TWI889340B - Method and system for predicting nutritional risk of critically ill patients using machine learning - Google Patents

Abstract

The present invention discloses a method and system for predicting the nutritional risk of critically ill patients by machine learning, which generates a nutritional risk prediction model by calculating and analyzing the key feature data of multiple patients with a machine learning model, and then inputs the key feature data of the patient to be tested into the nutritional risk prediction model for calculation and analysis to obtain the nutritional risk of the patient to be tested. Accordingly, the method and system for predicting the nutritional risk of critically ill patients by machine learning disclosed in the present invention can quickly and accurately evaluate the nutritional risk of patients on a regular or real-time basis, so as to achieve the effect of improving the quality of care, enhancing the prognosis of patients and reducing the burden of care.

Description

Method and system for predicting nutritional risk of critically ill patients using machine learning

The present invention relates to a method for applying machine learning technology to a clinical decision-making system, and more particularly to a method and system for predicting nutritional risks of critically ill patients using machine learning.

According to the American Society for Parenteral and Enteral Nutrition guideline, NRS2002 and NUTRIC scores should be used to define patients with high nutritional risk. However, the European Society for Parenteral and Enteral Nutrition guideline believes that there is no gold standard for nutritional risk and that anyone who stays in the intensive care unit (ICU) for more than 48 hours should be considered a high-risk patient.

Although studies have shown that for patients with high nutritional risk, active nutritional support strategies can help improve patient prognosis, critically ill patients are highly heterogeneous and there is currently no nutritional support strategy that can be universally used for all critically ill patients in clinical practice. Nutritional risk assessment can only be performed by dietitians; however, due to the limited number of dietitians and working hours, it is impossible to provide patients with continuous nutritional assessment and care.

The main purpose of the present invention is to provide a method and system for predicting the nutritional risk of critically ill patients by machine learning. The method can construct a nutritional risk prediction module for critically ill patients by combining artificial intelligence technology with clinical data, thereby achieving real-time and accurate assessment of the nutritional risk of critically ill patients, allowing nutritionists to quickly perform appropriate nutritional interventions, and reducing the manpower burden of medical staff and nutritionists.

Therefore, in order to achieve the above-mentioned purpose, the present invention discloses a method for predicting the nutritional risk of critically ill patients by machine learning, which mainly trains a machine learning model by key feature data of multiple patients to construct a nutritional risk prediction module, and uses the nutritional risk prediction module to evaluate the nutritional risk of a patient to be tested.

Among them, the patient was admitted to the intensive care unit.

Specifically, the method of predicting nutritional risk of critically ill patients using machine learning mainly includes the following steps:

Step a: Obtain key characteristic data of multiple patients and their nutritional risks as a data set, wherein the key characteristics include APACHE II score, albumin, age, body mass index and hemoglobin.

Step b: training a predetermined machine learning model with a portion of the data set in step a to generate the nutritional risk prediction model; and using another portion of the data set in step a to verify the prediction accuracy of the nutritional risk prediction model.

Step c: inputting a key characteristic data of a patient to be tested into the nutritional risk prediction model for analysis to obtain the nutritional risk of the patient to be tested.

Among them, the analysis, acquisition, training and other actions in the method of predicting nutritional risks of critically ill patients by machine learning disclosed in the present invention are executed by a processor.

Among them, the machine learning model is an algorithmic analysis method, such as XGBoost, CatBoost, LightGBM and logical regression.

Among them, the key data features are obtained from the patient 24 hours before entering the intensive care unit to 48 hours after entering the intensive care unit.

The key features are obtained by calculating and analyzing the characteristic data of the patients during hospitalization through the machine learning model. The characteristic data includes body position detection data, clinical detection data, drug use data, physiological parameter data, and critical illness index data; specifically, the characteristic data includes corresponding data of characteristic factors such as age, blood sugar, hemoglobin, body mass index, APACHE II score, SOFA score, gender, and medication status.

Another embodiment of the present invention discloses a system for predicting nutritional risks of critically ill patients using machine learning, which mainly includes a database and a processor, wherein the database is used to store characteristic data of multiple patients; the processor is connected to the database and can access the data in the database and perform calculation and analysis processing on the data.

The characteristic data is obtained during the patient's hospitalization period and includes a body position detection data, a clinical detection data, a drug use data, a physiological parameter data, and a critical illness indicator data.

Specifically, the processor includes a data receiving module and a nutritional risk prediction module; the data receiving module receives key characteristic data of a patient to be tested, and the key characteristics include APACHE II score, albumin, age, body mass index and hemoglobin as described above.

