Multi-index-based burn patient sepsis risk prediction system and method

Abstract

The invention relates to the technical field of medical data processing, in particular to a burn patient sepsis risk prediction system and method based on multiple indexes, and the method comprises the following steps: S1, obtaining multi-dimensional clinical data of a burn patient at the early stage of admission, the multi-dimensional clinical data comprises a physiological function score, a burn severity score, a complication index and a laboratory inspection index; s2, inputting the multi-dimensional clinical data into a pre-trained sepsis prediction logistic regression model, and obtaining a sepsis occurrence probability P of the patient; s3, outputting a sepsis occurrence probability P of the patient, comparing the probability P with a clinical risk threshold, and when the probability P is higher than the clinical risk threshold, performing high-risk early warning; the method is used for critical burn adult patients, and accurate sepsis risk prediction is achieved in the early hospital admission stage of the patients.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, in particular to a burn patient sepsis risk prediction system and method based on multiple indexes. BACKGROUND

[0002] Sepsis is a very serious and common complication of critically ill adult burn patients, and is also the main cause of death in such patients. Therefore, early identification of high-risk patients with sepsis and timely intervention are of great significance for reducing patient mortality and improving their prognosis.

[0003] Currently, the prediction of sepsis after burns in clinical practice mainly relies on the following methods: first, clinical experience judgment. Doctors make subjective judgments based on patients' vital signs such as body temperature, heart rate, and respiration, as well as clinical symptoms, relying on their own experience. However, this method lacks objective quantitative standards and often cannot confirm infection until it occurs, which can easily miss the valuable opportunity for early intervention. Second, single index monitoring. Common monitoring indicators include white blood cell count, procalcitonin (PCT), C-reactive protein (CRP), etc. However, these indicators have the problem of weak specificity and are easily affected by the inflammatory response caused by burns, resulting in a high false positive rate. Third, existing general scoring systems. For example, SOFA, APACHE II, etc. However, these scoring systems are not specifically designed for burn populations and do not include burn-specific severity indicators (such as Baux score, inhalation injury) and key changes in coagulation function, so there are certain limitations in prediction accuracy.

[0004] Most existing prediction methods only indicate the risk after infection occurs, and cannot provide early warning, which is lagging. Secondly, single indicators or general models have insufficient predictive efficiency for sepsis in burn patients, and have the problem of low accuracy. Finally, the comprehensiveness is poor, and there is currently a lack of a specialized prediction tool that can integrate physiological function, burn severity, nutritional status, coagulation dysfunction, and other multidimensional information.

[0005] Therefore, the present application provides a technical solution that is specifically designed for critically ill adult burn patients and can accurately predict sepsis risk in the early stages of patient admission. SUMMARY

[0006] To solve the above problems, the present application provides a burn patient sepsis risk prediction system and method based on multiple indexes, which is used for critically ill adult burn patients and meets the need for accurate sepsis risk prediction in the early stages of patient admission.

[0007] In order to achieve the above object, the technical scheme of the present application is as follows: a burn patient sepsis risk prediction method based on multiple indexes, comprising the following steps: S1, obtaining multi-dimensional clinical data of a burn patient at admission, the multi-dimensional clinical data including physiological function score, burn severity score, complication index and laboratory test index;

[0008] Among them, the physiological function score includes sequential organ failure estimation score and Barthel index score; the burn severity score adopts Baux score; the complication index includes whether there is inhalation injury and whether tracheostomy is performed; the laboratory test index includes white blood cell count, hemoglobin, platelet count, albumin, lactate dehydrogenase, activated partial thromboplastin time, prothrombin time, antithrombin-III and D-dimer, and the values of the laboratory test index have been converted into categorical variables according to predetermined rules;

[0009] S2, inputting the multi-dimensional clinical data into a pre-trained sepsis prediction logistic regression model to obtain the probability P of the patient developing sepsis;

[0010] S3, outputting the probability P of the patient developing sepsis, and comparing the probability P with a clinical risk threshold value, when the probability P is higher than the clinical risk threshold value, a high-risk warning is given.

