Artificial intelligence-based dynamic decision system and method for critical patient nutrition

Abstract

The application is a critical patient nutrition full-process dynamic decision system and method based on artificial intelligence, comprising: multi-source heterogeneous data acquisition and integration, hierarchical dynamic nutrition risk assessment (three-level architecture of preliminary screening-diagnosis-dynamic re-evaluation), evidence-based nutrition starting decision, dynamic monitoring and dose adjustment based on deep reinforcement learning, five-step nutrition intervention two-way intelligent switching, and graded early warning and report output. Each step operation is executed by a multi-model fusion artificial intelligence decision engine including a four-layer architecture. The scheme can convert critical nutrition guidelines into standardized calculation rules that can be automatically executed, realizing real-time dynamic identification of nutrition risk, standardized starting decision, automatic warning of re-feeding syndrome, and full-process closed-loop dynamic management. It is expected to improve the individualization and precision of critical patient nutrition support.

Description

Technical Field

[0001] This invention relates to the field of critical care nutrition support and artificial intelligence-assisted decision-making technology, specifically to a dynamic decision-making system and method for the entire process of nutrition for critically ill patients based on artificial intelligence. Background Technology

[0002] Malnutrition occurs in 30% to 50% of critically ill patients, significantly prolonging total hospital stay and ICU stay, and significantly increasing the risk of complications such as infection and ICU-acquired weakness. Furthermore, approximately 75% of critically ill patients experience feeding intolerance during enteral nutrition, leading to nutritional interruption, further exacerbating the negative energy-protein balance, prolonging ICU stay, and increasing mortality. Nutritional support is a crucial aspect of comprehensive critical care treatment and directly impacts clinical prognosis.

[0003] However, current clinical practice lacks standardized dynamic decision-making tools covering the entire process of "nutritional risk screening—initiation decision—tolerance monitoring—step-by-step switching," making it difficult to achieve individualized and precise nutritional management. Currently, clinical nutritional support management faces the following technical challenges: First, nutritional risk screening and assessment relies on manual periodic operations, which cannot meet the guidelines' requirement for standardized, dynamic, and real-time assessment within 24 to 48 hours of admission, and makes it difficult to accurately capture the continuous fluctuations in patients' nutritional status.

[0004] Second, the timing of initiating critical care nutrition, the intervention method, and the selection of initial dosage are highly dependent on the physician's clinical experience. There is a lack of quantitative and standardized intelligent decision-making support tools, which makes it difficult to efficiently implement the concept of individualized nutritional therapy.

[0005] Third, the nutritional support lacks closed-loop dynamic management throughout the process. It fails to achieve intelligent automatic parameter adjustment based on the refined standards of low-calorie, low-protein diet in the acute phase and gradual incremental diet in the recovery phase, resulting in low frequency and delayed response of manual adjustments and insufficient feeding accuracy and timeliness.

[0006] Fourth, there is a lack of automated early warning and coordinated intervention system for refeeding syndrome, which makes it impossible to monitor key indicators such as blood phosphorus in real time and automatically limit energy when they exceed the standard, posing a serious safety hazard.

[0007] Fifth, the clinically accepted five-step nutritional ladder treatment principle lacks digital implementation tools, making it impossible to dynamically monitor energy target achievement rates or intelligently trigger step switching, resulting in a poor match between intervention plans and patients' actual nutritional needs.

[0008] Sixth, existing AI nutrition models are fragmented, limited to risk prediction and dosage recommendation in a single feeding stage, and have not yet built an integrated intelligent decision-making system covering the entire process of screening, assessment, initiation, monitoring, adjustment and step-by-step switching, making it difficult to adapt to the standardized needs of modern critical care precision nutrition. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems existing in the prior art and provide an artificial intelligence-based dynamic decision-making system and method for the entire process of nutrition for critically ill patients. It constructs an integrated AI closed-loop decision-making system covering screening, assessment, initiation, monitoring, dynamic dosage adjustment, tiered switching, and risk warning. It includes seven collaborative functional modules, transforming critical care nutrition guidelines into standardized calculation rules that can be automatically executed, thus solving the defects of existing systems such as fragmentation, human lag, subjective decision-making, and lack of automated warning and tiered control.

