF94 force field, where MMFF94 represents the Merck Molecular Force Field version 94. The docking calculation uses Lamarckian genetic algorithm, population size Indicates the number of individuals in the genetic algorithm population, the maximum number of evaluations Indicates the maximum number of fitness function evaluations, mutation rate Indicates the probability of gene mutation in the genetic algorithm.

[0087] Step S4 is performed in parallel to model the analysis of the correlation between the flora and the nutrient metabolism. Step S4 selects the top 100 genera for analysis, covering key functional flora such as Bacteroides, Lactobacillus, Bifidobacterium, etc. Spearman correlation analysis is implemented using the scipy.stats module, where scipy.stats is the statistical module in the Python scientific computing library, and multiple test correction uses the Benjamini-Hochberg method to control the false discovery rate FDR < 0.05, where FDR represents False Discovery Rate. The random forest model is constructed using the scikit-learn library, where scikit-learn is a Python machine learning library, and the hyperparameters are optimized through grid search: number of decision trees , maximum depth , minimum split sample size . Model performance is evaluated by 5-fold cross-validation, and evaluation metrics include RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and (determination coefficient), where the calculation formula of is:

[0088] where represents the residual sum of squares, represents the total sum of squares.

[0089] Finally, step S5 undertakes the important mission of integrating multi-dimensional analysis results, generating optimal solutions, and implementing precise interventions. Step S5 receives the comprehensive analysis results from the previous three steps, including the nutrient risk time window, the drug synergistic matching scheme, and the flora optimization recommendations, and generates a Pareto optimal solution set containing 10-20 non-dominated solutions through the NSGA-II algorithm. Scheme selection uses the TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) method for multi-attribute decision analysis, and selects the scheme with the highest comprehensive score as the final execution scheme. The scoring formula of the TOPSIS method is:

[0090] where represents the scheme the overall score, the representation scheme the distance to the positive ideal solution, the representation scheme the distance to the negative ideal solution, the scheme number.

[0091] Specifically, the intervention is implemented through an integrated intelligent device system, including a Baxter Colleague 3CXE programmable infusion pump and an AdhereTech smart capsule system. The infusion pump supports 8-channel simultaneous infusion with a flow rate range of 0.1-999.9 ml / h, where ml / h represents milliliters per hour, a flow rate accuracy of ±1%, and a volume accuracy of ±2%. The smart capsule system supports Bluetooth 4.0 connection, battery life ≥6 months, and medication reminder accuracy ≥95%. Real-time monitoring is performed using an Abbott i-STAT portable biochemical analyzer to detect 13 indicators, including Na⁺ (sodium ion), K⁺ (potassium ion), Cl⁻ (chloride ion), glucose, BUN (blood urea nitrogen), and creatinine, with a detection time of 2-3 minutes and an accuracy CV <5%, where CV represents the coefficient of variation.

[0092] To ensure the continuous optimization of the intervention effect, step S5 establishes a closed-loop feedback mechanism. Step S5 automatically collects patient nutrition index data every 6-12 hours, and the monitoring results are fed back to step S2 to restart the prediction analysis in real time. The processing of the feedback data uses the sliding window average method:

[0093] where represents the moving average value at time t, is the window size set to 6, represents the measurement value at time t, is the time offset from 0 to . Based on the updated prediction results, step S5 dynamically adjusts the intervention parameters and repeats the intervention process, and the adjustment amplitude is calculated by the following formula:

[0094] where represents the adjustment amplitude, is the adjustment coefficient, is the new risk probability prediction value, is the previous risk probability prediction value.

[0095] This process continues until the patient's nutritional status reaches the expected target range, and the judgment criteria for achieving the target are:

[0096] wherein represents the actual value of the th nutritional index, represents the target value of the th nutritional index, is the total number of nutritional indexes, is the allowable relative error threshold value, forming a complete closed-loop control process of prediction-intervention-monitoring-adjustment.

[0097] The above-mentioned individualized prediction and active intervention mechanism based on molecular nutrition characteristics runs through the execution process of steps S1 to S5, dynamically adjusts the nutritional intervention strategy according to the genotypes, flora states and metabolic characteristics of different patients, and improves the intervention effect and treatment safety.

