Intelligent prediction system of energy-protein intake in critical care based on digital twin

Abstract

The present application discloses an intelligent prediction system for energy-protein intake of critical illness based on digital twin, which relates to the field of digital twin technology, including a patient characteristic acquisition module, a twin model update module, a candidate solution acquisition module, a nutrition support prediction module and a nutrition solution recommendation module; the patient characteristic acquisition module is used to acquire the physiological characteristics of the target patient to obtain the target physiological characteristics; the twin model update module is used to update the target digital twin model in real time according to the target physiological characteristics; the candidate solution acquisition module is used to acquire a preset nutrition solution; the nutrition support prediction module is used to input the preset nutrition solution into the current target digital twin model to obtain the predicted state of the first patient; the nutrition solution recommendation module is used to generate a recommended nutrition solution according to the current physiological characteristics of the target patient, the preset nutrition solution and the predicted state of the first patient, so as to improve the degree of adaptation between the metabolic state and the nutrition intake of the patient at each stage.

Description

Technical Field

[0001] The present application relates to the field of digital twin technology, and specifically to an intelligent prediction system for critical energy-protein intake based on digital twins. Background Art

[0002] The failure of critically ill patients to receive reasonable nutritional support therapy is one of the important factors leading to adverse outcomes such as prolonged ICU hospitalization, difficulty in weaning off the ventilator, and even death. Insufficient energy intake of patients will lead to accumulated energy debt, aggravate protein decomposition, and further lead to muscle wasting of patients, while excessive energy supply will cause blood sugar to rise in patients, aggravate stress hyperglycemia and insulin resistance. Insufficient protein intake of patients will cause malnutrition, ICU-acquired weakness, etc., while excessive protein intake will put tremendous pressure on the liver and kidneys, and lead to problems such as electrolyte imbalance and intestinal microbial imbalance. Therefore, the best nutritional therapy needs to provide patients with nutritional support with an intake close to the actual nutritional consumption, in order to adapt to the patient's pathophysiological indicators and improve the patient's nutritional metabolic indicators.

[0003] Current studies have shown that the metabolic characteristics of critically ill patients vary significantly at different stages of the disease. Taking the 2018 European Society of Nutrition and Metabolism staging of critically ill patients as an example, patients in the "early stage of the acute phase" are characterized by metabolic instability and severe increase in catabolism, while patients in the "late stage of the acute phase" have significant muscle atrophy and metabolic stabilization.

[0004] Traditional patient disease staging is mainly based on observing the patient's clinical manifestations and medical history, relying on the subjective experience of medical workers for rough estimation. However, this method is highly subjective and has low assessment accuracy. Although existing guidelines have recommended target energy and protein intake for critically ill patients, due to differences in genetic, immune and metabolic backgrounds of critically ill patients, even the clinical manifestations of the same disease are difficult to be exactly the same, which increases the difficulty for medical workers to formulate nutritional treatment plans. Summary of the invention

[0005] The purpose of the present invention is to solve the technical problem of how to formulate personalized nutritional support plans for critically ill patients during the treatment process. It provides an intelligent prediction system for critical energy-protein intake based on digital twins, which formulates highly adaptable nutritional support plans according to the patient's specific physiological characteristics and the stage of the disease, so as to achieve the technical effect of reducing the risk of poor prognosis for patients.

[0006] The present invention claims protection for a critical energy-protein intake intelligent prediction system based on digital twins, comprising a patient feature acquisition module, a twin model update module, a candidate solution acquisition module, a nutrition support prediction module, and a nutrition solution recommendation module;

[0007] The patient characteristic acquisition module is used to acquire physiological characteristics of a target patient to obtain target physiological characteristics; the physiological characteristics include static variable characteristics and dynamic variable characteristics;

[0008] The twin model updating module is used to update the target digital twin model of the target patient in real time according to the target physiological characteristics;

[0009] The candidate plan acquisition module is used to acquire a preset nutrition plan; the nutrition plan includes energy intake and protein intake;

[0010] The nutritional support prediction module is used to input each of the preset nutritional plans into the current target digital twin model to obtain the corresponding predicted state of the first patient;

[0011] The nutrition plan recommendation module is used to generate a recommended nutrition plan based on the current physiological characteristics of the target patient, each of the preset nutrition plans and the corresponding predicted state of the first patient.

[0012] In one embodiment of the present application, the twin model update module also includes a reference patient retrieval submodule and a real-time model training submodule; the reference patient retrieval submodule is used to query the first database according to the target physiological characteristics of the target patient at different times to obtain the corresponding first reference patient, and the similarity between the physiological characteristics of the first reference patient and the corresponding target physiological characteristics is not greater than a first preset threshold; the real-time model training submodule is used to update the target digital twin model in real time according to the target physiological characteristics and the corresponding physiological characteristics of all the first reference patients, and the initial state of the target digital twin model is obtained by transfer learning of the standard digital twin model based on the initial physiological characteristics of the target patient and the corresponding physiological characteristics of the first reference patient.

[0013] In one embodiment of the present application, the standard digital twin model includes a first digital twin sub-model, and the first digital twin sub-model includes an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer is used to input the physiological characteristics of the patient; the first hidden layer is used to extract the spatial feature vector of the target patient's corresponding space-time graph, and the second hidden layer is used to extract the time feature vector of the target patient's corresponding space-time graph; the output layer is used to obtain the predicted state of the first patient according to the spatial feature vector and the time feature vector;

[0014] The spatiotemporal graph includes graph structures corresponding to several moments, each graph structure includes nodes and edges, the nodes are obtained according to the physiological characteristics, and the edges are obtained according to the correlation between different physiological characteristics.

[0015] In one embodiment of the present application, all the dynamic variable features are divided into several feature sets according to the human body system, all the static variable features correspond to one feature set, each node of the space-time graph represents the corresponding feature set, and the edge represents the association relationship between different feature sets.

[0016] In one embodiment of the present application, the standard digital twin model is used to simulate the physiological characteristic values ​​of the patient at different times, and the dynamic variable characteristics include blood glucose concentration, plasma insulin concentration and insulin injection rate. The standard digital twin model also includes a second digital twin sub-model, and the second digital twin sub-model constructs a patient's blood glucose and insulin change prediction model according to a dynamic modeling method. The second digital twin sub-model includes:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] Where t represents the time, X(t) represents the rate of change of the patient's blood glucose concentration at the tth time, G(t) represents the patient's blood glucose concentration at the tth time, I(t) represents the patient's plasma insulin concentration at the tth time, Q(t) represents the rate of change of the patient's plasma insulin concentration at the tth time, and u ex (t) represents the insulin injection rate of the patient at time t, k1, k2, k3, k4, k5, k6, k7, α, basal and h represent the kinetic model parameters.