[0010] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: An AI-based dynamic decision-making system for the entire nutritional process of critically ill patients, comprising: The multi-source heterogeneous data acquisition and integration module is used to connect with hospital information systems, laboratory information systems, electronic medical record systems and bedside monitoring equipment to collect multi-dimensional clinical data of patients in real time. The stratified dynamic nutrition risk assessment module adopts a three-tier architecture of initial screening layer, diagnosis layer and dynamic reassessment layer, which is used to automatically perform nutrition risk screening and malnutrition diagnosis, and periodically refresh the risk stratification. The evidence-based nutrition initiation decision module is used to output standardized decisions, including initiation timing, intervention pathway, and initial dose, based on the quantitative rules of the built-in critical care nutrition guidelines. The dynamic monitoring and dose adjustment module is used to periodically output personalized dose adjustment schemes based on a deep reinforcement learning framework. The five-step nutrition intervention and management module is used to digitally implement the five-step nutrition diagnosis and treatment path, lock the patient's current intervention step in real time, and realize bidirectional intelligent switching between ascending and descending steps. The tiered early warning and report output module is used to generate multi-level risk real-time early warnings and daily nutritional diagnosis and treatment reports; The multi-model fusion artificial intelligence decision engine is connected to the hierarchical dynamic nutrition risk assessment module, the evidence-based nutrition initiation decision module, the dynamic monitoring and dosage adjustment module, and the five-step nutrition intervention and control module, respectively, and is used to perform calculations. The multi-model fusion artificial intelligence decision engine includes a multi-source data fusion encoding layer, a multi-prediction model integration layer, a guideline rule computable transformation layer, and an interpretable output layer.

[0011] Furthermore, in the stratified dynamic nutritional risk assessment module, the initial screening layer automatically performs NRS2002 scoring and mNUTRIC scoring. When the NRS2002 score is greater than or equal to 3 or the mNUTRIC score is greater than or equal to 5, the in-depth assessment is automatically triggered. The diagnostic layer automatically verifies the diagnostic criteria for malnutrition according to the GLIM standard. The dynamic reassessment layer refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours.

[0012] Furthermore, the dynamic monitoring and dosage adjustment module outputs a personalized dosage adjustment plan every 4 hours based on the DeepEN framework, including automatic handling of gastrointestinal tolerance, automated early warning linkage for refeeding syndrome, dosage adjustment linked to inflammation and glucose metabolism, and automatic determination of energy target achievement. Among them, the automated early warning linkage for refeeding syndrome monitors blood phosphorus index in real time. When blood phosphorus is less than 0.65 mmol / L, it immediately issues a graded warning and simultaneously recommends reducing energy to 20 kcal / h and supplementing with multivitamins.

[0013] Accordingly, the present invention also provides an artificial intelligence-based dynamic decision-making method for the entire nutritional process of critically ill patients, the method comprising: The process involves collecting and integrating multi-source heterogeneous data, connecting with hospital information systems, laboratory information systems, electronic medical record systems, and bedside monitoring equipment to gather multi-dimensional clinical data from patients in real time. The stratified dynamic nutritional risk assessment process adopts a three-tier architecture of initial screening, diagnosis, and dynamic reassessment. It automatically performs nutritional risk screening and malnutrition diagnosis, and periodically updates the risk stratification. Evidence-based nutrition initiation decision-making steps, based on built-in critical care nutrition guidelines and quantitative rules, output standardized decisions including initiation timing, intervention pathway, and initial dose; The dynamic monitoring and dosage adjustment process periodically outputs personalized dosage adjustment plans based on a deep reinforcement learning framework. The five-step nutritional intervention and management steps are digitally implemented to create a five-step nutritional diagnosis and treatment path, which can lock the patient's current intervention step in real time and realize two-way intelligent switching between step advancement and step demotion. The system includes a tiered early warning and reporting output process, generating multi-level real-time risk warnings and daily nutritional diagnosis and treatment reports. The calculations of the hierarchical dynamic nutritional risk assessment step, the evidence-based nutritional initiation decision step, the dynamic monitoring and dosage adjustment step, and the five-step nutritional intervention and control step are all executed by a multi-model fusion artificial intelligence decision engine.