[0011] Further, in S2, the sepsis prediction logistic regression model is as follows:

[0012] ;

[0013] Among them, is an intercept term, , , …, is the regression coefficient corresponding to each multi-dimensional clinical data, , , …, is the multi-dimensional clinical data after preprocessing and variable selection.

[0014] Further, the establishment method of the sepsis prediction logistic regression model comprises the following steps, S2.1, collecting the clinical data set of historical burn patients and the outcome label of whether sepsis occurs, and establishing a training data set based on the outcome label;

[0015] S2.2, quartile grouping is performed on continuous laboratory test indexes to convert them into ordered categorical variables; then a variable selection strategy is adopted to combine with the clinical evaluation standard to screen out the most predictive variable combination from the training data set;

[0016] S2.3, based on the screened variable combination, the final prediction model is trained using the logistic regression algorithm, and the regression coefficient of each variable is determined.

[0017] Further, in S1 to S3, the applicable patients are adult patients aged > 18 years old, and the patients are hospitalized within 24 hours of being burned and have not received antibiotic treatment.

[0018] Further, in S2.2, the method for screening the most predictive combination of variables from the training data set is as follows: binary Logit regression, stepwise regression, Lasso regression screening, Boruta feature variable screening, and screening of independent risk factors for sepsis after burn in adult critically ill patients combined with clinical experience.

[0019] Further, in S2.2 to S2.3, internal and external validation methods are also used to verify the prediction performance and clinical practical value of the model.

[0020] Further, internal validation uses Bootstrap self-sampling 1000 times for verification.

[0021] Further, in S2.2, considering that the SOFA score table content covers liver and kidney function, and the Baux score is also an important score in clinical practice, the variables of the final prediction model include SOFA score, D-dimer, and Baux score value, followed by Logistic regression and nomogram drawing.

[0022] Further, a burn patient sepsis risk prediction system based on multiple indicators includes a data acquisition module, a model establishment module, and a warning reminder module.

[0023] The data acquisition module is used to input and obtain multi-dimensional clinical data.

[0024] The model establishment module is used to input and train the sepsis prediction logistic regression model in the above-mentioned burn patient sepsis risk prediction method based on multiple indicators.

[0025] The warning reminder module is used to compare the probability P with the clinical risk threshold, and when the probability P is higher than the clinical risk threshold, a display reminder is given.

[0026] Further, the data acquisition module further includes a data screening module, which is used to convert laboratory test indicators into categorical variables according to the predetermined rules in the above-mentioned burn patient sepsis risk prediction method based on multiple indicators.

[0027] Further, the data acquisition module is also used to establish data association based on the sepsis prediction logistic regression model, compare the data association with the multi-dimensional clinical data, and when the multi-dimensional clinical data is missing, send a missing reminder instruction to the warning reminder module.

[0028] Further, the display reminding mode of the early warning reminding module includes text display reminding and chart display reminding.

[0029] The above scheme has the following beneficial effects:

[0030] Compared with the prior art, the present application has the following remarkable advantages:

[0031] Early prediction: only the data in the early stage of hospitalization is needed to predict, which saves valuable time for clinical intervention.

[0032] High precision: physiological function, burn severity, complications, coagulation function and other multi-dimensional indicators are fused, especially the coagulation-related indicators (AT-III, D-Dimer, etc.) processed by quartile grouping are introduced, which greatly improves the specificity and sensitivity of the prediction according to the pathophysiological characteristics of post-burn sepsis.

[0033] Strong specificity: the sepsis prediction model designed for burn patients is more targeted than the general scoring system, and the prediction result is more reliable.

[0034] Good practicability: the variables required by the model are all clinical routine collection items, without the need for additional detection, and are easy to popularize. Through line graph and other forms, the use method is simple and intuitive, and the doctor can complete the risk assessment within 1 minute.

[0035] Objective quantification: avoids the error of subjective judgment, provides data support for clinical decision-making, and helps to realize precision medicine.

