A method and system for generating malnutrition risk in patients with traumatic brain injury
By constructing a specific malnutrition risk prediction model and utilizing multi-dimensional clinical data and machine learning algorithms, the accuracy and interpretability issues of nutritional risk assessment for patients with traumatic brain injury (TBI) were resolved, achieving efficient and reliable prediction of malnutrition risk in TBI patients.
Patent Information
- Application Number
- CN202611447628.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing nutritional risk screening tools rely on subjective information for assessing patients with traumatic brain injury (TBI), which is difficult to obtain accurately. Furthermore, they lack integration of TBI-specific pathophysiological characteristics, resulting in insufficient accuracy and interpretability in the quantitative analysis of malnutrition risk.
A model for predicting the risk of malnutrition in patients with traumatic brain injury was constructed. By receiving multi-dimensional clinical data, machine learning algorithms were used for data preprocessing and feature selection, including missing value imputation, standardization and sample balancing. Features reflecting neurological function impairment, inflammatory and immune status and the intensity of treatment intervention were integrated to form a set of specific core features.
This approach enables the objective quantification of malnutrition risk in TBI patients, improves the sensitivity and specificity of predictions, addresses the issues of missing data and class imbalances, and enhances the stability and generalization ability of the model in complex clinical data environments.
Smart Images

Figure CN123120318A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of generating malnutrition risk information for patients with traumatic brain injury, and specifically to a method and system for generating malnutrition risk information for patients with traumatic brain injury. Background Technology
[0002] Traumatic brain injury (TBI) is organic damage to brain tissue caused by external force. Patients often present with stress-induced hypermetabolism, altered consciousness, and dysphagia. Malnutrition is a common complication of TBI, associated with infection, stress injury, delayed recovery, prolonged hospital stay, and increased mortality. Early identification of high-risk patients for malnutrition and timely implementation of targeted nutritional support and nursing interventions are crucial for reducing the incidence of complications, promoting neurological function recovery, and improving clinical outcomes.
[0003] Currently, most clinically used nutritional risk screening tools rely on subjective information such as patients' self-recollection of weight changes, appetite, and food intake, such as NRS-2002, MUST, and MNA-SF. Some tools focus primarily on disease severity and routine clinical indicators, such as the NUTRIC score. However, TBI patients often have impaired consciousness and speech, making it difficult to accurately obtain some subjective nutritional assessment information, and they are also susceptible to memory bias or communication impairments.
[0004] Although machine learning algorithms have begun to be used for medical risk prediction in recent years, existing models are mostly designed for the general population or chronic diseases such as cardiovascular disease, and lack specific optimization for the characteristics of TBI patient data. Specifically, existing general screening tools and prediction models have failed to fully integrate the pathophysiological characteristics of TBI, such as swallowing disorders and inflammatory immune dysregulation, resulting in insufficient accuracy, sensitivity, and interpretability in the quantitative analysis of malnutrition risk in TBI patients.
[0005] Therefore, there is an urgent need to construct a specific malnutrition risk prediction model for TBI patients, which comprehensively incorporates multi-dimensional information such as demographic factors, disease factors, treatment factors, and laboratory indicators, and uses data preprocessing technology to solve the problems of data missingness and imbalance, so as to achieve accurate quantitative prediction of malnutrition risk for TBI patients, and further transform it into a web calculator to improve its application convenience in early clinical screening, risk stratification management, and individualized nutritional intervention. Summary of the Invention
[0006] This application provides a method and system for generating malnutrition risk information for patients with traumatic brain injury, which can solve the technical problems in the prior art, such as low reliability due to reliance on subjective information, insufficient targeting due to lack of specific features, and poor data quality affecting model accuracy.
[0007] In a first aspect, embodiments of this application provide a method for predicting the risk of malnutrition in patients with traumatic brain injury, the method comprising: Receive clinical multidimensional data from patients with traumatic brain injury, including demographic indicators, disease-related indicators, treatment-related indicators, and laboratory indicators; The clinical multidimensional data are preprocessed, including data standardization. The preprocessed clinical multidimensional data were input into a pre-trained malnutrition risk prediction model to obtain the malnutrition risk probability value related to the patient's nutritional status. The malnutrition risk prediction model is based on a machine learning algorithm and is trained using historical training data that has undergone missing value imputation, data standardization, and sample class balancing. The historical training data contains a set of core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention. The set of core features includes features that reflect basic nutritional reserves, features that reflect the degree of neurological function impairment, features that reflect the inflammatory and immune status, and features that reflect the intensity of treatment intervention.
[0008] Secondly, embodiments of this application provide a method for constructing a malnutrition risk prediction model for patients with traumatic brain injury, including: Historical clinical sample data of patients with traumatic brain injury were obtained, and the historical clinical sample data had characteristics such as missing values, inconsistent data scaling, and uneven category distribution. To address the missing value feature, a random forest-based missing value imputation method was used to process the data; to address the data scaling inconsistency, Z-score was used for standardization; and to address the uneven class distribution feature, SMOTE technology was used to balance the historical clinical sample data and generate a balanced training set. The historical clinical sample data is linearly compressed using LASSO regression to obtain a first feature set; the historical clinical sample data is nonlinearly evaluated using the Boruta algorithm to obtain a second feature set; the intersection of the first feature set and the second feature set is determined as the core feature set; the machine learning model is trained using the balanced training set to obtain the malnutrition risk prediction model.