[0014] Furthermore, the stratified dynamic nutritional risk assessment steps include: the initial screening layer automatically performs NRS2002 scoring and mNUTRIC scoring, and automatically triggers in-depth assessment when the NRS2002 score is greater than or equal to 3 or the mNUTRIC score is greater than or equal to 5; the diagnostic layer automatically verifies the diagnostic criteria for malnutrition according to the GLIM standard, the diagnostic criteria including positive screening, at least one phenotypic indicator and at least one causal indicator of inflammation or insufficient intake; the dynamic reassessment layer refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours.

[0015] Furthermore, in the evidence-based nutrition initiation decision-making steps: the initiation timing is determined based on hemodynamics, lactate levels, and gastrointestinal function. For hemodynamically stable patients, enteral nutrition is initiated early within 48 hours of ICU admission. For patients in shock and using high-dose vasoactive drugs, initiation is delayed or parenteral nutrition is preferred. The intervention pathway strictly follows the five-step nutrition ladder to determine the intervention plan step by step. The initial dose follows the acute-phase protective feeding standard, with 6 to 8 kcal / kg / day of calories and 0.2 to 0.9 g / kg / day of protein in the first week, using a low initial gradient incremental strategy.

[0016] Furthermore, the dynamic monitoring and dosage adjustment steps, based on the DeepEN framework, output personalized dosage adjustment plans every 4 hours, including: automatic handling of gastrointestinal tolerance, monitoring signs of feeding intolerance and automatically recommending deceleration, continuous infusion, formula change, or suspension of feeding; automated early warning linkage for refeeding syndrome, real-time monitoring of serum phosphorus levels, and immediate graded warning when serum phosphorus is less than 0.65 mmol / L, simultaneously suggesting a reduction in energy intake to 20 kcal / h and supplementation with multivitamins; dosage adjustment linked to inflammation and glucose metabolism, dynamically adjusting calorie supply based on insulin infusion rate, CRP, and IL-6; and automatic determination of energy target achievement, calculating the daily intake as a percentage of the target energy, and automatically pushing tiered escalation suggestions when it is below 60% for 2 to 3 consecutive days.

[0017] Furthermore, the five-step nutritional intervention and management steps include: real-time locking of the patient's current intervention step; step-up determination, automatically generating an upgrade plan from oral nutrition supplementation to tube enteral nutrition to supplemental parenteral nutrition to total parenteral nutrition when the intake is less than 60% of the target energy for several consecutive days; and step-down determination, continuously monitoring intestinal function recovery indicators, and automatically determining the timing of resuming enteral nutrition even when the patient is in the total parenteral nutrition stage.

[0018] Furthermore, in the multi-model fusion artificial intelligence decision engine: the multi-source data fusion encoding layer uses deep learning to unify heterogeneous clinical data features; the multi-prediction model integration layer integrates the XGBoost model to predict early malnutrition, the NutriSighT Transformer model to predict insufficient enteral feeding, and the Dueling Double DQN reinforcement learning model to output individualized energy and protein targets; the guideline rule calculation conversion layer embeds the critical care nutrition guideline quantification threshold; and the interpretable output layer synchronously outputs the basis indicators, guideline clauses, and feature weights for each decision.

[0019] Furthermore, the tiered early warning and reporting output steps include: three-level real-time risk early warning, divided into three levels: alert, warning, and critical; automatic generation of daily nutrition diagnosis and treatment reports, including intake, target achievement rate, step status, and adjustment suggestions; and a large language model clinical decision support submodule, which transforms indicators and guidelines into standardized operational suggestions.

[0020] The beneficial effects of this invention are: 1. This invention, through a three-tiered architecture (initial screening—confirmation—dynamic reassessment) of real-time acquisition of multi-source heterogeneous data and hierarchical dynamic nutritional risk assessment, can achieve continuous dynamic monitoring of nutritional risks. This helps to solve the problem of lag in manual periodic assessment and meets the requirements of guidelines for standardized, dynamic, and real-time assessment within 24 to 48 hours of hospital admission.