[0036] Additional aspects and advantages of the application will be given in part in the following description, some of which will become apparent from the following description, or will be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A LASSO regression result diagram of a burn patient sepsis risk prediction system and method embodiment based on multiple indexes of the application;

[0038] Figure 2 Another LASSO regression result diagram of a burn patient sepsis risk prediction system and method embodiment based on multiple indexes of the application;

[0039] Figure 3 Importance diagram of different features of a burn patient sepsis risk prediction system and method embodiment based on multiple indexes of the application;

[0040] Figure 4 Change curve diagram of different features of a burn patient sepsis risk prediction system and method embodiment based on multiple indexes of the application;

[0041] Figure 5 Nomogram chart of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0042] Figure 6 ROC curve chart of the training set and test set of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0043] Figure 7 Calibration chart of the training set of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0044] Figure 8 Calibration chart of the test set of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0045] Figure 9 Clinical decision curve DCA chart of the training set and test set of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0046] Figure 10 Test set result confusion matrix of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application;

[0047] Figure 11 ROC curve chart of three models of the multi-index based burn patient sepsis risk prediction system and method embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] The specific embodiments will be described in detail below:

[0050] A multi-index based burn patient sepsis risk prediction method, comprising the following steps: S1, obtaining multi-dimensional clinical data of burn patients in the early stage of hospitalization, the multi-dimensional clinical data including physiological function score, burn severity score, complication index and laboratory test index;

[0051] wherein the physiological function score includes Sequential Organ Failure Assessment (SOFA) score and Barthel Index (BI) score; the burn severity score adopts Baux score; the complication index includes whether there is inhalation injury and whether tracheostomy is performed; the laboratory test index includes white blood cell count (WBC), hemoglobin (HGB), platelet count (PLT), albumin (ALB), lactate dehydrogenase (LDH), activated partial thromboplastin time (APTT), prothrombin time (PT), antithrombin-III (AT-III) and D-dimer (D-Dimer), wherein the values of the laboratory test index are all converted into categorical variables according to predetermined rules (such as the quartiles of the patient cohort).

[0052] S2, inputting the multi-dimensional clinical data into a pre-trained sepsis prediction logistic regression model to obtain a probability P of the patient developing sepsis, wherein the sepsis prediction logistic regression model is as follows:

[0053] ;

[0054] wherein, is an intercept term, , , …, is a regression coefficient corresponding to each multi-dimensional clinical data, , , …, is the multi-dimensional clinical data after preprocessing and variable selection.

[0055] And the establishment method of the sepsis prediction logistic regression model includes the following steps, S2.1, collecting the clinical data set of historical burn patients and the outcome label of whether sepsis occurs, and establishing a training data set based on the outcome label.

[0056] S2.2, quartile grouping is performed on continuous laboratory test indexes to convert them into ordered categorical variables; then a variable selection strategy is adopted to combine with clinical evaluation standards to screen out the most predictive variable combination from the training data set.

[0057] The method for screening out the most predictive variable combination from the training data set is as follows, binary Logit regression, stepwise regression, Lasso regression screening, Boruta feature variable screening are performed on the training data set, independent risk factors for sepsis occurrence after burn in adult critical patients are screened out by combining with clinical experience, a nomogram prediction model for the risk of sepsis occurrence after burn in adult critical burn patients is constructed by using R software, the area under the ROC curve is used to evaluate the discrimination of the model, and the calibration curve and the clinical decision curve are used to evaluate the calibration and clinical practicability of the model, respectively.

[0058] At the same time, considering that the SOFA score table in the clinic covers liver and kidney function, the Baux score is also an important score in the clinic, and the final prediction model includes SOFA score, D-dimer, Baux score value, and then performs Logistic regression and draws nomogram.

[0059] S2.3, based on the screened variable combination, using a logistic regression algorithm to train the final prediction model, and determining the regression coefficient of each variable.

[0060] S3, output the probability P of the patient developing sepsis, and compare the probability P with the clinical risk threshold (such as P>0.3), when the probability P is higher than the clinical risk threshold, a high-risk warning is performed.

[0061] In S2.2 to S2.3, internal validation (Bootstrap self-sampling 1000 times) and external validation (test set) are also used to verify the prediction performance and clinical practical value of the model.