The nutritional risk prediction module analyzes the key characteristic data of the patient to be tested using the nutritional risk prediction model to output the nutritional risk of the patient to be tested.

The data receiving module is capable of screening out the key characteristic data of the patient to be tested from the database.

In another embodiment of the present invention, the processor further includes a data processing module, which performs calculation and analysis on key feature data of multiple patients using a machine learning model to obtain the nutritional risk prediction model.

The machine learning model is XGBoost, CatBoost, LightGBM, or logical regression.

The present invention discloses a method and system for predicting the nutritional risk of critically ill patients by machine learning. The method uses a machine learning model to perform calculation and analysis on the key feature data of a plurality of patients and their nutritional risks to generate a nutritional risk prediction model. The key feature data of the patient to be tested is input into the nutritional risk prediction model for calculation and analysis to obtain the nutritional risk of the patient to be tested. Accordingly, the method and system disclosed in the present invention for predicting nutritional risk of critically ill patients by machine learning can quickly and accurately assess the nutritional risk of patients on a regular or real-time basis, allowing nutritionists to intervene appropriately and timely, thereby achieving the effects of improving the quality of care, enhancing patient prognosis, and reducing the burden of care.

The following will explain the terms used in the present invention. If they are not listed in the following description, they will be interpreted according to reference materials recognized by those with ordinary knowledge in the technical field to which the present invention belongs, such as dictionaries, literature or common knowledge.

The term "postural measurement data" refers to information related to height, weight, or body composition.

The term "clinical test data" refers to the information obtained by doctors from examining patients for the purpose of diagnosis, including biochemical values, text data, image data, etc.

The term "drug use data" refers to information about medications administered or administered to patients, including drug name, dosage, side effects, etc.

The term "physiological parameter data" refers to information obtained by measuring patients' blood pressure, heart rate, respiratory rate, etc. by machines at irregular or regular intervals for diagnosis or care purposes.

The term "critical illness index data" refers to the judgment result of the severity of the patient's illness, which usually requires a comprehensive statistical analysis of multiple parameters to obtain.

The term "key characteristic data" refers to a value, data, test result, analysis result or calculation result corresponding to a key characteristic. In the present invention, the key characteristics are APACHE II score, albumin, age, body mass index and hemoglobin.

The term "database" refers to a structured collection of information or data that is stored in a predetermined manner on a device or a device, such as a computer system, hard drive, or cloud space.

The term "processor" refers to hardware that can be used to execute a series of programs, instructions, or operate on data, such as computers, computer processing equipment, etc.

The term "machine learning" refers to using a machine learning model to learn and improve in data, find patterns and relationships, and make decisions and predictions based on its learning and analysis results. For example, the machine learning model listed in the present invention includes XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), LightGBM (Light Gradient Boosting Machine), and Logistic Regression (LR).

The term "Intensive Care Unit" refers to a ward provided to patients who are deemed to require additional care or are in serious condition. It belongs to different departments depending on the department.

The term "APACHE II" is the abbreviation of Acute Physiology and Chronic Health Evaluation II, which is a disease severity grading system. Specifically, APACHE II is based on the patient's age and 12 physiological test variables within 24 hours after admission to the intensive care unit, such as body temperature, heart rate, mean arterial pressure, respiratory rate, arterial oxygen partial pressure, inspired oxygen partial pressure, blood sodium, blood potassium, blood creatinine, white blood cell count, platelet count, Glasgow Coma Index, etc., and analyzes them to obtain an integer score between 0-71.

The term "SOFA" is an abbreviation for Sequential Organ Failure Assessment, which is a system used to assess the organ function or failure rate of patients in the intensive care unit. SOFA is based on the scores of six systems, including respiratory, cardiovascular, liver, coagulation, kidney, and nervous system, and the total score is added up.

The term "SHAP force plot" or "SHAP value" refers to the value assigned to each feature in a data set or a set of data when a given machine learning model produces a prediction for the data set or a set of data.

One embodiment of the present invention discloses a method for predicting nutritional risk of critically ill patients using machine learning, which comprises the following steps:

Step a: Obtain key characteristic data and nutritional risk of multiple patients as a data set, wherein the key characteristics include five, including APACHE II score, albumin, age, body mass index and hemoglobin; and nutritional risk is the result of at least one nutritionist's assessment based on the patient's condition, including high nutritional risk and low nutritional risk.