[0022] In one embodiment of the present application, the nutrition plan recommendation module also includes a plan generation submodule, and the plan generation submodule is used to obtain the recommended nutrition plan based on the output of the plan generation model of all the preset nutrition plans and the corresponding first patient predicted state; wherein the plan generation model is constructed based on a large language model, and a plan generation data set is obtained based on the physiological characteristics of all historical patients before using the nutrition plan, the actual nutrition plan used, the physiological characteristics after using the nutrition plan, and the corresponding standard nutrition plan, and the plan generation model is supervised and fine-tuned based on the plan generation data set.

[0023] In one embodiment of the present application, the nutrition plan recommendation module also includes a generation model update submodule, which is used to obtain the actual nutrition plan and the actual status of the target patient, obtain the plan difference based on the comparison result of the actual nutrition plan and the recommended nutrition plan, obtain the recovery status of the target patient after using the actual nutrition plan, construct a reward function of the reinforcement learning model based on the plan difference and the recovery status, and update the plan generation model according to the reinforcement learning model.

[0024] In one embodiment of the present application, the nutritional support prediction module also includes a low nutrition warning submodule, which is used to input the target physiological characteristics into a low nutrition probability prediction model to obtain a target probability of malnutrition in the target patient. When the target probability is greater than a first preset threshold, a first feedback information is generated. The first feedback information is used to remind medical staff to intake energy and / or protein for the target patient.

[0025] In one embodiment of the present application, the nutrition plan recommendation module also includes a plan screening submodule, which is used to obtain the output of the plan generation model to obtain a candidate nutrition plan and determine whether the candidate nutrition plan is greater than 1. If the number of candidate nutrition plans is greater than 1, each candidate nutrition plan is input into the current target digital twin model to obtain the corresponding second patient prediction state, and the second patient prediction state at each moment is input into the low nutrition probability prediction model to obtain the corresponding low nutrition probability. The moment when the low nutrition probability is greater than the first preset threshold is taken to obtain the corresponding supplementation moment, and the candidate nutrition plan corresponding to the maximum value of all the supplementation moments is taken as the recommended nutrition plan.

[0026] In one embodiment of the present application, the dynamic variable characteristics also include whether symptoms of gastrointestinal intolerance occur, the protein intake includes the protein intake provided by enteral nutrition support and the protein intake provided by parenteral nutrition support, and the energy intake includes the energy intake provided by enteral nutrition support and the energy intake provided by parenteral nutrition support.

[0027] This application has the following beneficial effects:

[0028] 1. Through multimodal and multi-scale digital twin technology, an individualized monitoring, analysis and prediction model is established according to the patient's real-time physiological characteristics to simulate the changing trend of the patient's physiological characteristics during the treatment process, so as to improve the accuracy of the assessment of the patient's disease stage. At the same time, when predicting the energy-protein intake required by the patient, the energy-protein intake is adjusted by "titration", and the intake is accurately adjusted according to the impact of different preset nutritional plans on the patient's case physiological state, so that it matches the patient's metabolic needs, and improves the matching degree between the patient's metabolic state and nutritional intake at each stage, which not only effectively improves the accuracy and positive impact of nutritional intervention, but also reduces the mortality rate of critically ill patients.

[0029] 2. The standard digital twin model is trained based on all the physiological characteristics of all historical patients in the historical medical records. The standard digital twin model is trained based on the physiological characteristics of all historical patients, so that it can comprehensively learn and extract the metabolic mechanism characteristics shared by patients with different physiological characteristics. The standard digital twin model is fine-tuned based on the target physiological characteristics of the target patient, so that the target digital twin model not only retains the general characteristics of the metabolic mechanism, but also combines the individual differences of the target patient, so that the prediction results can more accurately reflect the personalized physiological characteristics of the target patient and improve the prediction accuracy. Especially for rare cases, due to the complexity of the disease and significant individual differences, the amount of information on the patient's physiological characteristics is limited. Through transfer learning, the model parameters are trained and fine-tuned on the basis of the standard digital twin model, which improves the generalization ability and accuracy of the model.

[0030] 3. Establish a spatiotemporal graph of the patient based on physiological characteristics. The spatiotemporal graph not only describes the interaction between different physiological characteristics, but also reflects the changing trend of each physiological characteristic of the same patient at different stages of the disease. The first digital twin sub-model constructed according to the STGNN network model extracts feature vectors of the spatiotemporal graph from the spatial and temporal dimensions respectively, and trains the first digital twin sub-model based on the physiological characteristics of historical patients, so that the first digital twin sub-model can explore the correlation between different physiological characteristics at different stages of the disease, capture the patient's state fluctuations during the evolution of the disease, significantly improve the patient's state prediction accuracy and generalization ability, and effectively support the personalized optimization of clinical diagnosis and treatment decisions.

[0031] 4. Establishing the nodes of the space-time graph according to the human body system can reduce the number of nodes in the space-time graph, so as to reduce the amount of calculation for each forward propagation and reverse propagation, thereby achieving the technical effect of reducing the overall computational complexity of the digital twin model, improving computational efficiency and reducing the computational content occupied.

[0032] 5. A blood glucose and insulin change prediction model constructed according to the kinetic modeling method is added to the digital twin model, so that the digital twin model can predict the patient's blood glucose and insulin concentrations according to different nutritional plans. The second digital twin sub-model can clearly explain the causal relationship between the prediction results and the input conditions, thereby enhancing the interpretability of the digital twin model. This highly interpretable digital twin model provides medical personnel with a more intuitive understanding, improves the reliability of the digital twin model, and promotes the implementation of personalized treatment and precision medicine.

[0033] 6. The regimen generation model obtained according to the large model can obtain an optimized nutrition plan based on the preset nutrition plan and the corresponding predicted state of the first patient. On the one hand, the regimen generation model integrates the professional optimization experience of multidisciplinary experts in optimizing nutrition plans through pre-training and supervised fine-tuning stages, which reduces the difficulty of using the system. It can assist clinicians with different professional levels and different subject areas to judge the energy-protein intake required by the patient. On the other hand, the regimen generation model judges whether the preset plan is suitable for the patient based on the patient's state after using the preset nutrition plan, so that it can intelligently generate personalized nutrition plans based on the patient's individual physiological characteristics and disease course characteristics, improve the degree of adaptation of the nutrition plan to the patient, so as to achieve personalized nutrition plan recommendations that adapt to the patient's disease course, and achieve the technical effect of optimizing treatment effects, improving recovery rates, and reducing adverse prognoses.

[0034] 7. The regimen recommendation model updated and adjusted based on reinforcement learning constructs a reward function according to the difference between the nutritional regimen actually adopted by the clinician and the nutritional regimen output by the regimen recommendation model, as well as the patient's treatment outcomes, so that the regimen recommendation model can automatically adjust its recommendation strategy according to the feedback from the clinician and the needs of individual patients. This adaptive optimization mechanism ensures that the model can adapt to changes in patient status and the development of medicine, and continuously absorb new clinical feedback and data, thereby improving the reliability of the system and improving the treatment effects of patients.