[0009] Thirdly, embodiments of this application provide a system for predicting the risk of malnutrition in patients with traumatic brain injury, the system comprising: The data receiving module is used to receive clinical multidimensional data of patients with traumatic brain injury, including demographic indicators, disease-related indicators, treatment-related indicators, and laboratory indicators. The data preprocessing module is used to preprocess the clinical multidimensional data, and the preprocessing includes data standardization. The model calculation module is used to input the preprocessed clinical multidimensional data into the pre-trained malnutrition risk prediction model to obtain the malnutrition risk probability value related to the patient's nutritional status. An information output module is used to output the reference information; The malnutrition risk prediction model is based on a machine learning algorithm and is trained using historical training data that has undergone missing value imputation, standardization, and sample class balancing. The historical training data contains a set of core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention. The set of core features includes features that reflect basic nutritional reserves, features that reflect the degree of neurological function impairment, features that reflect the inflammatory and immune status, and features that reflect the intensity of treatment intervention.
[0010] The beneficial effects of the technical solutions provided in this application include: This application addresses the technical problem that traditional screening tools cannot accurately assess TBI patients due to impaired consciousness or aphasia by receiving multidimensional clinical data covering demographic, disease, treatment, and laboratory dimensions, thus eliminating reliance on patients' subjective recollections and achieving objective risk quantification without patient cooperation.
[0011] Furthermore, by training the model with a set of specific core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention, the limitations of general prediction models in not integrating TBI-specific pathophysiological features (such as dysphagia and the effects of mechanical ventilation) are overcome, significantly improving the sensitivity and specificity of predicting the risk of malnutrition in TBI patients.
[0012] Meanwhile, by using missing value imputation in the data preprocessing stage of model building to ensure data integrity, and by using SMOTE-based sample class balancing in the model building stage, the technical problems of information loss and uneven distribution of positive and negative samples in clinical data, which lead to difficulties in model training convergence or prediction bias, are effectively solved. This improves the stability and generalization ability of the model in complex clinical data environments, thereby ensuring that the reference information output in the application stage is accurate and reliable. Attached Figure Description
[0013] Figure 1 A flowchart illustrating a method for generating malnutrition risk information in patients with traumatic brain injury, provided in an embodiment of this application; Figure 2 A schematic diagram of a malnutrition risk prediction system for patients with traumatic brain injury provided in this application embodiment; Figure 3LASSO regression coefficient path diagram provided in the embodiments of this application; Figure 4 The Boruta feature importance evaluation graph provided for embodiments of this application; Figure 5 Subject operating characteristic curves for seven models provided for internal validation in this application embodiment; Figure 6 The internal verification calibration curves for seven models provided in this application embodiment; Figure 7 Clinical decision curves for seven models provided for internal validation in this application embodiment; Figure 8 SHAP bee colony graph (ranked by contribution of core indicators) provided for the embodiments of this application. Figure 9 The SHAP waterfall plot (individual patient indicator contribution plot) provided for the embodiments of this application; Figure 10 External validation subject operating characteristic curves provided for embodiments of this application; Figure 11 External verification calibration curves provided for embodiments of this application; Figure 12 External validation clinical decision curves provided for embodiments of this application; Figure 13 This is a schematic diagram of a web calculator interface provided in an embodiment of this application. Detailed Implementation
[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.
[0016] In a first aspect, embodiments of this application provide a method for generating malnutrition risk information for patients with traumatic brain injury, which is executed by a processor.
[0017] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic flowchart illustrating an embodiment of the method for generating malnutrition risk information in patients with traumatic brain injury according to this application. Figure 1 As shown, the methods for generating malnutrition risk information in patients with traumatic brain injury include: Step S1: Receive multidimensional clinical data from patients with traumatic brain injury. This multidimensional clinical data includes demographic indicators, disease-related indicators, treatment-related indicators, and laboratory indicators. This data can be automatically collected through the hospital information system or entered by medical staff, avoiding the reliance on the patient's subjective recollection in traditional assessments.
[0018] Step S2: Preprocess the clinical multidimensional data, including data standardization. Standardization eliminates dimensional differences, providing a standardized data foundation for subsequent model input. It should be noted that missing value imputation and sample class balancing (such as SMOTE technique) are performed during the construction phase of the pre-trained malnutrition risk prediction model to address the issues of missing data and scarcity of high-risk samples in the training data. In the application phase of this method (i.e., when assessing a single patient), only standardization is required; missing value imputation or sample balancing is neither necessary nor feasible.
[0019] Step S3: Input the preprocessed clinical multidimensional data into the pre-trained malnutrition risk prediction model to obtain reference information related to the nutritional status of the patient.
[0020] In this embodiment, by receiving multidimensional clinical data covering demographic, disease, treatment, and laboratory dimensions, the reliance on patients' subjective recollections is eliminated. This addresses the technical problem that traditional screening tools cannot accurately assess TBI patients due to impaired consciousness or aphasia, achieving objective risk quantification without patient cooperation. Simultaneously, by training the model using a set of specific core features reflecting neurological function impairment, inflammatory and immune status, and the intensity of treatment interventions, the deficiency of general prediction models in not integrating TBI-specific pathophysiological characteristics is overcome, significantly improving the sensitivity and specificity of nutritional risk prediction for TBI patients. Furthermore, it effectively solves the technical problems of information gaps and uneven distribution of positive and negative samples in clinical data, leading to difficulties in model training convergence or prediction bias, thus improving the model's stability and generalization ability in complex clinical data environments.
[0021] Furthermore, the pre-trained malnutrition risk prediction model is constructed through the following process: collecting historical clinical sample data of patients with traumatic brain injury to form an original dataset; dividing the original dataset into a training set and an internal validation set; training the machine learning model using the training set and validating the model performance using the internal validation set; optionally, using prospectively collected external data to perform external validation of the model to ensure the model's generalization ability.