[0062] A multi-index-based sepsis risk prediction system for burn patients, comprising: a data acquisition module for entering and obtaining multi-dimensional clinical data; the data acquisition module further comprises a data screening module, and the data screening module is used for converting laboratory test indexes into classification variables according to the predetermined rules in the multi-index-based sepsis risk prediction method for burn patients.

[0063] The data acquisition module is also used to establish data association based on the sepsis prediction logistic regression model, compare the data association with the multi-dimensional clinical data, and send a missing alert instruction to the early warning reminding module when the multi-dimensional clinical data is missing.

[0064] The model establishment module is used to enter and train the sepsis prediction logistic regression model in the multi-index-based sepsis risk prediction method for burn patients.

[0065] The early warning reminding module is used to compare the probability P with the clinical risk threshold, and when the probability P is higher than the clinical risk threshold, a display reminder is performed. The display reminding mode of the early warning reminding module includes text display reminding and chart display reminding.

[0066] The experimental method of the embodiment is as follows:

[0067] I. Inclusion criteria:

[0068] (1) This embodiment is illustrated by burn patients in the Affiliated Hospital of Nantong University, aged > 18 years old, and the patients were hospitalized within 24 hours of being burned and had not received antibiotic treatment;

[0069] (2) Meet the diagnostic criteria for severe and extremely severe burns in "Surgery" (9th edition);

[0070] (3) Patients with burn sepsis meet the diagnostic criteria of sepsis-3.0;

[0071] (4) The patient's related information and medical record are complete.

[0072] II. Exclusion criteria:

[0073] (1) Incomplete medical records;

[0074] (2) Patients with tumors, expected life <6 months;

[0075] (3) Patients with severe primary diseases of liver, kidney, hematopoietic system, etc. before admission, or with abnormal liver and kidney function, AST, ALT greater than 1.5, 2 times the normal value, creatinine greater than the upper limit of normal value; patients with mental illness;

[0076] (4) Pregnant, lactating or preparing for pregnancy women;

[0077] (5) Patients with severe active infection or major blood, kidney, metabolic, gastrointestinal or endocrine function disorders before admission;

[0078] (6) History of malignant tumor in the past 10 years (except non-melanoma skin cancer, cervical carcinoma in situ, breast ductal carcinoma in situ or stage 1 prostate cancer);

[0079] (7) Patients with a history of drug abuse, drug abuse or alcoholism;

[0080] (8) Patients who have not been treated for 6 weeks after major organ surgery, and patients with poor surgical wound healing;

[0081] III. Diagnosis method:

[0082] (1) Sepsis-3.0, which was published in February 2016, International Consensus Definition of Sepsis and Septic Shock, Third Edition (Sepsis 3.0). The definition of sepsis is the uncontrolled response of the body to infection leading to life-threatening organ dysfunction. In simple terms, the definition of sepsis 3.0 is: positive infection indicators and SOFA score greater than 2 points. In this study, the infection indicators are identified as positive bacterial culture results.

[0083] (2) The diagnostic criteria for severe and very severe burns are based on the definition in the 9th edition of Surgery: Severe burns: total area between 30% and 49% TBSA or third-degree burn area between 10% and 19% TBSA, or total area not exceeding 30% TBSA, but with one of the following conditions: ① severe systemic condition or shock; ② combined injury or combined injury (such as severe trauma, chemical poisoning, etc.); ③ moderate or severe inhalation injury. Very severe burns: total area more than 50% TBSA or third-degree burn area more than 20% TBSA.

[0084] Four, grouping method:

[0085] The collected patient data were randomly divided into training set and test set. A total of 137 cases of critically ill burn patients were included in the training set, and another 56 cases of critically ill burn patients were included in the test set. The training set patient data were used for the construction and internal evaluation of the clinical prediction model, and were used to verify the test set patients, establish prediction probability, and complete the verification of the model combined with the actual situation of the test set patients with burn sepsis. According to the diagnostic criteria of sepsis-3.0 proposed in 2016, the clinical data of the two groups of patients were divided into sepsis group and non-sepsis group according to whether they had burn sepsis.