[0021] 2. This invention transforms critical care nutrition guidelines into computable decision rules through an evidence-based nutrition initiation decision module. It can provide standardized and interpretable initiation decision support, including initiation timing, intervention pathway, and initial dose. This helps reduce over-reliance on physicians' clinical experience and efficiently implement the concept of individualized nutrition therapy.

[0022] 3. This invention uses the automated early warning and linkage logic of refeeding syndrome in the dynamic monitoring and dosage adjustment module to monitor blood phosphorus index in real time and immediately issue a graded warning when blood phosphorus is less than 0.65 mmol / L, and simultaneously recommend that energy be reduced to 20 kcal / h and supplement with multivitamins. This enables early identification and proactive intervention of refeeding syndrome, which helps to reduce the safety risks caused by refeeding syndrome.

[0023] 4. This invention, through an integrated closed-loop architecture covering the entire process from screening, assessment, initiation, monitoring to adjustment, combined with intelligent assessment and adjustment every 4 hours, can achieve closed-loop dynamic management of the entire nutritional support process. The design is expected to improve the individualization and precision of nutritional support for critically ill patients. The above effects are design expectations and await clinical validation.

[0024] 5. This invention embeds the five-step nutritional intervention and control module into the system decision-making logic, which integrates the five-step principle and the 60% energy target achievement principle. It can automatically track the step status and trigger bidirectional intelligent switching between step advancement and step demotion, which helps to improve the matching degree between the intervention plan and the patient's actual nutritional needs.

[0025] 6. This invention integrates guideline rules with multimodal prediction models through a four-layer architecture of a multi-model fusion artificial intelligence decision engine, and synchronously outputs decision-making basis through an interpretable output layer, which helps to improve the clinical acceptability of decisions and avoids the functional fragmentation defects of existing AI nutrition models. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of the critical care nutrition artificial intelligence closed-loop decision-making system of the present invention; Figure 2 This is a schematic diagram of the four-layer architecture of the multi-model fusion artificial intelligence decision engine of the present invention; Figure 3This is a flowchart illustrating the AI-powered automatic tiered nutritional ladder decision-making logic of the present invention. Figure 4 This is a flowchart of the closed-loop control method for the entire critical care nutrition process of the present invention. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] like Figure 1 As shown, this embodiment provides an artificial intelligence dynamic decision-making system for the entire process of critical care nutrition, including seven collaborative functional modules: a multi-source heterogeneous data acquisition and integration module, a hierarchical dynamic nutrition risk assessment module, an evidence-based nutrition initiation decision-making module, a dynamic monitoring and dosage adjustment module, a five-step nutrition intervention and control module, a multi-model fusion artificial intelligence decision engine, and a graded early warning and report output module.

[0029] The system hardware consists of an ICU monitoring terminal, a gastrointestinal data collection terminal, a Laboratory Information System (LIS) interface, a bedside tablet, and an intelligent nutrition infusion pump. The software includes a data preprocessing module, an AI-powered nutritional risk assessment module, a five-tiered decision engine, a tolerance monitoring module, an AI-driven adaptive treatment adjustment module, a daily target achievement review module, and an early warning and quality control storage module.

[0030] The multi-source heterogeneous data acquisition and integration module interfaces with the Hospital Information System (HIS), Laboratory Information System (LIS), Electronic Medical Record System (EMR), and bedside monitoring equipment, aggregating six categories of multidimensional evidence-based data in real time: The first category is basic anthropometric and medical history data; the second category is continuous vital signs data; the third category is a complete set of laboratory indicator data, with a focus on dynamically tracking serum phosphorus levels; the fourth category is critical care and nutritional score data, including APACHE II score, SOFA score, NRS2002 score, and mNUTRIC score; the fifth category is organ and ventilation therapy parameters; and the sixth category is records of feeding methods, infusion rates, intake, and gastrointestinal tolerance. These six categories of data provide a continuous underlying data stream for end-to-end intelligent computation.