[0022] Furthermore, the method may also include interpretability analysis of the model output results, using the SHAP method to generate the contribution degree and direction of each core indicator to individual risk prediction, and displaying it to medical staff in a visual form to enhance the clinical credibility and decision support value of the model output.
[0023] Furthermore, in one embodiment, the core feature set includes the following 11 indicators: Body Mass Index (BMI), Activities of Daily Living (ADL) score, Glasgow Coma Scale (GCS) score, dysphagia, gastrointestinal dysfunction, infection, comorbidity index, invasive mechanical ventilation, total cholesterol, hemoglobin, and inflammatory and immune composite indicators.
[0024] Specifically, the aforementioned indicators cover four dimensions: demographics, disease, treatment, and laboratory testing. This application innovatively integrates inflammation-immune composite indicators, comorbidity burden indicators, neurological deficit indicators, and treatment intervention indicators to form a specific malnutrition risk prediction model, overcoming the shortcomings of existing assessments that rely solely on traditional nutritional indicators (such as serum albumin and weight changes).
[0025] Furthermore, in one embodiment, the inflammatory immune composite index is the C-reactive protein to lymphocyte ratio; and the comorbidity index is the Charson comorbidity index.
[0026] Furthermore, in one embodiment, during the model building phase, the missing value imputation process employs the random forest imputation method.
[0027] Specifically, the random forest imputation method utilizes the random forest algorithm to predict and impute missing data. For clinical indicators with missing values, other complete indicators are used as feature inputs, and the missing indicator is used as the target variable to train a random forest regression or classification model to predict and impute the missing values. Compared to traditional methods such as mean imputation or deletion of missing samples, the random forest imputation method can utilize the correlation between features for high-precision imputation, avoiding information loss caused by simply deleting samples. It is particularly suitable for scenarios with complex correlations among multiple indicators in clinical data of traumatic brain injury.
[0028] In one example, the missing value imputation process employs the following strategy: variables with a missing value rate higher than a preset threshold (e.g., 20%) are directly deleted to avoid imputation bias caused by variables with high missing value rates; for variables with a missing value rate lower than or equal to the preset threshold, a random forest algorithm is used for imputation. This algorithm predicts missing values using other complete variables through an iterative random forest model, achieving higher accuracy compared to mean imputation or multiple imputation.
[0029] Furthermore, in one embodiment, the data standardization process employs the Z-score standardization method.
[0030] Specifically, the Z-score standardization method is the process of converting clinical indicators of different dimensions into standard normally distributed data with a mean of 0 and a standard deviation of 1. The calculation formula is: z = (x The formula is μ) / σ, where x is the original index value, μ is the sample mean of the index, and σ is the sample standard deviation of the index. This processing eliminates the influence of dimensional differences among different test items such as body mass index, hemoglobin, and inflammatory markers on model training, accelerating model convergence and improving prediction stability. It should be noted that missing value imputation and sample class balancing (such as SMOTE technique) are performed during the construction phase of the pre-trained malnutrition risk prediction model to address the issues of missing data and scarcity of high-risk samples in the training data. These steps are not required during the application phase of this method.
[0031] Furthermore, in the construction phase of the malnutrition risk prediction model, a SMOTE-based synthetic minority class oversampling technique is used to balance the sample classes of historical training data. Specifically, the SMOTE-based synthetic minority class oversampling technique is a process of synthesizing new samples by linear interpolation of minority class samples (usually malnourished patients). This includes: randomly selecting a minority class sample in the feature space, finding its k nearest neighbor minority class samples, and then randomly generating new synthetic samples along the lines connecting these two neighbors until the number of minority and majority class samples reaches a preset balance ratio. This process effectively solves the class imbalance problem caused by the low proportion of high-risk malnutrition samples in clinical data, prevents prediction bias towards the majority class during model training, and enhances the model's ability to identify high-risk patients.
[0032] In this embodiment, during the model construction phase, a random forest imputation method is employed to accurately fill in missing data using the correlation between features. This avoids information loss caused by simply deleting samples and solves the technical problem of incomplete model input due to missing key indicators in clinical data of traumatic brain injury. By using Z-score normalization, the impact of dimensional differences between different clinical indicators on model training is eliminated, improving the convergence speed and predictive stability of the model in complex clinical data environments. It should be noted that sample class balancing (such as SMOTE) is performed during the construction phase of the pre-trained malnutrition risk prediction model to address the scarcity of high-risk samples in the training data; this step is not required during the application phase of this method. Through the combined use of the above data processing steps (multiple imputation, normalization, and SMOTE class balancing are used during training, while only normalization is used when applying the model), the training quality and generalization ability of the machine learning model are effectively improved. This solves the technical problems of missing key indicators, class imbalance, and dimensional inconsistencies in clinical data of traumatic brain injury, providing a data foundation for the accuracy of subsequent nutritional risk prediction.
[0033] Furthermore, in one embodiment, the malnutrition risk prediction model is established using 11 core prediction indicators as input variables. The model can be implemented using existing algorithms such as logistic regression, decision trees, extreme gradient boosting, support vector machines, gradient boosting machines (GBM), k-nearest neighbors, and neural networks.
[0034] The gradient boosting machine algorithm is an iterative ensemble learning algorithm. Specifically, it sequentially trains multiple weak learners (usually decision trees), each weak learner fitting the residuals of the previous model, and finally weightedly combines the predictions of multiple weak learners to obtain the final prediction output. In this embodiment, the gradient boosting machine algorithm is used to learn the nonlinear mapping relationship between multidimensional clinical data of patients with traumatic brain injury and their malnutrition risk status. Compared with linear models such as logistic regression, the gradient boosting machine can better capture the complex interactions between various clinical indicators; compared with a single decision tree, the gradient boosting machine effectively reduces the risk of overfitting and improves the model's generalization ability through an ensemble strategy.