Step b: using a portion of the data set in step a as training data and inputting it into a machine learning model for computational training to generate a nutritional risk prediction model; and using another portion of the data set as validation data and performing computational analysis of the machine learning model to verify the prediction accuracy of the nutritional risk prediction model.

Step c: obtaining a key characteristic data of a patient to be tested, and analyzing the key characteristic data using the nutritional risk prediction model to obtain the nutritional risk of the patient to be tested.

Among them, the key feature data of each patient comes from the data within a predetermined period before and after each patient is admitted to the ICU. For example, the predetermined period is from 24 hours before the patient enters the ICU to 48 hours after entering the ICU.

The nutritional risk of the patient to be tested obtained by the method of predicting the nutritional risk of critically ill patients by machine learning disclosed in the present invention can be a word, a number, a graphic, or a combination of any two of the above. For example, the nutritional risk is a score of 0-9, a word of high or low, or a graphic of a smiling face or a crying face.

In one embodiment of the present invention, the machine learning model is an algorithm commonly used by those skilled in the art, such as XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), LightGBM (Light Gradient Boosting Machine), and Logistic Regression (LR).

In an embodiment of the present invention, the patient is admitted to an intensive care unit.

In another embodiment of the present invention, the method for predicting nutritional risk of critically ill patients by machine learning further comprises a pre-processing step for selecting the key feature, and the pre-processing step is:

(1) Obtain the characteristic data and nutritional risk of multiple patients during hospitalization. The characteristic data can be divided into five categories: postural measurement data, clinical test data, drug use data, physiological parameter data, and critical illness index data. Specifically, the characteristic data includes age, blood sugar, hemoglobin, body mass index, APACHE II score, SOFA score, gender, medication status, etc. As mentioned above, the nutritional risk of each patient is obtained by his nutritionist based on a judgment standard.

(2) Using a portion of the feature data and the nutritional risk data to train a machine learning model to generate a nutritional risk prediction model prototype; wherein the machine learning model is an algorithmic analysis method, such as XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), LightGBM (Light Gradient Boosting Machine), Logistic Regression (LR) or other algorithms well known in the art to which the present invention belongs.

(3) Analyze the importance of each characteristic factor in the nutritional risk prediction model prototype in terms of its impact on nutritional risk, so as to obtain the key characteristics disclosed in the present invention.

In the above-mentioned pre-processing step (2), part of the characteristic data and the nutritional risk data may be used to verify the nutritional risk prediction model prototype.

In another embodiment of the present invention, a system for predicting the nutritional risk of critically ill patients by machine learning is disclosed, which is used to execute or operate the above method to obtain a nutritional risk assessment result of a patient. The system for predicting the nutritional risk of critically ill patients by machine learning mainly includes a database and a processor, wherein:

The database is used to store characteristic data and nutritional risks of multiple patients during hospitalization, wherein the characteristic data includes body position detection data, clinical detection data, drug use data, physiological parameter data, and critical illness indicator data.

The processor is connected to the database in a wired or wireless manner to access data from the database and analyze and process the data to generate a nutritional risk prediction result of the patient. Specifically, the processor has a data receiving module, a data processing module and a nutritional risk prediction module, wherein:

The data receiving module obtains and/or screens a key characteristic data of a patient to be tested from an external unit or the database, wherein the key characteristic data is APACHE II score, albumin content, age, body mass index and hemoglobin; and the external unit can be a computer, a tablet, a medical care system, etc.

The data processing module obtains key feature data and nutritional risks of multiple patients from the database, and performs an algorithmic analysis program using a machine learning model to generate a nutritional risk prediction model, wherein the machine learning model includes XGBoost, CatBoost, LightGBM, logical regression, etc.; and the algorithmic analysis program includes a training program and a verification program, and the training program and the verification program respectively use at least part of the key feature data and nutritional risks as processing objects.

The nutritional risk prediction module analyzes the key characteristic data of the patient to be tested using the nutritional risk prediction model to output the nutritional risk of the patient to be tested.

In the next embodiment of the present invention, the key feature data is obtained from 24 hours before a patient enters the intensive care unit to 48 hours after entering the intensive care unit.

In order to illustrate the technical features and efficacy of the present invention, several test results are described in detail below. All tests have been approved by the review committee of Taichung Veterans General Hospital.

The SHAP diagram in the following example is used to illustrate the relationship between nutritional risks and traits.