[0086] Five, research variable factors:

[0087] Various predictors have shown different results in previous studies. By consulting the data and combining the results of the clinical laboratory report, the white blood cell count (White blood cell, WBC), hemoglobin (Haemoglobin, Hb), platelet count (Platelet, PLT), albumin (Albumin, ALB), albumin / globulin ratio (A / G), lactate dehydrogenase (Lactate dehydrogenase, LDH), fasting blood glucose (Blood glucose, Glu), activated partial thromboplastin time (APTT), prothrombin time (PT), D-dimer (D-Dimer), and antithrombin-Ⅲ (AT-Ⅲ) values were selected as reference indicators. At the same time, the predictive significance of age, gender, burn cause, whether combined with inhalation injury, whether treated with mechanical ventilation, burn area, third-degree burn area, Baux score, and BI score on adult severe burn patients with burn sepsis was analyzed.

[0088] Six, the verification process is as follows:

[0089] (1) According to the inclusion and exclusion criteria, 78 cases of adult severe and critical burn patients with burn sepsis and 118 cases of adult severe and critical burn patients with non-burn sepsis who were treated in the Affiliated Hospital of Nantong University from January 2012 to October 2022 were included, and were randomly divided into training set (70%) and test set (30%).

[0090] (2) The general clinical data and laboratory serological test clinical indicators of the patients were analyzed at baseline, and the serum experimental related indicators were divided into Q1, Q2, Q3, and Q4 according to the quartile. The data of the patients in the training set were screened by binary Logit regression, stepwise regression, Lasso regression, and Boruta feature variable screening. Finally, the independent risk factors for the occurrence of burn sepsis in adult critical patients after burn were screened out combined with clinical experience.

[0091] (3) The nomogram prediction model for the incidence of burn sepsis in adult critical burn patients was constructed using R software. The area under the ROC curve (Area Under Curve, AUC value) was used to evaluate the discrimination of the model. The calibration curve and clinical decision curve were used to evaluate the calibration and clinical utility of the model, respectively. The baseline conditions of the patients in the training set and the test set were compared, as shown in Tables 1 to 4.

[0092] At the same time, internal validation (Bootstrap self-sampling 1000 times) and external validation (test set) were used to verify the prediction performance and clinical utility of the model.

[0093] Table 1. Comparison of baseline conditions of patients in training set and test set

[0094]

[0095]

[0096]

[0097] The demographic and clinical characteristics of 193 patients (137 in the training set and 56 in the test set) are shown in Table 1. The two groups were comparable in age distribution (p = 0.419), with approximately 80% of patients aged <60 years. The burn area distribution was balanced (p = 0.463), with a majority of patients in the 30-79.9% interval (76.2%). Notably, there were no statistically significant differences between the two groups in terms of full-thickness burn area ≥20% TBSA (54.4% vs 53.3%, p = 0.625) and the incidence of inhalation injury (83.9% vs 77.4%, p = 0.308).

[0098] Table 2. Comparison of baseline characteristics between sepsis and non-sepsis patients

[0099]

[0100]

[0101]

[0102] The data in Table 2 reveals significant differences between the sepsis group (97 cases) and the non-sepsis group (96 cases): demographic characteristics: the proportion of patients aged ≥60 years was higher in the sepsis group (25.8% vs 13.5%, p = 0.033), and the proportion of males was higher (73.2% vs 59.4%, p = 0.042); burn characteristics: the proportion of patients with full-thickness burns ≥20% TBSA was significantly higher in the sepsis group (66.0% vs 42.7%, p = 0.001), and the proportion of patients with 90-99.9% large-area burns was 23.7% (vs 3.1%, p < 0.001); laboratory indicators: the sepsis group showed significantly higher proportions of patients in the high-value intervals for WBC (Q4 49.5% vs 1.0%), LDH (Q4 48.5% vs 2.1%), and DD (Q4 47.4% vs 3.1%) (all p < 0.001), while the proportion of patients in the low-value interval for the protective indicator ALB (Q1 47.4% vs 0%) was higher, and the difference between the two groups was statistically significant, allowing for feature selection.