[0035] 8. The low nutrition warning submodule uses a low nutrition probability prediction model to accurately predict the timing of nutrition intake for critically ill patients, reducing the risk of patients accumulating energy debt due to untimely nutrition intake.

[0036] 9. The current target digital twin model is used to simulate the physiological characteristics of the target patient after using each candidate nutrition plan, and the low nutrition probability prediction model is used to determine the next time the target patient needs to take energy-protein after using the corresponding candidate nutrition plan, that is, the replenishment time. Taking the candidate nutrition plan corresponding to the maximum value of the replenishment time as the recommended nutrition plan is conducive to reducing the workload of medical staff on the basis of ensuring the normal physiological metabolism of the target patient.

[0037] 10. Based on the simulation of the patient's disease course changes by the target digital twin model, corresponding preset nutrition plans are set according to different nutrition support pathways. The plan generation model generates recommended protein intake and energy intake through enteral nutrition support according to the first patient predicted state after the target patient uses different preset nutrition plans, as well as protein intake and energy intake provided by parenteral nutrition support. It assists medical staff in adjusting the appropriate nutrition support pathway according to the patient's state, so as to assist patients with enteral nutrition intolerance to transition from parenteral nutrition support to enteral nutrition support, thereby improving the quality of patient prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0039] Figure 1 This is a schematic diagram of the structure of a critical energy-protein intake intelligent prediction system based on digital twins involved in an embodiment of the present application;

[0040] Figure 2 A flowchart of an implementable method for generating a recommended nutrition plan according to an embodiment of the present application;

[0041] Figure 3 A schematic diagram of a space-time diagram corresponding to the first digital twin model involved in an embodiment of the present application;

[0042] Figure 4 This is a schematic diagram of the structure of a standard digital twin model involved in an embodiment of the present application;

[0043] Figure 5 A schematic diagram of an update process of a solution generation model involved in an embodiment of the present application;

[0044] Figure 6 A schematic diagram of the structure of an electronic device involved in an embodiment of the present application;

[0045] Symbols in the figure: o_j represents the graph structure at time t_j, o_j+1 represents the graph structure at time t_ j+1, o_j+2 represents the graph structure at time t_ j+2, and o_j+3 represents the graph structure at time t_ j+3. DETAILED DESCRIPTION

[0046] The present invention provides a critical energy-protein intake intelligent prediction system based on digital twins. In order to make the above-mentioned purposes, features and advantages of the present application more obvious and understandable, the technical scheme in the embodiment of the present application will be clearly and completely described in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, with reference to the terms "one embodiment", "some embodiments", "implementation", "embodiment", "illustrative embodiment", "example", "specific example" or "some examples", the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but only indicates that the specific features, structures or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. And the specific features, structures or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0047] It should be noted that similar reference numerals and letters represent similar items in the following figures, so once an item is defined in one figure, it does not need to be further defined and explained in the subsequent figures. At the same time, in the description of this application, the terms "first", "second" and other relational terms are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0048] The present invention claims protection for a critical illness energy-protein intake intelligent prediction system based on digital twins, with reference to the attached Figure 1 and attached Figure 2 As shown, it includes a patient feature acquisition module, a twin model update module, a candidate plan acquisition module and a nutrition plan recommendation module, but of course it is not limited to this.

[0049] In this embodiment, the patient feature acquisition module includes a static feature acquisition submodule. The static feature acquisition submodule is used to obtain static variable features of the target patient. The static variable features include age, gender, personal medical history, diagnosis, medical history records and genetic test results, etc., of course, it is not limited to this. The target patient is a patient who will be predicted for energy-protein intake.

[0050] In this embodiment, the patient feature acquisition module also includes a dynamic feature acquisition submodule. The dynamic special diagnosis acquisition submodule is used to obtain the dynamic variable features of the target patient in real time. The dynamic variable features include the use of parenteral enteral nutrition support therapy drugs, circulatory function characteristics, respiratory dynamics characteristics, microcirculation and tissue perfusion characteristics, gastrointestinal function characteristics, glucose metabolism characteristics, liver and kidney function characteristics, coagulation function characteristics, infection-related indicator characteristics, organ function evaluation results, and metabolic network characterization characteristics, of course, it may not be limited to this.

[0051] It should be noted that the use of parenteral enteral nutrition support therapy drugs includes the name of the enteral enteral nutrition drug, the dosage of the drug, the start and end time of the drug, the flow rate of tube feeding or intravenous infusion, the historical energy intake and the historical protein intake, which may not be limited to this. The circulatory function characteristics include the continuous cardiac output monitoring cardiac output index (CI) based on the pulse wave indicator, the intrathoracic blood volume index (ITBI), the global diastolic volume index (GEDI), the global ejection fraction (GEF), the extravascular lung water index (EVLWI), the pulmonary vascular permeability index (PVPI), the pulse indicated cardiac output index (PCCI), the stroke volume index (SVI), the stroke volume variability index (SVV), the pulse pressure variability rate (PPV), the arterial systolic pressure (Apsys), the arterial diastolic pressure (Apdia), the mean arterial pressure (MAP), the maximum pressure increase rate (dPmax) and the peripheral vascular resistance index (SVRI), which may not be limited to this. The respiratory dynamics characteristics include tidal volume (VT), minute ventilation, positive end-inspiratory pressure (PEEP), oxygen partial pressure (PO2), carbon dioxide partial pressure (PCO2) and oxygen saturation (SpO2), of course, it may not be limited thereto. The microcirculation and tissue perfusion characteristics include arterial blood lactate, which can sensitively respond to tissue hypoxia. The gastrointestinal function characteristics can be obtained by quantitatively detecting serum citrulline levels and gastrointestinal-related clinical manifestations by high-performance liquid phase quadrupole tandem mass spectrometry (UPLC-TQD) at 7 a.m. every day. The glucose metabolism characteristics include blood glucose concentration, plasma insulin concentration and insulin injection rate, such as blood glucose data per minute recorded by a continuous glucose monitoring (CGM) system and a patch used in conjunction with the CGM system. The liver and kidney function characteristics include serum creatinine (Cr), glomerular filtration rate, alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), γ-glutamyl transpeptidase (GGY), total bilirubin (TB), direct bilirubin (DB), prealbumin (PA), albumin (Alb), cholinesterase (CHE) and daily urine volume and intake, of course, it may not be limited to this. The coagulation function characteristics include prothrombin time (PT), activated partial thromboplastin time (APTT), thrombin time (TT), fibrinogen (FIB), activated whole blood clotting time (ACT) and platelet count (PLT), of course, it may not be limited to this. The infection-related indicator characteristics include peripheral blood leukocyte and neutrophil counts, procalcitonin (PCT), C-reactive protein (C-RP), interleukin 6 (IL-6), TNF-a and T lymphocyte subsets, of course, it may not be limited to this. The organ function evaluation results can be obtained based on the daily SOFA score.The characterization characteristics of the metabolic network can be obtained by collecting blood samples from patients after an overnight fast using a blood collection tube without anticoagulants, taking the upper serum after centrifugation and placing it in an EP tube and storing it in a -80°C refrigerator, and then obtaining the differential metabolites found using a high performance liquid chromatography-quadrupole tandem mass spectrometer (UPLC-TQD).