Example 1: Subject data

A total of 1,994 patients were included in the study. All of them met the screening criteria of being over 20 years old, having respiratory failure requiring ventilator support, and receiving complete nutritional risk assessment. Patients who had been admitted to the intensive care unit (ICU) for less than 3 days or were pregnant were excluded. All data had been de-identified.

The data collection period for all subjects was between 24 hours before and 48 hours after entering the ICU. The collected data included age, gender, anthropometric data (height and weight), biochemical data (albumin, liver function tests, kidney function tests, etc.), and other parameters (Acute Physiology and Chronic Health Evaluation (APACHE) II score, SOFA score, comorbidities), etc.

The nutritional risk of each subject was determined by consensus decision by six experienced nutritionists. The results of the nutritional risk determination were as follows: 701 subjects (35.16%) had a high nutritional risk, and 1,293 subjects had a low nutritional risk. The reliability (consistency) evaluated by Fleiss' kappa was 0.64.

The subjects were divided into groups according to nutritional risk, and the data between the two groups were analyzed. The results are shown in Table 1. The analysis and comparison of the two groups of data were performed using t-test or chi-square test; the values are expressed as mean ± standard deviation.

According to Table 1, compared with the subjects in the low nutritional risk group, the subjects in the high nutritional risk group had the following characteristics: older age (72.83 ± 14.58 vs. 61.68 ± 15.87, p ＜ 0.01), lower weight (58.58 ± 12.58 vs. 64.44 ± 14.20, p ＜ 0.01), lower body mass index (22.72 ± 4.52 vs. 24.28 ± 4.79, p ＜ 0.01) and lower albumin (2.58 ± 0.57 vs. 3.14 ± 0.64, p ＜ 0.01).

Furthermore, compared with the subjects with low nutritional risk, the subjects in the high nutritional risk group had higher disease severity, such as APACHE II and SOFA scores; and the clinical outcomes, such as ICU days, ventilator use days, hospital stay days, and mortality, were worse in the high nutritional risk group than in the low nutritional risk group.

Table 1: Patients’ demographic characteristics, severity scores, and clinical outcomes Characteristic Factors All (n = 1994) High nutritional risk group (n = 701) Low nutritional risk group (n = 1293) P-value Demographic characteristics Age (years) 65.60±16.32 72.83±14.58 61.68±15.87 <0.001** Gender (female) 706 (35.41%) 262 (37.38%) 444 (34.34%) 0.192 Weight (kg) 62.38±14.02 58.58±12.85 64.44±14.20 <0.001** Body mass index 23.73±4.75 22.72±4.52 24.28±4.79 <0.001** Albumin (mg/dl) 2.92±0.67 2.58±0.57 3.14±0.64 <0.001** Complications (n,%) diabetes 658 (33.0%) 256 (36.52%) 402 (31.09%) 0.016* Cirrhosis 157 (7.87%) 70 (9.99%) 87 (6.73%) 0.013* Uremia 633 (31.75%) 287 (40.94%) 346 (26.76%) <0.001** Central nervous system diseases 407 (20.41%) 142 (20.26%) 265 (20.49%) 0.946 Chronic lung disease 300 (15.05%) 126 (17.97%) 174 (13.46%) 0.009** Immunocompromising diseases 175 (8.78%) 76 (10.84%) 99 (7.66%) 0.021* Any malignant tumor, including lymphoma and leukemia, except malignant tumors of the skin 661 (33.15%) 273 (38.94%) 388 (30.01%) <0.001** Congestive heart failure 358 (17.95%) 143 (20.4%) 215 (16.63%) 0.042* Chronic lung disease 300 (15.05%) 126 (17.97%) 174 (13.46%) 0.009** Disease severity score APACHE II Scoring 23.78±7.80 29.37±5.62 20.42±6.96 <0.001** SOFA Rating 7.55±3.91 9.54±3.69 6.46±3.59 <0.001** Clinical prognosis Number of days in ICU (days) 10.34±10.14 13.36±10.62 8.70±9.48 <0.001** Respirator use days (days) 5.27±10.97 7.49±12.43 4.06±9.89 <0.001** Hospital stay (days) 5.27±10.97 7.49±12.43 4.06±9.89 <0.001** Hospital mortality rate 27.01±27.34 31.42±25.69 24.62±27.91 <0.001** Intensive care unit (n,%) <0.001** Medical 1432 (71.82%) 562 (80.17%) 870 (67.29%) Operation 562 (28.18%) 139 (19.83%) 423 (32.71%)

Example 2: Constructing a prediction model (I)

As shown in Figure 1, 80% of the subjects and their feature factors in Example 1 were randomly selected as the training data set, and a 5-fold cross-validation was used; the remaining 20% of the subject data was used as the test data set; among them, the classifiers used to train the prediction model include XGBoost, CatBOOST, LightGBM, and Logistic Regression.