[0103] Table 3. Results of binary Logit regression analysis

[0104]

[0105] From Table 3, four independent risk factors were identified: SOFA score (OR = 1.95, 95% CI 1.045-3.638): each increase of 1 point, the risk of sepsis increased by 95%; WBC grade (OR = 10.387, 95% CI 1.245-86.685): high-grade patients increased the risk by 9.4 times; ALB grade (OR = 0.04, 95% CI 0.003-0.607): high grade had significant protective effect; LDH grade (OR = 16.752, 95% CI 1.489-188.446): the highest grade of patients increased the risk by 15.8 times. The goodness-of-fit index of the model was excellent (Nagelkerke R²=0.977), indicating that these factors could explain 97.7% of the variation in the occurrence of sepsis.

[0106] Table 4. Results of stepwise regression analysis (n=193)

[0107]

[0108] Table 5. Model summary table

[0109]

[0110] From Table 4 and Table 5, the model contains four variables: WBC grade (β = 0.130, p < 0.001): standardized coefficient 0.292; ALB grade (β = -0.155, p < 0.001): standardized coefficient -0.345; LDH grade (β = 0.149, p < 0.001): standardized coefficient 0.332; D-Dimer grade (β = 0.048, p = 0.003): standardized coefficient 0.108 The model determination coefficient R²=0.850, adjusted R²=0.847, D-W value 2.226 indicates no autocorrelation, and VIF <3 shows no multicollinearity problem.

[0111] Meanwhile, based on the LASSO regression results of Figure 1 and Figure 2 , the most predictive variables were selected by L1 regularization: when the value of λ is taken as the minimum value (logλ≈-3), 7 non-zero coefficient variables are retained, and as λ increases, the variables are gradually compressed to 4 core predictors. The selected λ value corresponds to the simplest model within 1 standard error, balancing the prediction accuracy and model complexity.

[0112] And based on Figure 3 and Figure 4The Boruta feature variable screening results of the Baux score, D-Dimer, and SOFA score were identified as the true relevant features (green), the irrelevant variables were identified as noise (red), and the provisional variables were identified as noise (blue). The results were consistent with the previous statistical methods, which enhanced the credibility of the feature selection.

[0113] Meanwhile, considering that the SOFA score table covers liver and kidney function, and the Baux score is also an important score in clinical practice, the final variables included in the model were SOFA score, D-Dimer, and Baux score value. Logistic regression was performed and a nomogram was drawn, as shown in Figure 5

[0114] Based on the Figure 5 nomogram, when the Total Points (total score) is in the interval below 80 points, the corresponding risk value is usually less than 0.2 (i.e., 20%), which belongs to the low-risk stage, indicating that the patient's prognosis is relatively good. When the total score enters the interval of 80 to 140 points, the risk enters a stage of rapid climb, which is the most sensitive part of the relationship. Within this interval, a small increase in the total score will cause a significant jump in the risk value, and the risk probability will increase rapidly from about 20% to about 70%. This means that patients in this score interval may have a significant change in prognosis risk due to a slight deterioration of their clinical indicators. When the total score exceeds 140 points, it enters the high-risk plateau, where the risk value is generally more than 0.7 (70%) and gradually approaches 0.9 (90%). In this stage, the risk increase brought by the further increase in the total score will be relatively slower, but the patient is already in a very dangerous situation.

[0115] Based on the ROC curve of the training set and the test set, as shown in Figure 6 , the AUC of the training set was 0.942 (95% CI: 0.906-0.978), the sensitivity was 96.3%, and the specificity was 94.7%; the AUC of the test set was 0.942 (0, 879-1.000), the sensitivity was 92.0%, and the specificity was 90.3%. The AUC values of the two groups of data were close and significantly higher than 0.5, indicating that the model had strong ability to distinguish between positive and negative examples, and the performance consistency of the training set and the test set was good, suggesting that the model had stable generalization ability.