[0035] In an exemplary study, the study participants were 846 patients with traumatic brain injury admitted to the neurosurgery department of a tertiary-level Class A hospital between October 2022 and October 2024. Inclusion criteria included: ① meeting the diagnostic criteria for traumatic brain injury, with a clear history of head trauma confirmed by cranial CT or MRI; ② age ≥ 18 years; ③ length of hospital stay > 48 hours. Exclusion criteria included: ① patients with pre-existing malignant tumors, end-stage organ failure, or other serious conditions at admission; ② patients who died or discharged themselves within 48 hours of admission. The outcome event was the occurrence of malnutrition, assessed using the Global Leadership in Malnutrition Initiative (GLIM) criteria.
[0036] In this embodiment, 36 candidate predictive variables were initially collected and divided into four dimensions according to their clinical meaning: demographic factors, disease factors, treatment factors, and laboratory indicators. Specifically, this includes: gender, age, BMI, Barthel score, type of TBI, severity of TBI, GCS score, Charlson's Comorbidity Index, brain herniation, infection, multiple injuries, dysphagia, gastrointestinal dysfunction, ICU admission, invasive mechanical ventilation, use of sedatives and analgesics, use of vasoactive drugs, tertiary or higher-level surgery, feeding route, C-reactive protein, white blood cells, lymphocytes, neutrophils, hemoglobin, albumin, prealbumin, total protein, globulin, total cholesterol, fasting blood glucose, neutrophil-to-lymphocyte ratio (NLR), neutrophil-to-albumin ratio (NAR), CLR, C-reactive protein-to-albumin ratio (CAR), glucose-to-lymphocyte ratio (GLR), and C-reactive protein-albumin-lymphocyte index (C-Reactive Protein-to-Lymphocyte ratio). Protein-Albumin-Lymphocyte index, CALLY).
[0037] In the model building phase, this embodiment compared the performance of various machine learning algorithms, including Logistic Regression (LR), Decision Tree (DT), GBM, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Neural Network (NN). A fixed random seed of 42 was set when partitioning the dataset, and the retrospective data was divided into a training set (n=593) and an internal validation set (n=253) in a 7:3 ratio. The training set was used for selecting core predictive indicators and building the risk prediction model, while the internal validation set was used to evaluate the model's ability to identify the risk of malnutrition in TBI patients.
[0038] Based on comprehensive evaluation using the internal validation set, GBM was ultimately selected as the optimal implementation model in this embodiment due to its best balance between discrimination, calibration, and clinical effectiveness. Examples of the optimal hyperparameters for each model are shown in Table 1 below: Table 1 Examples of optimal hyperparameters for each model
[0039] In this embodiment, a malnutrition risk prediction model is constructed using the gradient boosting machine algorithm, which can effectively capture the complex nonlinear relationships among various clinical indicators of patients with traumatic brain injury, improving prediction accuracy compared to linear models. By employing a multi-level validation strategy using training sets and internal validation sets, the model's fitting and generalization abilities are effectively evaluated, preventing overfitting and ensuring the model's stability and applicability in different clinical scenarios.
[0040] Furthermore, in one embodiment, since TBI malnutrition is influenced by multiple factors, the relevant clinical data may exhibit multicollinearity. Therefore, this invention employs a two-stage collaborative feature selection strategy of "linear compression followed by nonlinear evaluation." First, Least Absolute Shrinkage and Selection Operator (LASO) regression is used to linearly compress all candidate features, eliminating irrelevant redundant variables to obtain the first feature set. Subsequently, the Boruta Random Forest algorithm is used to nonlinearly evaluate the feature importance of all candidate features, obtaining the second feature set. Finally, the intersection of the two algorithm selection results is taken to ensure that the predictive factors ultimately included in the model possess both feature importance and simplicity.
[0041] Figure 3The diagram shows the LASSO regression coefficient path. The horizontal axis represents log(λ), and the vertical axis represents the regression coefficients of each candidate variable. As the penalty parameter λ increases, the regression coefficients of each candidate variable gradually shrink towards 0, and the number of variables entering the model decreases. The numbers at the top of the diagram indicate the number of variables with non-zero regression coefficients at different λ levels. The red vertical line represents λ.min, the penalty parameter corresponding to the minimum cross-validation error. From the position of the red line in the diagram, λ.min is located at log(λ) = 4 and The coefficients are between 3 and 4, therefore, 12 non-zero coefficient variables are retained at this point. The blue vertical line represents λ.1se, which is the maximum λ corresponding to the minimum cross-validation error within one standard error range. Its penalty strength is greater, and about 3 variables are retained at this position in the figure. Therefore, compared with the two parameter selection principles, λ.min retains more variables and focuses more on maintaining the model's predictive information, while λ.1se retains only a few of the most representative variables and emphasizes model simplicity. Considering that this study aims to fully screen potential malnutrition risk predictors, λ.min is used to screen out 12 important predictors.
[0042] Figure 4 The Boruta feature importance evaluation plot is shown. The horizontal axis represents candidate predictor variables, and the vertical axis represents variable importance. The box plot reflects the distribution of importance scores for each variable during multiple iterations of the Boruta algorithm. Boruta uses randomly shuffled "shadow variables" as a reference; the green box plot represents variables confirmed by the Boruta algorithm to have significant predictive value. As can be seen from the plot, Boruta ultimately identified 22 important variables, whose overall importance is higher than the highest importance level of the shadow variables.
[0043] Finally, the intersection of the first feature set obtained from LASSO screening and the second feature set obtained from Boruta screening was used to determine 11 core predictive indicators. Specifically, the core predictive indicators finally included in the model are: BMI, invasive mechanical ventilation, CLR, GCS score, activities of daily living score, total cholesterol, Charlson comorbidity index, gastrointestinal dysfunction, dysphagia, hemoglobin, and infection.