[0052] In this embodiment, the twin model update module is used to update the target digital twin model of the target patient in real time according to the target physiological characteristics of the target patient. The target digital twin model is used to simulate the real-time state of the target patient. The target physiological characteristics include static variable characteristics and dynamic variable characteristics of the target patient, but of course it may not be limited to this.

[0053] In this embodiment, before predicting the energy-protein intake of the target patient, an initial digital twin model corresponding to the target patient is constructed according to the initial physiological characteristics of the target patient. The initial physiological characteristics of the target patient are obtained. The initial physiological characteristics are the values ​​of each physiological characteristic of the patient before the first prediction of the energy-protein intake is made. The initial digital twin model is used to simulate the initial state of the target patient. The first database is queried according to the initial physiological characteristics of the target patient to obtain a first reference patient. The similarity between the physiological characteristics of the first reference patient and the initial physiological characteristics of the target patient is not greater than a first preset threshold. For example, if the patient needs to start predicting the energy-protein intake after admission, the initial physiological characteristics can be the values ​​of each physiological characteristic of the patient at the time of admission, or it can be a time series data composed of the values ​​of each physiological characteristic of the patient from admission to the start of the first prediction.

[0054] It should be noted that the first database includes all the physiological characteristics of historical patients during the treatment process in the historical medical records, but it is certainly not limited to this. The feature vectors of each physiological characteristic of each historical patient and the initial physiological characteristics of the target patient can be extracted separately through the neural network model; the cosine similarity between the feature vector of each historical patient and the feature vector of the target patient is calculated respectively, and the historical medical records are screened according to the comparison results of each cosine similarity and the first preset threshold to obtain the first reference patient. The specific value of the first preset threshold can be obtained by pre-setting. If the initial feature is time series data composed of various physiological characteristics, a 3D convolutional neural network, a recurrent neural network, etc. can be used to extract the feature vector of the time series data.

[0055] In this embodiment, the standard digital twin model is trained for transfer learning based on the physiological characteristics of the first reference patient and the initial physiological characteristics of the target patient to obtain the initial state of the target digital twin model of the target patient. When the dynamic feature acquisition submodule acquires the new physiological characteristics of the target patient, the twin model update module is used to update the current target digital twin model based on the new physiological characteristics of the target patient. The update training method of the target digital twin model is similar to the training method of the initial digital twin model, and will not be described in detail here.

[0056] It should be noted that the standard digital twin model is trained based on all physiological characteristics of all historical patients in the historical medical records. The standard digital twin model is trained based on the physiological characteristics of all historical patients, so that it can comprehensively learn and extract the metabolic mechanism characteristics shared by patients with different physiological characteristics. The standard digital twin model is fine-tuned based on the target physiological characteristics of the target patient, so that the target digital twin model not only retains the general characteristics of the metabolic mechanism, but also combines the individual differences of the target patient, so that the prediction results can more accurately reflect the personalized physiological characteristics of the target patient and improve the prediction accuracy. Especially for rare cases, due to the complexity of the disease and significant individual differences, the amount of information on the patient's physiological characteristics is limited. Through transfer learning, the parameters of the model are trained and fine-tuned on the basis of the standard digital twin model, which improves the generalization ability and accuracy of the model.

[0057] In this embodiment, the standard digital twin model includes a first digital twin sub-model. The first digital twin sub-model is used to predict physiological processes with unknown physiological mechanisms, and the first digital twin sub-model can be constructed according to a machine learning model.

[0058] It should be noted that the first digital twin model can be constructed based on a machine learning model. For example, a neural network model is trained based on historical patients to obtain physiological characteristics that can be used to simulate patients under different conditions.

[0059] In this embodiment, the first digital twin model is constructed according to the spatiotemporal graph neural network model. Figure 3 As shown, a spatiotemporal graph is constructed based on the physiological characteristics. In the graph at the same time, the edges between different nodes are used to describe the causal relationship or strong correlation between the nodes. In the graph at different times, the node features of the same node are used to describe the node features of the patient at the corresponding time.

[0060] In a feasible implementation, when constructing the spatiotemporal graph, each of the physiological characteristics is used as a node in the graph, and the associations between different physiological characteristics are used as edges between corresponding nodes.

[0061] It should be noted that the values ​​of static variable features are the same at different times, while the values ​​of dynamic variable features may be the same or different at different times.

[0062] It should be noted that the training data set of the second digital twin model is obtained based on the physiological characteristics of historical patients during treatment in the historical medical records. The training data set can be divided into a training set and a test set according to a preset division ratio, the training set is used to train the first digital twin model, and the test set is used to test the performance indicators of the second digital twin model.

[0063] It should be noted that a neural network layer, such as a fully connected layer, can be selected as the state update function. The feature vector of the spatiotemporal graph can be extracted by using Graph Convolutional Network (GCN), Graph Recurrent Network (GRN) and Graph Attention Network (GAT). In the process of extracting the spatial features of the patient's physiological characteristics, the node v has a hidden state at the t+1th time. The calculation methods include:

[0064] ;

[0065] Where f represents the state update function of the hidden state, that is, the local transfer function; x v Represents the node features of node v; The features of all edges adjacent to the node v can be obtained by presetting according to existing medical knowledge, or by fitting during model training by GCN, GRN or GAT; Represents the hidden state of the neighbor node of node v at time t; Represents the node features of all neighbor nodes of node v.

[0066] In this embodiment, the temporal features of the spatiotemporal graph are extracted by the transformer-XL model. The specific process includes: dividing the time series of physiological features corresponding to each historical patient into multiple subsequences; encoding the position of each subsequence according to the time sequence; extracting the feature vector of the temporal features from each subsequence in turn according to the time sequence; fusing the temporal features extracted from each subsequence with the temporal features of the previous subsequence to obtain the temporal feature vector of the subsequence, so that the first digital twin model can learn the long-distance dependency of the physiological features of the patient at different times.

[0067] It should be noted that the first digital twin model also includes an output layer, which is used to convert the feature vector of the spatiotemporal graph extracted by the first digital twin model into the output of the model. The output layer can be preset according to the output result of the digital twin model of the patient. For example, when the first patient prediction state output by the digital twin model is a classification task of whether the patient will have a poor prognosis, the corresponding output layer can select a linear layer containing a softmax activation function to achieve the mapping output from the feature vector to the output result. The difference between the predicted physiological characteristics and the real physiological characteristics is calculated according to the preset loss function, and the first digital twin sub-model is trained according to the gradient descent algorithm.