[0031] The stratified dynamic nutritional risk assessment module follows the guidelines' mandatory assessment requirement of 24 to 48 hours after admission and adopts a three-tiered automated assessment architecture. The initial screening tier automatically performs NRS2002 and mNUTRIC scoring; when the NRS2002 score is greater than or equal to 3 or the mNUTRIC score is greater than or equal to 5, in-depth assessment is automatically triggered. The diagnostic tier automatically verifies the diagnostic criteria for malnutrition according to the GLIM standard, specifically a positive screening result plus at least one phenotypic indicator plus at least one inflammatory or insufficient intake etiology indicator. The dynamic reassessment tier refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours, achieving uninterrupted nutritional status monitoring and replacing the lag problem of manual periodic assessment.

[0032] The evidence-based nutrition initiation decision-making module incorporates quantitative rules from critical care nutrition guidelines and outputs three-dimensional standardized decisions through a machine learning model. Initiation timing: Based on hemodynamics, lactate levels, and gastrointestinal function, enteral nutrition is initiated early within 48 hours of ICU admission for hemodynamically stable patients. Parenteral nutrition is delayed or preferred for patients in shock and using high-dose vasoactive drugs. Intervention pathway: Intervention plans are determined strictly according to the five-step nutrition ladder. Initial dosage: Acute-phase protective feeding standards are implemented, with 6-8 kcal / kg / day of calories and 0.2-0.9 g / kg / day of protein in the first week, using a low-start, gradient-increase strategy.

[0033] The dynamic monitoring and dosage adjustment module, based on the DeepEN deep reinforcement learning framework, outputs personalized dosage adjustment plans every 4 hours, integrating four layers of safety and metabolic control logic. The first layer is automatic handling of gastrointestinal intolerance: monitoring signs of feeding intolerance such as abdominal distension, vomiting, and increased gastric residual volume, and automatically recommending slowing down, continuous infusion, formula change, or temporary suspension of feeding. The second layer is automated early warning linkage for refeeding syndrome: real-time monitoring of serum phosphorus levels; when serum phosphorus is less than 0.65 mmol / L, an immediate tiered warning is issued, simultaneously recommending an energy intake reduction to 20 kcal / h and supplementation with multivitamins, matching the consensus requirements for refeeding in severe cases. The third layer is dosage adjustment linked to inflammation and glucose metabolism: dynamically adjusting calorie supply based on insulin infusion rate, CRP, and IL-6. The fourth layer is automatic energy target achievement determination: calculating the daily intake as a percentage of the target energy; if it is below 60% for 2 to 3 consecutive days, automatically pushing tiered escalation suggestions.

[0034] The five-step nutritional intervention and management module digitally implements the five-step nutritional treatment pathway, enabling bidirectional intelligent switching. This module continuously monitors the patient's current intervention step. The upgrade logic automatically generates an upgrade plan from oral nutrition supplementation (ONS) to enteral nutrition via tube feeding (EN), to supplemental parenteral nutrition (SPN), and finally to total parenteral nutrition (TPN) when the patient's energy intake is less than 60% of the target for several consecutive days. The downgrade logic continuously monitors intestinal function recovery indicators, automatically determining the timing for resuming enteral nutrition even during the total parenteral nutrition stage, avoiding long-term reliance on intravenous nutrition.

[0035] like Figure 2As shown, the multi-model fusion AI decision engine employs a four-layer architecture to integrate guideline rules with multimodal predictions. The multi-source data fusion encoding layer uses deep learning to unify heterogeneous clinical data features, transforming clinical data from different systems and formats into a unified feature representation. The multi-prediction model integration layer integrates three prediction models: the XGBoost model for predicting early malnutrition risk; the NutriSighT Transformer model for predicting enteral feeding inadequacy risk; and the Dueling Double DQN reinforcement learning model for outputting individualized energy and protein targets. The guideline rule computational conversion layer embeds all quantitative thresholds from critical care nutrition guidelines, transforming qualitative recommendations in the guidelines into automatically executable computational rules. The interpretable output layer synchronously outputs the supporting indicators, guideline clauses, and feature weights for each decision, enabling clinicians to understand the source and basis of the decision and ensuring clinical reliability.