[0044] Explanation of the innovativeness of the core forecasting indicators: Current assessments of malnutrition risk in patients with traumatic brain injury typically rely more on traditional nutritional indicators such as BMI, serum albumin, weight change, and food intake, and rarely integrate inflammatory-immune composite indicators, comorbidity burden indicators, neurological deficit indicators, and treatment interventions systematically. This invention innovatively incorporates indicators such as CLR, Charlson's comorbidity index, GCS score, invasive mechanical ventilation, gastrointestinal dysfunction, and swallowing disorders into the same risk prediction model, thereby forming a specific malnutrition risk prediction model.
[0045] Regarding demographic factors, this invention incorporates BMI and activities of daily living (ADL) scores as core predictive indicators. BMI, as a core indicator reflecting the body's nutritional reserves, directly reflects an individual's protein reserves and muscle mass. Traumatic brain injury (TBI) patients experience a high metabolic and catabolic state due to traumatic stress, leading to a significant increase in the body's energy demands. If TBI patients suffer from swallowing difficulties or gastrointestinal dysfunction resulting in insufficient exogenous energy supply, low BMI patients, lacking adequate protein and muscle reserves, are unable to withstand stress-induced losses and are highly susceptible to negative nitrogen balance, rapidly progressing to malnutrition. Patients with low ADL may also have limb motor dysfunction; they are often bedridden or have low activity levels, and skeletal muscles, lacking gravitational load and mechanical contraction stimulation, develop disuse atrophy. As the body's most important protein reserve, continuous loss of muscle leads to a decrease in endogenous protein reserves, easily resulting in malnutrition under stress.
[0046] Regarding disease factors, this invention incorporates indicators such as GCS score, Charlson comorbidity index, infection, gastrointestinal dysfunction, and dysphagia. A low GCS score predicts more severe brain damage. On the one hand, severe TBI triggers a violent stress response, leading to excessive excitation of the sympathetic nervous system, a cascade release of catecholamines, cortisol, and inflammatory factors, and a surge in energy consumption. On the other hand, a low GCS score is accompanied by impaired consciousness; patients are unable to eat orally and rely on tube feeding or parenteral nutrition support, and insufficient energy supply from nutritional support is particularly common in acute TBI patients with impaired consciousness. TBI patients with a high Charlson comorbidity index often have chronic diseases such as diabetes and cardiovascular disease. The coexistence of multiple diseases can lead to a persistent chronic low-grade inflammation, manifested as a mild and persistent increase in the levels of pro-inflammatory cytokines (such as IL-6 and TNF-α). Simultaneously, comorbid patients require multiple medications, and drug interactions and gastrointestinal side effects can suppress appetite and digestive and absorptive functions. Infection is a common complication during hospitalization for patients with total inflammatory lesions (TBI), which can further exacerbate inflammation and protein catabolism, thereby increasing the risk of malnutrition. Infection depletes the body's protein and energy reserves by triggering systemic inflammatory response syndrome. Compared with common infections, multidrug-resistant infections face challenges such as limited first-line antibiotic treatment options and longer infection control periods, resulting in a more prolonged state of high catabolism. Swallowing disorders directly affect patients' oral feeding safety and nutrient intake, while gastrointestinal dysfunction affects enteral nutrient tolerance and absorption; both reflect key aspects of nutrient intake and absorption disorders in TBI patients.
[0047] Regarding therapeutic factors, this invention incorporates invasive mechanical ventilation as a core predictive indicator. During ventilation, increased intrathoracic pressure is transmitted to the abdominal cavity, leading to increased intra-abdominal pressure. This increased intra-abdominal pressure inhibits gastrointestinal motility and slows emptying, causing gastrointestinal dysfunction and resulting in delayed or interrupted feeding. This interruption of exogenous nutrient supply makes it difficult for patients to meet their target energy requirements. Furthermore, the lung-gut axis mechanism indicates that mechanical ventilation can also affect nutrient absorption by reducing intestinal barrier integrity and gut microbiota homeostasis through changes in oxygen supply and neuroendocrine regulation mechanisms.
[0048] In terms of laboratory indicators, this invention incorporates CLR, hemoglobin, and total cholesterol. CLR, or C-reactive protein / lymphocyte ratio, is a composite indicator reflecting inflammatory burden and immunotrophic status. In the acute phase of total inflammatory disease (TBI), patients often experience traumatic inflammatory responses, increased stress metabolism, and immune dysfunction. Using C-reactive protein or lymphocyte count alone is insufficient to fully reflect the cumulative effect of inflammatory depletion and immunosuppression. This invention incorporates CLR into the predictive system, enabling simultaneous characterization of the nutritional and immune imbalance represented by elevated inflammation levels and decreased lymphocyte counts. Hemoglobin reflects protein nutrition and the body's oxygen-carrying capacity, while total cholesterol reflects lipid metabolism and energy reserve status.
[0049] Furthermore, in one embodiment, obtaining reference information related to the patient's nutritional status includes: interpreting the output of the malnutrition risk prediction model using the SHAP method to generate indicator contribution information; the indicator contribution information includes the contribution degree and direction of each indicator in the core feature set to the malnutrition risk probability value.
[0050] Specifically, the SHAP method is an interpretability technique based on the Shapley value principle in game theory, used to explain the output of machine learning models. In this embodiment, for each patient with traumatic brain injury, the SHAP method calculates the marginal contribution of each core feature (such as Glasgow Coma Scale score, dysphagia, C-reactive protein to lymphocyte ratio, etc.) to the patient's malnutrition risk prediction. The contribution level represents the strength of each indicator's influence on the risk prediction result, and the direction of contribution indicates whether each indicator increases or decreases the probability of nutritional risk.