The training datasets were trained with the above four classifiers and 5-fold cross validation, and the results are shown in Table 2. The test datasets were also put into the above four classifiers for calculation, and the test results are shown in Figures 2 and 3.

Table 2: Results of training and testing the prediction model using all features with different classifiers Classifier Accuracy Sensitivity Specificity Accuracy AUROC 5-fold Cross-Validation XGBoost 0.832 ± 0.039 0.915 ± 0.020 0.780 ± 0.044 0.868 ± 0.027 0.928 ± 0.023 Catboost 0.832 ± 0.039 0.916 ± 0.018 0.771 ± 0.059 0.865 ± 0.031 0.932 ± 0.024 LightGBM 0.803 ± 0.039 0.897 ± 0.022 0.780 ± 0.046 0.856 ± 0.027 0.925 ± 0.025 Logistic Regression 0.737 ± 0.069 0.872 ± 0.040 0.655 ± 0.069 0.796 ± 0.040 0.863 ± 0.035 Test XGBoost 0.779 0.876 0.801 0.850 0.921 Catboost 0.803 0.888 0.837 0.869 0.926 LightGBM 0.784 0.876 0.823 0.857 0.923 Logistic Regression 0.713 0.849 0.688 0.792 0.852

From the results in Table 2, we can see that the accuracy, sensitivity, specificity and accuracy of algorithms: XGBoost and Catboost are similar and better than other algorithms, especially the accuracy, sensitivity and AUROC performance in 5-fold cross validation. Algorithm: LightGBM has the second best prediction performance, while algorithm Logistic regression has the worst prediction performance.

From the results in Figures 2 and 3, we can see that the prediction net benefits of models such as Catboost, XGBoost, and LightGBM are higher than Logistic Regression.

Example 3: Feature Importance Analysis

From the results of Example 2, we can see that Catboost, XGBoost, and LightGBM have better prediction results. Therefore, in this example, Catboost will be used as an example, and the SHAP diagram will be used to analyze the importance of the characteristic factors from the subjects on the impact of nutritional risks. The results are shown in Figures 4 and 5.

From the results of Figure 4, we can see the top 20 characteristic factors that can affect nutritional risk, and the top 5 characteristic factors are APACHE II score, albumin, age, body mass index (BMI) and hemoglobin. In addition, from the results of Figure 5, we can see that APACHE II score and age are positively correlated with nutritional risk; while albumin, body mass index and hemoglobin are negatively correlated with nutritional risk.

Example 4: Constructing a Prediction Model (II)

Referring to the training and testing process of Example 2, and based on the results of feature importance in Example 3, the top five important factors were selected as the model prediction model construction factors in this example. That is, in this example, there are five feature factors in the training data set and the test data set, namely APACHE II score, albumin, age, body mass index (BMI) and hemoglobin. The training and testing results are shown in Table 3 below.

Table 3: Results of training and testing the prediction model using five feature factors with different classifiers Classifier Accuracy Sensitivity Specificity Accuracy AUROC 5-fold Cross-Validation XGBoost 0.825 ± 0.037 0.910 ± 0.020 0.784 ± 0.057 0.866 ± 0.028 0.929 ± 0.020 Catboost 0.833 ± 0.043 0.916 ± 0.020 0.779 ± 0.061 0.868 ± 0.033 0.934 ± 0.018 LightGBM 0.796 ± 0.033 0.892 ± 0.020 0.779 ± 0.049 0.852 ± 0.022 0.923 ± 0.018 Logistic Regression 0.786 ± 0.053 0.892 ± 0.029 0.730 ± 0.047 0.835 ± 0.032 0.902 ± 0.030 Test XGBoost 0.774 0.872 0.801 0.847 0.919 Catboost 0.803 0.891 0.809 0.862 0.923 LightGBM 0.777 0.872 0.816 0.852 0.913 Logistic Regression 0.720 0.837 0.766 0.812 0.890

Comparing the results of Table 2 and Table 3, for the same classifier, the performance of nutritional risk prediction using 5 characteristic factors is as good as the performance of nutritional risk prediction using all characteristic factors; that is, the results of this example can confirm that the following 5 characteristic factors: APACHE II score, albumin, age, body mass index (BMI) and hemoglobin are the key characteristic factors in the model for predicting patient nutritional risk disclosed in the present invention.