[0116] Further Figure 7 and Figure 8 ​The calibration curves of the training and test sets show that the model predictions are highly consistent with actual observations: Training set: Brier score 0.062, slope 0.98 (ideal value 1); Test set: Brier score 0.071, slope 0.95; Key interval: threshold probability 10-90%; Peak benefit 35.2% (40% threshold), meeting the accuracy requirements for clinical decision-making.

[0117] Then from Figure 9 The DCA plot of clinical decision curves for the training and test sets shows that the model has a significant net benefit: within the threshold probability range, both the training set (blue curve) and the test set (pink curve) are higher than the two reference lines "All" (all intervention, gray curve) and "None" (no intervention, black curve) in the 10-80% interval; in terms of peak net benefit, the test set shows better performance at around 40% threshold (i.e., it can reduce the corresponding number of erroneous decisions per 100 cases); and in terms of clinical applicability, the model covers the commonly used clinical intervention threshold range of 20-50%, which can provide effective reference support for clinical decision-making within this range.

[0118] By constructing a test set confusion matrix, such as Figure 10 As shown, the model's specific classification performance in the binary classification task is as follows: True positives (TP) numbered 24, meaning the model successfully identified 24 patients who actually had the target disease (such as sepsis); True negatives (TN) numbered 27, indicating the model correctly excluded 27 individuals who were not infected; False positives (FP) numbered 5, meaning 5 healthy individuals were incorrectly predicted as infected, reflecting a certain level of false alarms; False negatives (FN) numbered 2, indicating that only 2 individuals who actually had the disease were missed by the model. Based on this, the model exhibits high diagnostic accuracy: its specificity is as high as 84.4% (27 / 32), indicating that the model has high reliability in identifying negative samples and can effectively exclude non-target cases; at the same time, the precision reaches 82.8% (24 / 29), which means that when the model makes a positive prediction, the result has high credibility. In addition, the model's recall rate also reached 92.3% (24 / 26), which is particularly outstanding, indicating that the model can capture the vast majority of true positive cases and reduce the risk of missed diagnoses to an extremely low level. These metrics demonstrate that the model not only excels in overall accuracy but also achieves a good balance in its ability to identify positive and negative samples. High recall ensures high disease detection rates, while high specificity and precision guarantee the reliability of diagnostic results. This balanced and excellent performance makes the model a highly valuable diagnostic tool for clinical applications.

[0119] Based on the above results, ROC curves for the three models were plotted, as follows: Figure 11The performance of the models is ranked as follows: the model built in this study (Model A) has an AUC of 0.942, the SOFA score model (Model B) has an AUC of 0.935, and the Baux score model (Model C) has an AUC of 0.775. It can be seen that the model built in this study has the largest AUC value and the best discrimination.

[0120] The data was then input into the sepsis prediction logistic regression model for verification, resulting in Table 6, which shows the binary Logit regression analysis results as follows:

[0121]

[0122] The model formula is: ln(p / 1-p) = -5.808 + 0.479*SOFA + 1.472*DD_Degree-0.012*Baux (where p represents the probability of Sepsis being 1, and 1-p represents the probability of Sepsis being 0).

[0123] Table 6 shows the binary Logit regression analysis results of the final sepsis prediction model. The model includes three key predictors: SOFA score, D-Dimer classification (DD_Degree), and Baux score. The analysis shows that SOFA score and D-Dimer classification have significant predictive value for the occurrence of sepsis (p<0.001), while Baux score does not show statistical significance in this model (p=0.407).

[0124] Specifically, for every 1-point increase in SOFA score, the risk of sepsis increases by 61.5% (OR=1.615, 95% CI 1.308-1.994). The predictive effect of D-Dimer is more significant, with high-grade (Q4) patients having a risk of sepsis that is 4.359 times higher than low-grade (Q1) patients (95% CI 2.428-7.828). The model goodness-of-fit index performs well, with Nagelkerke R² reaching 0.709, indicating that the model can explain 70.9% of the variation in the occurrence of sepsis.