[0036] The tiered early warning and reporting output module comprises three functional parts. The first part provides real-time early warning for three levels of risk: alert, warning, and critical, triggering corresponding notification mechanisms for different risk levels. The second part automatically generates daily nutritional treatment reports, including the patient's daily intake, energy target achievement rate, current status in the nutritional ladder, and adjustment suggestions provided by the system. The third part is a large language model-based clinical decision support submodule, which transforms complex laboratory indicators and guideline provisions into standardized operational suggestions that medical staff can directly implement.

[0037] like Figure 4 As shown, the AI-powered dynamic decision-making method for the entire critical care nutrition process in this embodiment includes the following steps: Step S1: Data acquisition. The multi-source heterogeneous data acquisition and integration module acquires multi-dimensional clinical data of patients in real time, including basic anthropometric measurements and medical history, continuous vital signs, a full set of laboratory indicators, critical illness and nutritional scores, organ and ventilation treatment parameters, and feeding-related records.

[0038] Step S2: Nutritional risk identification. Within 24 to 48 hours after admission, the stratified dynamic nutritional risk assessment module automatically completes the initial screening using NRS2002 and mNUTRIC scores. If the initial screening is positive (NRS2002 score greater than or equal to 3 or mNUTRIC score greater than or equal to 5), proceed to step S3.

[0039] Step S3: Nutritional Assessment. The system automatically completes the mNUTRIC score review and GLIM standard assessment, verifies the diagnostic criteria for malnutrition according to the GLIM standard, and determines the patient's nutritional risk stratification and malnutrition diagnosis.

[0040] Step S4: Initiation decision: The multi-model fusion artificial intelligence decision engine comprehensively assesses the patient's hemodynamic status, lactate level, gastrointestinal function and nutritional risk stratification, and automatically generates initiation method suggestions according to the five-step principle, including initiation timing, intervention path and initial dose.

[0041] Step S5: Protocol Execution. After the clinician reviews and confirms the system-generated initiation decision, a detailed nutrition implementation plan is generated, including parameters such as nutritional formula, infusion rate, target calories and protein content, and is executed through an intelligent nutrition infusion pump.

[0042] Step S6: Dynamic monitoring. The system automatically collects monitoring data every 4 hours to assess the patient's gastrointestinal tolerance and metabolic status, including key indicators such as blood phosphorus, blood glucose, insulin infusion rate, CRP, and IL-6.

[0043] Step S7: Intelligent Adjustment. Based on dynamic monitoring data, a deep reinforcement learning framework automatically generates adjustment suggestions, including dose adjustment, infusion rate adjustment, formula change, and step-by-step switching suggestions. When serum phosphorus is detected to be less than 0.65 mmol / L, the system immediately triggers a graded warning for refeeding syndrome, and simultaneously suggests reducing energy intake to 20 kcal / h and supplementing with multivitamins.

[0044] Step S8: Daily Assessment. The system generates a daily nutritional target achievement rate report, calculates the proportion of the patient's daily intake to the target energy, and assesses whether a step-up switch is triggered. If the intake is below 60% of the target energy for 2 to 3 consecutive days, the system automatically pushes a step-up suggestion; if intestinal function is detected to have recovered, the system automatically determines when to downgrade the step.

[0045] Step S9: Closed-loop feedback. The system forms a closed-loop management of "monitoring-assessment-decision-adjustment". The dynamic review layer refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours to ensure that the nutritional support plan is continuously optimized as the patient's condition changes.