[0068] It should be noted that a space-time graph of the patient is established based on physiological characteristics. The space-time graph not only describes the interaction between different physiological characteristics, but also reflects the changing trend of each physiological characteristic of the same patient at different stages of the disease. The first digital twin sub-model constructed according to the STGNN network model extracts feature vectors of the space-time graph from the spatial dimension and the time dimension respectively, and trains the first digital twin sub-model according to the physiological characteristics of historical patients, so that the first digital twin sub-model can explore the correlation between different physiological characteristics at different stages of the disease, capture the patient's state fluctuations during the evolution of the disease, significantly improve the patient's state prediction accuracy and generalization ability, and effectively support the personalized optimization of clinical diagnosis and treatment decisions.

[0069] In a feasible embodiment, the physiological features are divided into 5 feature sets according to the dynamic variable features, including a circulation-respiration feature set, a liver and kidney function feature set, a metabolic feature set, an intestinal feature set and a nutritional therapy feature set, but it may not be limited thereto. The static variable feature is used as a feature set, and the 6 feature sets are used as corresponding nodes in the space-time graph. The indicators corresponding to each feature set are used as node features of the corresponding nodes. For example, the node features of the liver and kidney function feature set include serum creatinine (Cr), glomerular filtration rate, alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), γ-glutamyl transpeptidase (GGY), total bilirubin (TB), direct bilirubin (DB), prealbumin (PA), albumin (Alb), cholinesterase (CHE) and daily urine volume and intake, but it may not be limited thereto. The edges between nodes are obtained according to the relationship between different feature sets. For example, the relationship between the liver and kidney function feature set and the nutritional therapy feature set.

[0070] It should be noted that the specific division method and number of the feature set can be pre-set according to the specific indicators of the dynamic variables to determine the node features specifically contained in each node. For example, the circulation-respiration feature set includes the circulation function characteristics, the respiratory dynamics characteristics, the microcirculation and tissue perfusion characteristics, the coagulation function characteristics, and the organ function evaluation results related to the respiratory circulation, of course, it may not be limited to this; the liver and kidney function characteristics include the liver and kidney function characteristics and the organ function evaluation results related to the liver and kidney functions, of course, it may not be limited to this; the metabolic feature set includes the infection-related indicator characteristics, the metabolic network characterization characteristics, and the organ function evaluation results related to metabolism, of course, it may not be limited to this; the intestinal feature set includes the use of parenteral and enteral nutrition support therapy drugs, gastrointestinal function characteristics, and the organ function evaluation results related to the intestine, of course, it may not be limited to this; the nutritional therapy feature set includes the sugar metabolism characteristics and the organ function evaluation results related to nutritional therapy, of course, it may not be limited to this.

[0071] In this embodiment, the method of updating the hidden state of the node during the feature extraction process of the spatiotemporal graph and the training method of the second digital twin model are similar to the above method and will not be described in detail here.

[0072] It should be noted that since the space-time graph neural network model transmits information through edges when extracting feature vectors in the spatial dimension, establishing the nodes of the space-time graph according to the system function can reduce the number of nodes in the space-time graph. The reduction in the number of nodes in the space-time graph means a reduction in the number of edges in the space-time graph. The number of nodes involved in each round of information transmission of the corresponding space-time graph neural network model will also be reduced, reducing the amount of computation for each forward and backward propagation, thereby achieving the technical effect of reducing the amount of computation in each round and the overall computational complexity, improving computational efficiency and reducing the computational content occupancy.

[0073] In a possible implementation, refer to the attached Figure 4 As shown, the standard digital twin model also includes a second digital twin model. The second digital twin model is used to predict physiological processes with known physiological mechanisms, and the second digital twin model can be constructed by dynamic modeling. Taking the changes in blood glucose and insulin under stressful conditions with known physiological mechanisms as an example, the second digital twin model for predicting the patient's blood glucose level can be constructed based on the Bergman minimum model.

[0074] In this embodiment, the second digital twin model corresponding to the changes in blood sugar and insulin under stress state includes:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] Where t represents the time, X(t) represents the rate of change of the patient's blood glucose concentration at the tth time, G(t) represents the patient's blood glucose concentration at the tth time, I(t) represents the patient's plasma insulin concentration at the tth time, Q(t) represents the rate of change of the patient's plasma insulin concentration at the tth time, and u ex (t) represents the insulin injection rate of the patient at time t, k1, k2, k3, k4, k5, k6, k7, α, basal and h represent the kinetic model parameters.

[0080] It should be noted that the parameters of the kinetic model are fitted according to the physiological characteristics of historical patients. The blood glucose and insulin change prediction model constructed according to the kinetic modeling method is added to the digital twin model, so that the digital twin model predicts the patient's blood glucose and insulin concentrations according to different nutritional plans. The second digital twin sub-model can clearly explain the causal relationship between the prediction results and the input conditions, enhancing the interpretability of the digital twin model. This highly interpretable digital twin model provides medical personnel with a more intuitive understanding, improves the reliability of the digital twin model, and promotes the implementation of personalized treatment and precision medicine.

[0081] In a feasible embodiment, the candidate solution acquisition module is used to acquire a preset nutritional solution. The preset nutritional solution can be obtained by presetting a plurality of different dosages of protein-energy combinations. For example, preset nutritional solution 1 includes protein intake of A1 and protein intake of B1, and preset nutritional solution 2 includes protein intake of A2 and protein intake of B2.

[0082] In this embodiment, the nutritional support prediction module is used to input each of the preset nutritional plans into the current target digital twin model respectively to obtain the corresponding first patient predicted state.

[0083] It should be noted that the predicted state of the first patient can be the prediction result of the target digital twin model on the target patient's various physiological indicators after using the corresponding preset nutritional plan. It can also be an evaluation result of the prediction result output by the target digital twin model, such as a scoring result of the target patient's state. Or it can be a classification result of the state of the target patient after using the corresponding preset nutritional plan based on the prediction result output by the target digital twin model, such as whether the target patient will have a poor prognosis, and whether the prognosis grading result of the target patient is excellent, good or poor.

[0084] In this embodiment, the recommended plan generation module is used to generate a recommended nutritional plan based on all the preset nutritional plans and the corresponding first patient predicted status.

[0085] It should be noted that the recommended solution generation module can filter out the recommended nutrition solution from all the preset nutrition solutions. Take the predicted state of the first patient as the prediction result of each physiological indicator of the target patient after the target digital twin model uses the corresponding preset nutrition solution as an example. The recommended solution generation module can score the prediction results of each target digital twin model. The higher the score, the better the prognosis of the patient. The preset nutrition solution corresponding to the maximum value of the scoring result is taken as the recommended nutrition solution. The recommended solution generation module can also use a preset scoring threshold to take all the preset nutrition solutions whose scoring results are not less than the scoring threshold as the recommended nutrition solution. The recommended solution generation module can also be other feasible methods.