[0051] In an exemplary experiment, Figure 8The SHAP beehive diagram is shown, and the results indicate that indicators contributing significantly to malnutrition prediction include BMI, invasive mechanical ventilation, CLR, GCS score, and activities of daily living (ADL) score. Specifically, high BMI, high GCS score, high ADL score, high total cholesterol, and high hemoglobin tend to decrease the probability of malnutrition prediction; while invasive mechanical ventilation, high CLR, high Charson's comorbidity index, presence of gastrointestinal dysfunction, dysphagia, and infection tend to increase the probability of malnutrition prediction.
[0052] To further clarify how the model is applied at the individual level, this embodiment uses a SHAP waterfall plot to display the individual prediction results for randomly selected patients.
[0053] Figure 9 An individual SHAP waterfall plot of a TBI patient is shown. In this case, invasive mechanical ventilation, gastrointestinal dysfunction, and the Charlson Comorbidity Index were significant risk factors, while BMI and total cholesterol were risk-reducing factors. The contribution information of the indicators generated in this example can help healthcare professionals identify the main sources of malnutrition risk in this patient and provide a reference for targeted nutritional interventions and care measures.
[0054] In this embodiment, the SHAP method is used to interpret the output of the malnutrition risk prediction model, transforming the prediction results of the "black box" machine learning model into the specific contribution and direction of each core indicator to individual risk. This solves the technical problem of low clinical reliability of machine learning models. Healthcare professionals can intuitively understand the sources of risk for each patient through individual SHAP force maps. For example, a patient's high risk may be primarily driven by swallowing difficulties and low hemoglobin, while another patient's high risk may be primarily driven by severe infection and high CLR value. This interpretable output provides quantifiable evidence for healthcare professionals to develop individualized nutritional intervention plans. Furthermore, by outputting nutritional risk probability values and level classifications, hierarchical risk management is achieved, which helps in the rational allocation of clinical resources and the development of early intervention strategies, enhancing the clinical credibility and decision support value of the model output.
[0055] Furthermore, in one embodiment, the reference information includes one or more of the following: malnutrition risk probability value and indicator contribution information.
[0056] Specifically, the malnutrition risk probability value is a continuous value between 0 and 1 output by the malnutrition risk prediction model, representing the likelihood of a patient developing malnutrition. The indicator contribution information is explanatory information output in conjunction with the above probability value or classification, used to explain the basis of the risk prediction results.
[0057] Secondly, this application also provides a system for generating nutritional risk information for patients with traumatic brain injury. (See reference...) Figure 2 , Figure 2 This is a schematic diagram of the structure of the first embodiment of the nutritional risk information generation system for patients with traumatic brain injury according to this application. Figure 2 As shown, the system includes: Data receiving module 1 is used to receive clinical multidimensional data of patients with traumatic brain injury, including demographic indicators, disease-related indicators, treatment-related indicators and laboratory indicators. Data preprocessing module 2 is used to perform data standardization processing on the clinical multidimensional data; Model calculation module 3 is used to input preprocessed clinical multidimensional data into a pre-trained malnutrition risk prediction model to obtain reference information related to the nutritional status of the patient. Information output module 4 is used to output the reference information; The malnutrition risk prediction model is based on a machine learning algorithm and is trained using historical training data that has undergone sample class balancing. The historical training data contains a set of core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention.
[0058] Specifically, the system can be deployed on a hospital server or cloud computing platform, and each module can be implemented through software programs, hardware circuits, or a combination of software and hardware. The data receiving module can interface with the hospital's electronic medical record (EMR) system or laboratory information system (LIS) to achieve automatic data acquisition; the model calculation module can load pre-trained model files to achieve real-time inference.
[0059] Furthermore, in one embodiment, the system further includes an interactive front-end interface for implementing a web-based calculator function.
[0060] Specifically, the interactive front-end interface is developed based on the Shiny framework in the RStudio environment, providing a data entry window, a calculation trigger button, and a result display area. Medical staff can access this web calculator through a browser, input the patient's 11 core indicator values (such as body mass index, GCS score, CLR value, etc.) in the data entry window; after clicking the calculation trigger button, the front end sends the data to the back-end model calculation module; the result display area is used to display a visualization chart (such as a force diagram or waterfall chart) of the SHAP contribution of each indicator of nutritional risk probability value.
[0061] To improve the clinical applicability and convenience of the prediction model, this embodiment uses the Shiny framework in the RStudio environment to transform 11 core prediction indicators and their corresponding risk calculation logic into an interactive web calculator.
[0062] Specifically, Figure 13 The web calculator page is shown. Its PC interface consists of two main parts: ① Patient core indicator input area, where medical staff can input indicators such as BMI, invasive mechanical ventilation, CLR, GCS score, activities of daily living score, total cholesterol, Charlson's comorbidity index, gastrointestinal dysfunction, dysphagia, hemoglobin, and infection according to the patient's actual situation through numerical input boxes, drop-down selection boxes, or sliders; ② Predictive assessment report area, where the system automatically outputs the percentage of malnutrition risk, risk warnings, and indicator contribution charts after receiving input parameters.
[0063] This web-based calculator transforms the TBI malnutrition risk prediction model into an intuitive tool, facilitating the rapid identification of high-risk patients before transfer assessments and nutritional consultations in neurosurgical wards and ICUs.