[0046] like Figure 3 As shown, the decision-making logic for escalation and escalation of the five-step nutritional intervention is as follows: The system monitors the patient's energy intake target achievement rate and intestinal function status in real time. When the patient's energy intake is less than 60% of the target for several consecutive days, the system automatically determines that an escalation is needed, generating an escalation plan in the order of oral nutrition supplementation (ONS) → enteral nutrition via tube feeding (EN) → supplemental parenteral nutrition (SPN) → total parenteral nutrition (TPN). When the system continuously monitors improvements in the patient's intestinal function recovery indicators, even if the patient is in the total parenteral nutrition stage, the system can automatically determine the timing for resuming enteral nutrition and generate a de-escalation plan in the reverse order of TPN → SPN → EN → ONS, avoiding long-term dependence on intravenous nutrition.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic decision-making system for the entire nutritional process of critically ill patients based on artificial intelligence, characterized in that, The system includes: The multi-source heterogeneous data acquisition and integration module is used to connect with hospital information systems, laboratory information systems, electronic medical record systems and bedside monitoring equipment to collect multi-dimensional clinical data of patients in real time. The stratified dynamic nutrition risk assessment module adopts a three-tier architecture of initial screening layer, diagnosis layer and dynamic reassessment layer, which is used to automatically perform nutrition risk screening and malnutrition diagnosis, and periodically refresh the risk stratification. The evidence-based nutrition initiation decision module is used to output standardized decisions, including initiation timing, intervention pathway, and initial dose, based on the quantitative rules of the built-in critical care nutrition guidelines. The dynamic monitoring and dose adjustment module is used to periodically output personalized dose adjustment schemes based on a deep reinforcement learning framework. The five-step nutrition intervention and management module is used to digitally implement the five-step nutrition diagnosis and treatment path, lock the patient's current intervention step in real time, and realize bidirectional intelligent switching between ascending and descending steps. The tiered early warning and report output module is used to generate multi-level risk real-time early warnings and daily nutritional diagnosis and treatment reports; The multi-model fusion artificial intelligence decision engine is connected to the hierarchical dynamic nutrition risk assessment module, the evidence-based nutrition initiation decision module, the dynamic monitoring and dosage adjustment module, and the five-step nutrition intervention and control module, respectively, and is used to perform calculations. The multi-model fusion artificial intelligence decision engine includes a multi-source data fusion encoding layer, a multi-prediction model integration layer, a guideline rule computable transformation layer, and an interpretable output layer.

2. The system according to claim 1, characterized in that, In the stratified dynamic nutritional risk assessment module, the initial screening layer automatically performs NRS2002 scoring and mNUTRIC scoring. When the NRS2002 score is greater than or equal to 3 or the mNUTRIC score is greater than or equal to 5, the in-depth assessment is automatically triggered. The diagnostic layer automatically verifies the malnutrition diagnosis criteria according to the GLIM standard. The dynamic reassessment layer refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours.

3. The system according to claim 1, characterized in that, The dynamic monitoring and dosage adjustment module outputs a personalized dosage adjustment plan every 4 hours based on the DeepEN framework, including automatic handling of gastrointestinal tolerance, automated early warning linkage for refeeding syndrome, dosage adjustment linked to inflammation and glucose metabolism, and automatic determination of energy target achievement. Among them, the automated early warning linkage for refeeding syndrome monitors blood phosphorus index in real time. When blood phosphorus is less than 0.65 mmol / L, it immediately issues a graded warning and simultaneously recommends reducing energy to 20 kcal / h and supplementing with multivitamins.

4. A dynamic decision-making method for the entire nutritional process of critically ill patients according to any one of claims 1-3, characterized in that, The method includes: The process involves collecting and integrating multi-source heterogeneous data, connecting with hospital information systems, laboratory information systems, electronic medical record systems, and bedside monitoring equipment to gather multi-dimensional clinical data from patients in real time. The stratified dynamic nutritional risk assessment process adopts a three-tier architecture of initial screening, diagnosis, and dynamic reassessment. It automatically performs nutritional risk screening and malnutrition diagnosis, and periodically updates the risk stratification. Evidence-based nutrition initiation decision-making steps, based on built-in critical care nutrition guidelines and quantitative rules, output standardized decisions including initiation timing, intervention pathway, and initial dose; The dynamic monitoring and dosage adjustment process periodically outputs personalized dosage adjustment plans based on a deep reinforcement learning framework. The five-step nutritional intervention and management steps are digitally implemented to create a five-step nutritional diagnosis and treatment path, which can lock the patient's current intervention step in real time and realize two-way intelligent switching between step advancement and step demotion. The system includes a tiered early warning and reporting output process, generating multi-level real-time risk warnings and daily nutritional diagnosis and treatment reports. The calculations of the hierarchical dynamic nutritional risk assessment step, the evidence-based nutritional initiation decision step, the dynamic monitoring and dosage adjustment step, and the five-step nutritional intervention and control step are all executed by a multi-model fusion artificial intelligence decision engine.