[0086] In this embodiment, through multimodal and multi-scale digital twin technology, an individualized monitoring, analysis and prediction model is established according to the patient's real-time physiological characteristics to simulate the changing trend of the patient's physiological characteristics during the treatment process, so as to improve the accuracy of the assessment of the patient's disease stage. At the same time, when predicting the energy-protein intake required by the patient, the energy-protein intake is adjusted in a "titration" manner, and the intake is accurately adjusted according to the impact of different preset nutritional plans on the patient's case physiological state, so that it matches the patient's metabolic needs, and improves the matching degree between the patient's metabolic state and nutritional intake at each stage, which not only effectively improves the accuracy and positive impact of nutritional intervention, but also reduces the mortality rate of critically ill patients.

[0087] In this embodiment, the recommended solution generation module inputs all the preset nutrition solutions and the corresponding predicted state of the first patient into a solution generation model, and the solution generation model is constructed and trained according to a large prediction model, and the large language model is obtained, for example, according to the GPT series, BERT, T5, etc. The solution generation model optimizes all preset nutrition solutions according to the corresponding predicted state of the first patient to obtain the recommended nutrition solution.

[0088] It should be noted that the solution generation model can be constructed based on a large language model, and text data can be collected from the Internet to obtain a text corpus, and the solution generation model can be pre-trained based on the text prediction library. A standard nutrition plan is manually set according to the status of each historical patient. The physiological characteristics of each historical patient before using the nutrition plan, the actual nutrition plan used, and the physiological characteristics after using the nutrition plan are used as input features of the corresponding samples, the corresponding standard nutrition plan is used as the real nutrition plan, and the output features corresponding to each sample are used as the predicted nutrition plan. The pre-trained solution generation model is supervised and fine-tuned according to the real nutrition plan and the predicted nutrition plan, so that the solution generation model can optimize the nutrition plan for the patient according to the patient's physiological characteristics and the nutrition plan used.

[0089] It should be noted that the number of recommended nutrition plans output by the plan generation model is an integer greater than 1, that is, the plan generation model can output only one recommended nutrition plan, or can output multiple recommended nutrition plans. The number of the standard nutrition plans corresponding to each of the historical patients can be an integer not less than 0. If the standard nutrition plan corresponding to the historical patient is 0, the corresponding nutrition plan actually used is used as the standard nutrition plan.

[0090] In this embodiment, the regimen generation model obtained according to the large model can obtain an optimized nutrition regimen according to the preset nutrition regimen and the corresponding predicted state of the first patient. On the one hand, the regimen generation model integrates the professional optimization experience of multidisciplinary experts in optimizing nutrition regimens through pre-training and supervised fine-tuning stages, which reduces the difficulty of using and operating the system, and can assist clinicians with different professional levels and different subject areas to judge the energy-protein intake required by the patient. On the other hand, the regimen generation model judges whether the preset regimen is suitable for the patient according to the patient's state after using the preset nutrition regimen. If there is a preset regimen suitable for the patient, the corresponding preset nutrition regimen is used as the recommended nutrition regimen. If all preset regimens are not suitable for the patient, a new nutrition regimen is obtained after optimizing the preset regimen, and the new nutrition regimen is used as the recommended nutrition regimen, thereby intelligently generating a personalized nutrition regimen according to the individual physiological characteristics and disease course characteristics of the patient, improving the degree of adaptation of the nutrition regimen to the patient, so as to achieve personalized nutrition regimen recommendations adapted to the patient's disease course, and achieve the technical effects of optimizing treatment effects, improving recovery rates, and reducing adverse prognoses.

[0091] In this embodiment, refer to the attached Figure 5As shown, the nutrition regimen recommendation module also includes a generation model update submodule, and the generation model update submodule is used to update the regimen generation model through a reinforcement learning model. The regimen generation model is used as an intelligent agent, and the actual state change of the target patient is used as an environment. The reward function of the reinforcement learning model is constructed based on the comparison result between the actual nutrition regimen of the target patient and the recommended nutrition regimen output by the regimen generation model and the recovery status of the target patient.

[0092] It should be noted that the difference in the plans can be obtained based on whether the actual nutrition plan is consistent with the recommended nutrition plan; the distance between the actual nutrition plan and the recommended nutrition plan can also be calculated based on the Euclidean distance, or other feasible methods can be used. The recovery status can be obtained by manually judging the status label of the target patient after using the actual nutrition plan through a preset status label; the recovery status can also be calculated based on the Euclidean distance according to the target patient's physiological characteristics before using the actual nutrition plan, the physiological characteristics after using the actual nutrition plan, and the expected standard physiological characteristics.

[0093] In this embodiment, when the actual nutrition plan is consistent with the recommended nutrition plan, that is, the clinician adopts the nutrition plan recommended by the plan generation model, if the target patient's recovery is improved or cured, it means that the suggestion of the plan generation model is correct, and the reward function rewards the plan generation model. When the actual nutrition plan is inconsistent with the recommended nutrition plan, that is, the clinician does not adopt the nutrition plan recommended by the plan generation model, if the target patient's recovery is improved or cured, it means that the suggestion of the plan generation model is wrong, and the reward function punishes the plan generation model. If the target patient's recovery is no obvious change, deterioration or death, the rationality of the actual nutrition plan and the recommended nutrition plan can be evaluated by medical experts or evaluation auxiliary systems, and the reward function is preset according to the recovery situation and the rationality to reward or punish the plan, and the corresponding numerical value.

[0094] It should be noted that the specific value of the reward and the specific value of the penalty given by the reward function to the scheme generation model can be obtained by presetting. For example, when the actual nutrition scheme is consistent with the recommended nutrition scheme and the recovery of the target patient is improved, the reward value given by the reward function to the numerical value of the scheme generation model is 1; when the actual nutrition scheme is consistent with the recommended nutrition scheme and the recovery of the target patient is cured, the reward value given by the reward function to the numerical value of the scheme generation model is 100; when the actual nutrition scheme is inconsistent with the recommended nutrition scheme and the recovery of the target patient is improved or cured, the reward value given by the reward function to the scheme generation model is -1. When the recovery of the target patient has no obvious change, if the rationality of the actual nutrition scheme is greater than the rationality of the recommended nutrition scheme, the reward value given by the reward function to the scheme generation model is -1; if the rationality of the actual nutrition scheme is not greater than the rationality of the recommended nutrition scheme, the reward value given by the reward function to the scheme generation model is 1. When the target patient's recovery condition is deteriorating, if the rationality of the actual nutrition plan is greater than that of the recommended nutrition plan, the reward value given by the reward function to the plan generation model is -2; if the rationality of the actual nutrition plan is not greater than that of the recommended nutrition plan, the reward value given by the reward function to the plan generation model is 0. When the target patient's recovery condition is death, the reward value given by the reward function to the plan generation model is -100.