[0064] In this embodiment, a systematic modular deployment automates the nutritional risk assessment process, reducing human error and improving clinical efficiency. By providing a web-based calculator interface, complex machine learning models are encapsulated into a convenient tool that healthcare professionals can use without programming knowledge. They only need to input routine clinical indicators to obtain quantifiable risk scores and interpretable evidence, significantly improving the ease of application of the technology in clinical settings. This facilitates the rapid development of early screening, risk stratification management, individualized nutritional intervention plans, and nursing decision support.
[0065] Furthermore, the model evaluation was conducted from three dimensions: discrimination, calibration, and clinical effectiveness. Discrimination was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, precision, and F1 score; calibration was evaluated using the calibration curve and Brier score; and clinical effectiveness was evaluated using the clinical decision curve.
[0066] Figure 5 The receiver operating characteristic (ROC) curves for seven models on the internal validation set are shown. The horizontal axis in the figure represents 1. Specificity, with the ordinate representing sensitivity, is calculated by placing the ROC curve closer to the top left corner and increasing the area under the ROC curve (AUC), indicating better model discrimination ability. The results show that the AUC of all seven models is above 0.80, suggesting good overall discrimination ability. Among the various algorithms, Gradient Boosting Machine (GBM) had the highest AUC at 0.869 (95% CI: 0.822–0.915), indicating its superior discriminative ability on its internal validation set. This was followed by Extreme Gradient Boosting Machine (GBM) with an AUC of 0.866 (95% CI: 0.820–0.912); Logistic Regression had an AUC of 0.847 (95% CI: 0.798–0.895); Support Vector Machine (SVM) had an AUC of 0.839 (95% CI: 0.789–0.889); Neural Network (NN) had an AUC of 0.829 (95% CI: 0.777–0.881); K-Nearest Neighbors (KNN) had an AUC of 0.825 (95% CI: 0.772–0.878); and Decision Tree (GRM) had an AUC of 0.822 (95% CI: 0.772–0.872). Overall, GBM exhibited the highest AUC, demonstrating the best discriminative ability.
[0067] Figure 6 The calibration curves of seven models on the internal validation set are shown. The horizontal axis represents the model's average predicted probability, and the vertical axis represents the actual observed incidence rate in the corresponding group. The gray 45° dashed line represents the ideal calibration state; the closer the model curve is to this dashed line, the more consistent the predicted probability is with the actual probability, and the better the calibration performance. The Brier score is also used to quantitatively evaluate calibration performance; a value closer to 0 indicates a smaller prediction error. The results show that the Brier scores of the seven models range from 0.144 to 0.170. The GBM model has the lowest score (Brier=0.144), indicating the best overall consistency between its predicted and actual probabilities. This is followed by the XGBoost model (Brier=0.145), the LR model (0.160), the DT and SVM models (both 0.166), the KNN model (0.167), and the NNet model (Brier=0.170). From the calibration curve trends, all models generally revolve around the ideal reference line, but there are some fluctuations in certain predicted probability ranges. GBM and XGBoost generally show relatively better calibration performance.
[0068] Figure 7Clinical decision curves for seven models on an internal validation set are shown. The horizontal axis represents the threshold probability, i.e., the probability at which further assessment or intervention is initiated when the model-predicted risk of malnutrition reaches a certain level; the vertical axis represents net benefit. The black dashed line "No Intervention" represents no intervention for any study subjects, with a net benefit of 0; the gray dashed line "All Intervention" represents intervention for all study subjects. Generally, when the model's decision curve is above both the "All Intervention" and "No Intervention" reference lines, it indicates that applying the model for clinical decision-making within the corresponding threshold probability range can yield additional net benefit. The results show that most models exhibit some clinical net benefit over a wide threshold probability range. The decision curves of the GBM and XGBoost models are generally at a relatively high level, especially showing stable net benefit within the threshold probability range of approximately 0.20–0.80, suggesting that using these two models to assist in identifying high-risk malnutrition populations has good potential clinical application value.
[0069] In summary, the risk model built based on the TBI-specific core indicator system exhibits good discrimination, calibration, and clinical effectiveness. In this embodiment, the gradient booster model constructed using 11 core predictive indicators as inputs achieved an AUC of 0.869 (95% CI: 0.822–0.915), accuracy of 0.802, sensitivity of 0.815, specificity of 0.793, precision of 0.746, F1 score of 0.779, and Brier score of 0.144 on the internal validation set (n=253), demonstrating clinical net benefit over a relatively wide risk threshold range. This embodiment uses the gradient booster as the optimal implementation model, and the specific evaluation indicators are shown in Table 2 below: Table 2 Internal Validation Evaluation Indicators of the Gradient Booster Model
[0070] We prospectively enrolled 310 patients with traumatic brain injury admitted to the neurosurgery department of a tertiary hospital in Wuhan, Hubei Province, between November 2023 and October 2024. The inclusion and exclusion criteria and data collection methods were the same as those used in the model construction phase.
[0071] Figure 10 The receiver operating characteristic (ROC) curves of the GBM model on the external validation set are shown. The horizontal axis in the figure represents 1. Specificity is represented by the vertical axis, which is sensitivity. The gray diagonal dashed line represents the level of random classification with no discriminative ability. The closer the ROC curve is to the upper left corner, the larger the area under the ROC curve (AUC) value, indicating a stronger ability of the model to distinguish between malnourished and non-malnourished individuals. The results show that the GBM model's AUC on the external validation set is 0.885 (95% CI: 0.849–0.922), significantly higher than 0.5, indicating that the model has good discriminative ability, i.e., it can effectively distinguish between high-risk and low-risk individuals with malnutrition. Compared with the internal validation results, the GBM model's AUC on the internal validation set is 0.869 (95% CI: 0.822–0.915), and the AUC on the external validation set is 0.885. The two are quite close, and the 95% confidence intervals overlap significantly, suggesting that the model maintains good discriminative performance in independent external populations without significant performance degradation, demonstrating good generalization ability and stability. The external validation evaluation metrics are shown in Table 3 below. Table 3 External Validation Evaluation Indicators for Gradient Boosting Machine Model
[0072] The external validation AUC was 0.885 (95% CI: 0.849–0.922), and the Brier score was 0.143, indicating that the predicted probability was in good agreement with the actual observation results. Figure 11 The external validation calibration curve is shown. The Brier score for the GBM model external validation is 0.143, indicating that the predicted probabilities are in good agreement with the actual observations. Figure 12 The clinical decision curve shows that, within the risk threshold range of 0.11 to 0.92, the model's net benefit curve is higher than that of the "all intervention" and "no intervention" strategies, indicating that this core indicator system has good application value in real clinical scenarios.