[0098] In summary, the present application provides a nutritional intervention system and method for tumor radiotherapy and chemotherapy patients, which realizes individualized molecular feature recognition, accurate nutritional state prediction, drug interaction optimization, flora metabolism modeling and intelligent active intervention through the synergistic cooperation of molecular nutrition fingerprint collection, spatiotemporal prediction modeling, drug-nutrition synergistic optimization, intestinal flora-host metabolism modeling and active nutritional intervention execution. The present application can be widely applied in the fields of tumor clinical nutrition management, individualized medical treatment, etc., and has good practicability and popularization value.

[0099] Although this document demonstrates the specific application of the present application through specific examples, these examples are only for illustrative purposes and do not mean to limit the protection scope of the present application. The protection scope of the present application is defined by the claims, and any appropriate modification, equivalent replacement or improvement based on the principle of the present application shall be considered to fall within the protection scope of the present application. Therefore, the protection scope of the present application should be given the broadest interpretation to cover all such modifications, equivalent structures and functions.

Claims

1. A nutritional intervention system for cancer patients undergoing radiotherapy and chemotherapy, characterized in that, include: The molecular nutrition fingerprinting module is used to detect single nucleotide polymorphism sites of CYP2D6, MTHFR, and COMT nutrition metabolism-related genes, gut microbiota composition, and blood metabolite concentrations in patients, generating personalized molecular nutrition fingerprint vectors. The spatiotemporal prediction modeling module is used to build a bidirectional LSTM network to receive patients' historical nutritional status data, predict the trend of changes in nutritional indicators, and perform risk classification. The drug-nutrient synergistic optimization module is used to calculate the interaction between nutrients and the metabolites of chemotherapy drugs, and generate a synergistically optimized nutrient ratio scheme. The gut microbiota-host metabolism modeling module is used to construct a network of associations between the microbiota and metabolites, predict nutrient bioavailability, and formulate synergistic configuration schemes. The proactive nutrition intervention execution module is used to generate the optimal intervention plan, control the nutrition delivery equipment to perform precise interventions, and make real-time adjustments.

2. A nutritional intervention method for cancer patients undergoing radiotherapy and chemotherapy, characterized in that, Includes the following steps: Step S1: Detect patient gene polymorphisms, gut microbiota composition and blood metabolites, and integrate multi-omics data to generate a personalized molecular nutrition fingerprint vector; Step S2: Run a bidirectional LSTM network to predict the trend of changes in the patient's nutritional status, calculate the probability of malnutrition, and perform risk classification; Step S3: Run the molecular docking algorithm to calculate the binding energy between nutrients and chemotherapy drug metabolites, and optimize the nutrient ratio; Step S4: Construct a microbial community and metabolite association network to predict nutrient bioavailability and optimize the synergistic configuration of nutrients and probiotics; Step S5: Generate the optimal intervention plan by combining the above analysis results, control the delivery equipment to perform nutritional intervention and collect effect data for feedback optimization.

3. The nutritional intervention system according to claim 1, characterized in that, The molecular nutrient fingerprinting module employs genome-wide association analysis, 16S rRNA gene sequencing, and liquid chromatography-mass spectrometry.

4. The nutritional intervention system according to claim 1, characterized in that, The spatiotemporal prediction modeling module includes a bidirectional LSTM network, which dynamically adjusts the prediction weights using a self-attention mechanism.

5. The nutritional intervention system according to claim 1, characterized in that, The drug-nutrient synergistic optimization module uses the AutoDock algorithm to identify interaction pairs between nutrients and drug metabolites, and uses a genetic algorithm to optimize the nutrient ratio.

6. The nutritional intervention system according to claim 1, characterized in that, The gut microbiota-host metabolism modeling module establishes a random forest regression model to predict nutrient bioavailability.

7. The nutritional intervention system according to claim 1, characterized in that, The active nutrition intervention execution module uses the NSGA-II algorithm to generate the optimal solution set and achieves precise nutrient delivery through a programmable infusion pump and intelligent capsule system.

8. The nutritional intervention method according to claim 2, characterized in that, Steps S2 to S4 are executed in parallel based on the molecular nutrient fingerprint vector generated in step S1.

9. The nutritional intervention method according to claim 2, characterized in that, Step S5 periodically collects nutritional index data and feeds it back to step S2, forming a closed-loop optimization mechanism.

Images