Furthermore, from the test results shown in Table 3, it can be seen that the nutritional risk prediction model obtained by using 5 characteristic factors has an accuracy rate of 86.2% in predicting the nutritional risk of patients.

Example 5: SHAP (Shapley additive explanations) analysis

The SHAP force plot was used to analyze the effects of the five characteristics disclosed in the present invention: APACHE II score, albumin content, age, body mass index and hemoglobin on the prediction model. The results are shown in Figures 6 to 10, where each point in each figure represents a subject. When the SHAP force plot exceeds 0, it means that the nutritional risk increases.

Combining the results of Figures 6 to 10, it can be seen that when the age is about 70 years old or below and the APACHE II score is about 25 or below, the nutritional risk of the subject is predicted to be low; and when the body mass index is about less than 20, the albumin is about less than 3 mg/dl, and the hemoglobin is less than about 11 g/dl, the nutritional risk of the subject is predicted to be high.

without

Figure 1 is a flowchart of the model construction in Example 2. Figure 2 is a ROC curve obtained after the test data is calculated using different algorithms. Figure 3 is a decision curve obtained after the test data is analyzed using different algorithms. Figure 4 is the result of ranking the influence of characteristic factors on nutritional risk. Figure 5 shows the predicted influence of characteristic factors, where red represents a higher total SHAP value and a greater influence on the prediction; blue represents a lower total SHAP value and a smaller influence on the prediction. Figure 6 is the result of using SHAP to analyze the influence of age on the prediction of nutritional risk. Figure 7 shows the results of using SHAP to analyze the impact of APACHE II score on predicting nutritional risk. Figure 8 shows the results of using SHAP to analyze the impact of body mass index on predicting nutritional risk. Figure 9 shows the results of using SHAP to analyze the impact of albumin on predicting nutritional risk. Figure 10 shows the results of using SHAP to analyze the impact of hemoglobin on predicting nutritional risk.

without

Claims (6)

A method for predicting nutritional risk of critically ill patients by machine learning, comprising analyzing key feature data of a patient to be tested by a nutritional risk prediction model to obtain the nutritional risk of the patient to be tested; wherein: the key features are APACHE II score, albumin, age, body mass index and hemoglobin; the nutritional risk prediction model is obtained by the following steps: step a: obtaining key feature data of a plurality of patients and their nutritional risk as a data set; step b: using at least a part of the data set as a training data to train a machine learning model to generate a nutritional risk prediction model, wherein the machine learning model is selected from XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), LightGBM (Light Gradient Boosting Machine) and Logistic Regression (LR); the above analysis, acquisition and training operations are performed by a processor. A method for predicting nutritional risk of critically ill patients by machine learning as described in claim 1, wherein step b further includes using another part of the data set as a test data to analyze the prediction accuracy of the nutritional risk prediction model. A method for predicting nutritional risk of critically ill patients using machine learning as described in claim 1, wherein the patients are patients admitted to an intensive care unit. A method for predicting nutritional risk of critically ill patients using machine learning as described in claim 1, wherein the key feature data is collected during a time period from 24 hours before the patient enters the intensive care unit to 48 hours after the patient enters the intensive care unit. A system for predicting the nutritional risk of critically ill patients by machine learning comprises: a database that collects characteristic data of a plurality of patients during hospitalization and their nutritional risk, wherein the characteristic data comprises a body position detection data, a clinical detection data, a drug use data, a physiological parameter data, and a critical disease index data; a processor that is connected to the database in a wired or wireless manner and can transmit data between each other, and has a data receiving module, a data processing module, and a nutritional risk prediction module; wherein: the data receiving module receives a key characteristic data of a patient to be tested, and the key characteristic data is APACHE II score, albumin, age, body mass index and hemoglobin; the data processing module uses a machine learning model to calculate and analyze the key feature data of multiple patients and their nutritional risks to generate the nutritional risk prediction model, wherein the machine learning model is selected from a group consisting of XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), LightGBM (Light Gradient Boosting Machine) and Logistic Regression (LR); the nutritional risk prediction module uses the nutritional risk prediction model to analyze the key feature data of the patient to be tested to generate the nutritional risk of the patient to be tested. A system for predicting nutritional risk of critically ill patients using machine learning as described in claim 5, wherein the data receiving module selects the key feature data of the patient to be tested from the database.