[0095] It should be noted that a value function can be constructed based on the Q-Learning algorithm, Sarsa algorithm or policy gradient algorithm and the reward, and the solution generation model can be updated and trained in a single step according to the value function to achieve continuous optimization of the nutrition solution recommendation strategy.

[0096] It should be noted that the regimen recommendation model updated and adjusted based on reinforcement learning constructs a reward function according to the difference between the nutritional regimen actually adopted by clinicians and the nutritional regimen output by the regimen recommendation model, as well as the patient's treatment outcomes, so that the regimen recommendation model can automatically adjust its recommendation strategy according to the feedback from clinicians and the needs of individual patients. This adaptive optimization mechanism ensures that the model can adapt to changes in patient status and the development of medicine, and continuously absorb new clinical feedback and data, thereby improving the reliability of the system and improving the treatment effects of patients.

[0097] In a feasible embodiment, the nutrition support prediction module also includes a low nutrition warning submodule, which is used to input the target physiological characteristics into a low nutrition probability prediction model to obtain a target probability of malnutrition in the target patient. When the target probability is greater than a first preset threshold, a first feedback information is generated. The first feedback information is used to remind medical staff to take in energy and / or protein for the target patient.

[0098] It should be noted that the low nutrition probability prediction model can be constructed through a machine learning model, and a low nutrition label can be set for each physiological state of historical patients during the treatment process. For example, the physiological characteristics of historical patients with malnutrition are used as positive samples, that is, the probability of malnutrition in the positive samples is 1, and the physiological characteristics of historical patients without malnutrition are used as negative samples, that is, the probability of malnutrition in the negative samples is 0. The low nutrition probability prediction model is trained according to the difference between the probability of malnutrition of the patient output by the low nutrition probability prediction model and the probability of malnutrition corresponding to the low nutrition label.

[0099] It should be noted that the specific value of the first preset threshold can be obtained by manual preset.

[0100] In this embodiment, the low nutrition warning submodule realizes accurate prediction of the timing of nutrition intake for critically ill patients through the low nutrition probability prediction model, thereby reducing the risk of patients accumulating energy debt due to untimely nutrition intake.

[0101] In this embodiment, the nutrition plan recommendation module also includes a plan screening submodule, which is used to obtain the output of the plan generation model to obtain a candidate nutrition plan and determine whether the candidate nutrition plan is greater than 1. If the number of candidate nutrition plans is greater than 1, each candidate nutrition plan is input into the current target digital twin model to obtain the corresponding second patient prediction state, and each of the second patient prediction states is input into the low nutrition probability prediction model to obtain the corresponding low nutrition probability. The moment when the low nutrition probability is greater than the first preset threshold is taken to obtain the corresponding supplementation moment, and the candidate nutrition plan corresponding to the maximum value of all the supplementation moments is taken as the recommended nutrition plan.

[0102] It should be noted that when the solution generation model outputs multiple candidate nutrition solutions, it means that multiple energy-protein intake combinations meet the physiological characteristics and disease stage requirements of the target patients. On the basis of ensuring the normal physiological metabolism of critically ill patients, the less the energy-protein intake, the higher the frequency of energy-protein intake that medical staff need to perform for the target patients. The current target digital twin model simulates the physiological characteristics of the target patient after using each of the candidate nutrition solutions, and the low nutrition probability prediction model is used to determine the next time the target patient needs to perform energy-protein intake after using the corresponding candidate nutrition solution, that is, the supplementation time. Taking the candidate nutrition solution corresponding to the maximum value of the supplementation time as the recommended nutrition solution is beneficial to reducing the workload of medical staff on the basis of ensuring the normal physiological metabolism of the target patient.

[0103] It should be noted that the scheme screening submodule also includes obtaining a preset supplement time. The preset supplement time includes the preset time for the target patient to supplement energy-protein next time. For example, if the target patient currently supplements energy-protein at 8:00 and the next time to supplement energy-protein is 12:00, then the preset supplement time corresponding to the target patient is 12:00. The scheme screening submodule inputs each candidate nutrition scheme into the current target digital twin model to obtain the corresponding second patient prediction state. The second patient prediction state is screened according to the preset supplement time to obtain the low nutrition probability of the target patient after using the corresponding candidate nutrition scheme and before the preset supplement time. The recommended nutrition scheme is obtained according to the comparison result of the mean of the low nutrition probability of the target patient after using the corresponding candidate nutrition scheme and before the preset supplement time and the first preset threshold. The candidate nutrition scheme whose mean of low nutrition probability is greater than the first preset threshold is taken as the recommended nutrition scheme, which assists critically ill patients to regularly take in energy-protein during the recovery process while reducing the nursing workload of medical staff.

[0104] In a feasible embodiment, the dynamic variable feature also includes whether symptoms of gastrointestinal intolerance occur, the protein intake includes the protein intake provided by enteral nutrition support and the protein intake provided by parenteral nutrition support, and the energy intake includes the energy intake provided by enteral nutrition support and the energy intake provided by parenteral nutrition support. The enteral nutrition support is a method of providing nutrition through the gastrointestinal tract. The parenteral nutrition support is to provide nutrition to the patient through intravenous infusion.

[0105] In this embodiment, based on the target digital twin model simulating the changes in the patient's disease course, corresponding preset nutrition plans are set according to different nutrition support pathways. The plan generation model generates recommended protein intake and energy intake through enteral nutrition support according to the first patient predicted state of the target patient after using different preset nutrition plans, as well as protein intake and energy intake provided by parenteral nutrition support, to assist medical staff in adjusting the appropriate nutrition support pathway according to the patient's state, so as to assist patients with enteral nutrition intolerance to transition from parenteral nutrition support to enteral nutrition support, thereby improving the quality of patient prognosis.

[0106] See attached Figure 6 As shown, an embodiment of the present application provides an electronic device, including: a processor and a memory, the processor and the memory are interconnected and communicate with each other through a communication bus and / or other forms of connection mechanisms (not shown), the memory stores a computer program executable by the processor, and when the computing device is running, the processor executes the computer program to execute the system in any optional implementation of the above embodiments.

[0107] The embodiment of the present application provides a storage medium, and when the computer program is executed by the processor, the system in any optional implementation of the above embodiment is executed. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, disk or optical disk.

[0108] In the embodiments provided in the present application, it should be understood that the disclosed system can be implemented in other ways. The system embodiments described above are merely schematic. For example, the division of the modules is only a logical function division, and can be implemented in another way. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some communication interface, system or unit, which can be electrical, mechanical or other forms.

[0109] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0110] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0111] Flowcharts are used herein to illustrate the steps of the method of the embodiments of the present disclosure. It should be understood that the preceding or following steps are not necessarily performed precisely in order. On the contrary, various steps may be evaluated in reverse order or simultaneously. At the same time, other operations may also be added to these processes.