[0073] This embodiment also recorded indicators such as sensitivity, specificity, accuracy, and F1 score. External validation results showed that the model maintained high specificity while also meeting the sensitivity requirements for clinical screening, effectively reducing the false negative rate.
[0074] In this embodiment, multi-dimensional performance verification data confirms that the technical solution of this application is superior to existing general tools in terms of discrimination, calibration and clinical effectiveness. External verification data also proves that the model has good generalization ability, can adapt to clinical samples at different time periods, and improves prediction accuracy and stability.
[0075] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0076] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.
[0077] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.
[0078] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.
[0079] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.
[0080] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.
[0081] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for predicting the risk of malnutrition in patients with traumatic brain injury, characterized in that, The method includes: Receive clinical multidimensional data from patients with traumatic brain injury, including demographic indicators, disease-related indicators, treatment-related indicators, and laboratory indicators; The clinical multidimensional data are preprocessed, including data standardization. The preprocessed clinical multidimensional data were input into a pre-trained malnutrition risk prediction model to obtain the malnutrition risk probability value related to the patient's nutritional status. The malnutrition risk prediction model is based on a machine learning algorithm and is trained using historical training data that has undergone missing value imputation, data standardization, and sample class balancing. The historical training data contains a set of core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention. The set of core features includes features that reflect basic nutritional reserves, features that reflect the degree of neurological function impairment, features that reflect the inflammatory and immune status, and features that reflect the intensity of treatment intervention.
2. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The core feature set Including the following indicators: Body mass index, activities of daily living score, Glasgow Coma Scale score, dysphagia, gastrointestinal dysfunction, infection, comorbidity index, invasive mechanical ventilation, total cholesterol, hemoglobin, and inflammatory and immune composite indicators.
3. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 2, characterized in that, The inflammatory immune composite index is the ratio of C-reactive protein to lymphocytes; The comorbidity index is the Charson Comorbidity Index.
4. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The missing value imputation process uses the random forest imputation method.
5. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The data standardization process employs the Z-score standardization method.
6. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The malnutrition risk prediction model is constructed based on the gradient booster algorithm.
7. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The malnutrition risk prediction model is constructed through the following steps: Data was imputed using random forest and standardized using Z-score. SMOTE technology was used to balance the class of historical clinical sample data to generate a balanced training set. LASSO regression was used to perform linear feature compression on all candidate features in the historical clinical sample data to obtain the first feature set. The Boruta algorithm is used to perform a nonlinear importance evaluation on all candidate features in the historical clinical sample data to obtain a second feature set; The intersection of the first feature set and the second feature set is determined as the core feature set; The machine learning model is trained using the balanced training set and the core feature set to obtain the malnutrition risk prediction model.
8. The method for predicting the risk of malnutrition in patients with traumatic brain injury as described in claim 1, characterized in that, The obtained reference information related to the patient's malnutrition status includes: The SHAP method is used to interpret the output of the malnutrition risk prediction model and generate indicator contribution information. The indicator contribution information includes the degree and direction of contribution of each indicator in the core feature set to the probability value of malnutrition risk.
9. A method for constructing a model to predict the risk of malnutrition in patients with traumatic brain injury, characterized in that, include: Historical clinical sample data of patients with traumatic brain injury were obtained, and the historical clinical sample data had characteristics such as missing values, inconsistent data scaling, and uneven category distribution. To address the missing value feature, a random forest-based missing value imputation method was used to process the data; to address the data scaling inconsistency, Z-score was used for standardization; and to address the uneven class distribution feature, SMOTE technology was used to balance the historical clinical sample data and generate a balanced training set. The historical clinical sample data is linearly compressed using LASSO regression to obtain a first feature set; the historical clinical sample data is nonlinearly evaluated using the Boruta algorithm to obtain a second feature set; the intersection of the first feature set and the second feature set is determined as the core feature set. The machine learning model is trained using the balanced training set to obtain the malnutrition risk prediction model.
10. A system for predicting the risk of malnutrition in patients with traumatic brain injury, characterized in that, The system includes: The data receiving module is used to receive clinical multidimensional data of patients with traumatic brain injury, including demographic indicators, disease-related indicators, treatment-related indicators, and laboratory indicators. The data preprocessing module is used to preprocess the clinical multidimensional data, and the preprocessing includes data standardization. The model calculation module is used to input the preprocessed clinical multidimensional data into the pre-trained malnutrition risk prediction model to obtain the malnutrition risk probability value related to the patient's nutritional status. An information output module is used to output the reference information; The malnutrition risk prediction model is based on a machine learning algorithm and is trained using historical training data that has undergone missing value imputation, standardization, and sample class balancing. The historical training data contains a set of core features that reflect neurological function impairment, inflammatory and immune status, and the intensity of treatment intervention. The set of core features includes features that reflect basic nutritional reserves, features that reflect the degree of neurological function impairment, features that reflect the inflammatory and immune status, and features that reflect the intensity of treatment intervention.