5. The method according to claim 4, characterized in that, The stratified dynamic nutritional risk assessment steps include: the initial screening layer automatically performs NRS2002 scoring and mNUTRIC scoring, and automatically triggers in-depth assessment when the NRS2002 score is greater than or equal to 3 or the mNUTRIC score is greater than or equal to 5; the diagnostic layer automatically verifies the diagnostic criteria for malnutrition according to the GLIM standard, which includes a positive screening result, at least one phenotypic indicator, and at least one causal indicator of inflammation or insufficient intake; the dynamic reassessment layer refreshes all real-time indicators and updates the risk stratification every 4 to 6 hours.

6. The method according to claim 4, characterized in that, In the evidence-based nutrition initiation decision-making step: The timing of initiation is determined based on hemodynamics, lactate levels, and gastrointestinal function. Enteral nutrition is initiated early within 48 hours of ICU admission for patients with stable hemodynamics. Parenteral nutrition is delayed or preferred for patients in shock who are using high-dose vasoactive drugs. The intervention pathway strictly follows the five-step nutrition ladder to determine the intervention plan step by step. The initial dose follows the acute phase protective feeding standard, with 6 to 8 kcal / kg / day of calories and 0.2 to 0.9 g / kg / day of protein in the first week, using a low initial gradient incremental strategy.

7. The method according to claim 4, characterized in that, The dynamic monitoring and dosage adjustment steps are based on the DeepEN framework and output personalized dosage adjustment plans every 4 hours, including: automatic handling of gastrointestinal tolerance, monitoring signs of feeding intolerance and automatically recommending slowing down, continuous infusion, changing formula or pausing feeding; automated early warning linkage for refeeding syndrome, real-time monitoring of serum phosphorus levels, and immediate graded warning when serum phosphorus is less than 0.65 mmol / L, simultaneously suggesting energy reduction to 20 kcal / h and supplementation with multivitamins; dosage adjustment linked to inflammation and glucose metabolism, dynamically adjusting calorie supply based on insulin infusion rate, CRP and IL-6; automatic determination of energy target achievement, calculating the daily intake as a percentage of the target energy, and automatically pushing tiered escalation suggestions when it is below 60% for 2 to 3 consecutive days.

8. The method according to claim 4, characterized in that, The five-step nutritional intervention and management steps include: real-time locking of the patient's current intervention step; step-up determination, when the intake of energy is less than 60% of the target for several consecutive days, automatically generating an upgrade plan from oral nutrition supplementation to tube feeding enteral nutrition to supplemental parenteral nutrition to total parenteral nutrition; step-down determination, continuously monitoring intestinal function recovery indicators, and automatically determining the timing of resuming enteral nutrition even in the total parenteral nutrition stage.

9. The method according to claim 4, characterized in that, In the multi-model fusion artificial intelligence decision engine: the multi-source data fusion encoding layer uses deep learning to unify heterogeneous clinical data features; the multi-prediction model integration layer integrates the XGBoost model to predict early malnutrition, the NutriSighT Transformer model to predict insufficient enteral feeding, and the Dueling Double DQN reinforcement learning model to output individualized energy and protein targets; the guideline rule calculation conversion layer embeds the quantitative threshold of critical care nutrition guidelines; and the interpretable output layer synchronously outputs the basis indicators, guideline clauses, and feature weights for each decision.

10. The method according to claim 4, characterized in that, The tiered early warning and reporting output steps include: three-level real-time risk early warning, divided into three levels: alert, warning, and critical; automatic generation of daily nutrition diagnosis and treatment reports, including intake, target achievement rate, step status, and adjustment suggestions; and a large language model clinical decision support submodule, which transforms indicators and guidelines into standardized operational suggestions.

Images