[0112] Unless otherwise defined, all terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present disclosure belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an idealized or highly formal sense, unless explicitly defined as such herein.

[0113] The above is a detailed introduction to the critical energy-protein intake intelligent prediction system based on digital twins. This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only an embodiment of this application, which is only used to help understand the critical energy-protein intake intelligent prediction system based on digital twins of this application, and is not used to limit the scope of protection of this application; at the same time, for those skilled in the art, this application may have various changes and variations. Any modifications and equivalent substitutions made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. The intelligent prediction system of energy-protein intake for critical illness based on digital twin is characterized by: It includes patient feature acquisition module, twin model update module, candidate solution acquisition module, nutrition support prediction module and nutrition solution recommendation module; The patient characteristic acquisition module is used to acquire physiological characteristics of the target patient to obtain target physiological characteristics; the physiological characteristics include static variable characteristics and dynamic variable characteristics; the dynamic variable characteristics include blood glucose concentration, plasma insulin concentration and insulin injection rate; The twin model update module is used to update the target digital twin model of the target patient in real time according to the target physiological characteristics; the initial state of the target digital twin model is obtained by transfer learning the standard digital twin model according to the initial physiological characteristics of the target patient and the physiological characteristics of the corresponding first reference patient; The first reference patient includes a patient in the first database whose similarity between the physiological characteristics and the initial physiological characteristics is not greater than a first preset threshold; The standard digital twin model is used to simulate the physiological characteristic values ​​of the patient at different times, and includes a second digital twin sub-model; The second digital twin model constructs a patient's blood sugar and insulin change prediction model based on a kinetic modeling method, including: ； ； ； ； Where t represents the time, X(t) represents the rate of change of the patient's blood glucose concentration at the tth time, G(t) represents the patient's blood glucose concentration at the tth time, I(t) represents the patient's plasma insulin concentration at the tth time, Q(t) represents the rate of change of the patient's plasma insulin concentration at the tth time, and u ex (t) represents the insulin injection rate of the patient at time t, k1, k2, k3, k4, k5, k6, k7, α, basal and h represent the kinetic model parameters; The candidate plan acquisition module is used to acquire a preset nutrition plan; the nutrition plan includes energy intake and protein intake; The nutritional support prediction module is used to input each of the preset nutritional plans into the current target digital twin model to obtain the corresponding predicted state of the first patient; The nutrition plan recommendation module is used to generate a recommended nutrition plan based on the current physiological characteristics of the target patient, each of the preset nutrition plans and the corresponding predicted state of the first patient.

2. According to claim 1, the intelligent prediction system for critical energy-protein intake based on digital twin is characterized in that: The twin model update module also includes a reference patient retrieval submodule and a real-time model training submodule; the reference patient retrieval submodule is used to query the first database according to the target physiological characteristics of the target patient at different times to obtain a corresponding first reference patient, and the similarity between the physiological characteristics of the first reference patient and the corresponding target physiological characteristics is not greater than a first preset threshold; The real-time model training submodule is used to update the target digital twin model in real time according to the target physiological characteristics and the corresponding physiological characteristics of all the first reference patients.

3. The critical care energy-protein intake intelligent prediction system based on digital twin according to claim 2 is characterized in that: The standard digital twin model includes a first digital twin sub-model, and the first digital twin sub-model includes an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer is used to input the physiological characteristics of the patient; the first hidden layer is used to extract the spatial feature vector of the space-time graph corresponding to the target patient, and the second hidden layer is used to extract the temporal feature vector of the space-time graph corresponding to the target patient; the output layer is used to obtain the predicted state of the first patient according to the spatial feature vector and the temporal feature vector; The spatiotemporal graph includes graph structures corresponding to several moments, each graph structure includes nodes and edges, the nodes are obtained according to the physiological characteristics, and the edges are obtained according to the correlation between different physiological characteristics.

4. The critical energy-protein intake intelligent prediction system based on digital twin according to claim 3 is characterized in that: All the dynamic variable features are divided into several feature sets according to the human body system, all the static variable features correspond to one feature set, each node of the space-time graph represents a corresponding feature set, and the edge represents the association relationship between different feature sets.

5. The critical care energy-protein intake intelligent prediction system based on digital twin according to any one of claims 1 to 4, characterized in that: The nutrition plan recommendation module also includes a plan generation submodule, which is used to obtain the recommended nutrition plan based on the output of the plan generation model of all the preset nutrition plans and the corresponding predicted state of the first patient; wherein the plan generation model is constructed based on a large language model, and a plan generation data set is obtained based on the physiological characteristics of all historical patients before using the nutrition plan, the nutrition plan actually used, the physiological characteristics after using the nutrition plan and the corresponding standard nutrition plan, and the plan generation model is supervised and fine-tuned based on the plan generation data set.

6. The critical energy-protein intake intelligent prediction system based on digital twin according to claim 5 is characterized in that: The nutrition plan recommendation module also includes a generation model update submodule, which is used to obtain the actual nutrition plan and the actual status of the target patient, obtain the plan difference based on the comparison result of the actual nutrition plan and the recommended nutrition plan, obtain the recovery status of the target patient after using the actual nutrition plan, construct a reward function of the reinforcement learning model based on the plan difference and the recovery status, and update the plan generation model according to the reinforcement learning model.

7. The critical energy-protein intake intelligent prediction system based on digital twin according to claim 5 is characterized in that: The nutritional support prediction module also includes a low nutrition warning submodule, which is used to input the target physiological characteristics into a low nutrition probability prediction model to obtain a target probability of malnutrition in the target patient. When the target probability is greater than a first preset threshold, a first feedback information is generated. The first feedback information is used to remind medical staff to intake energy and / or protein for the target patient.

8. The critical care energy-protein intake intelligent prediction system based on digital twin according to claim 7 is characterized in that: The nutrition plan recommendation module also includes a plan screening submodule, which is used to obtain the output of the plan generation model to obtain a candidate nutrition plan and determine whether the candidate nutrition plan is greater than 1. If the number of candidate nutrition plans is greater than 1, each candidate nutrition plan is input into the current target digital twin model to obtain the corresponding second patient prediction state, and the second patient prediction state at each moment is input into the low nutrition probability prediction model to obtain the corresponding low nutrition probability. The moment when the low nutrition probability is greater than the first preset threshold is taken to obtain the corresponding supplementation moment, and the candidate nutrition plan corresponding to the maximum value of all the supplementation moments is taken as the recommended nutrition plan.

9. The critical energy-protein intake intelligent prediction system based on digital twin according to claim 6 is characterized in that: The dynamic variable characteristics also include whether symptoms of gastrointestinal intolerance occur, the protein intake includes the protein intake provided by enteral nutrition support and the protein intake provided by parenteral nutrition support, and the energy intake includes the energy intake provided by enteral nutrition support and the energy intake provided by parenteral nutrition support.

Images