[0125] It is worth noting that although the Baux score had predictive value in univariate analysis (see Table 2), its predictive role was superseded by the SOFA score and D-Dimer in the multivariate model. This may be because the age factor in the Baux score is already included in the SOFA score, while the influence of burn area is reflected by more direct inflammatory markers such as D-Dimer. The model formula ln(p / 1-p)=-5.808 + 0.479*SOFA + 1.472*DD_Degree-0.012*Baux provides a quantitative risk assessment tool for clinical practice, where the intercept term -5.808 indicates that the baseline risk is extremely low when all predictors are 0 (OR=0.003).

[0126] Therefore, this study constructed a burn sepsis risk prediction model by systematically analyzing data from 193 patients with severe burns (137 in the training set and 56 in the test set). Table 1 shows that the two groups of patients were comparable in terms of baseline characteristics such as age and burn area (p>0.05), ensuring the reliability of subsequent analyses. Table 2 reveals significant characteristics of sepsis patients: elderly patients (≥60 years old, 25.8% vs 13.5%), male patients (73.2% vs 59.4%), and patients with larger third-degree burn areas (≥20% TBSA, 66.0% vs 42.7%) were more prone to sepsis, and were also accompanied by significantly elevated levels of WBC, LDH, and D-Dimer (all >47% in Q4) and decreased ALB (47.4% in Q1).

[0127] Multivariate statistical analysis, using binary logit regression in Table 3, identified four independent risk factors: SOFA score (OR=1.95), WBC grade (OR=10.387), LDH grade (OR=16.752), and the protective ALB grade (OR=0.04). Stepwise regression in Table 4 further validated the predictive value of WBC, ALB, LDH, and D-Dimer grades (all p<0.01), with a model explanatory power of 84.7% (adjusted R²). During feature selection, LASSO regression (… Figure 1 , 2 ) and Boruta algorithm ( Figure 3 , 4 The SOFA score, D-Dimer score, and Baux score were used to corroborate each other and were ultimately selected to construct the model based on clinical significance.

[0128] Column chart ( Figure 5 These three indicators are integrated into a visualization tool, allowing clinicians to quickly assess risk through simple scoring. Model validation shows excellent performance: the AUC on both the training and test sets reaches 0.942 (…). Figure 6), the calibration curve slope was close to 1 (0.98 for training set and 0.95 for test set), and the Brier score was lower than 0.08 ( Figures 7-8 ), indicating that the prediction was highly consistent with the actual risk, and the decision curve demonstrated its clinical utility ( Figure 9 ), and the net benefit was significantly higher than the "all intervention" or "no intervention" strategy within the threshold interval of 20-50%. The confusion matrix of the test set ( Figure 10 ) further showed that the sensitivity was 92.3% and the specificity was 84.4%, and the positive / negative predictive values were both over 82%. In the test set, compared with the existing scoring systems ( Figure 11 ), the model of this study (AUC 0.942) was significantly better than the SOFA score (0.935) and the Baux score (0.775) alone.

[0129] In summary, this study established a burn sepsis prediction model with excellent performance through rigorous statistical analysis and multidimensional verification. Its innovations are reflected in: 1. Integrating multiple organ function, coagulation status and burn severity; 2. Using multiple methods to ensure the reliability of feature selection; 3. Realizing convenient clinical application through nomogram. The model provides a powerful tool for early intervention of burn sepsis.

[0130] Obviously, the above embodiments are only examples for clarity, and do not limit the embodiments. Based on the above description, those skilled in the art can make other different forms of changes or modifications. Here, it is not necessary and impossible to exhaust all the embodiments. The obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for predicting the risk of sepsis in burn patients based on multiple indicators, characterized in that, Includes the following steps: S1. Obtain multi-dimensional clinical data of burn patients upon admission. The multi-dimensional clinical data includes physiological function scores, burn severity scores, complication indicators, and laboratory test indicators. The physiological function scores included the sequential organ failure estimation score and the Barthel index score; the burn severity score used the Baux score; the complication indicators included whether there was inhalation injury and whether a tracheotomy was performed; the laboratory test indicators included white blood cell count, hemoglobin, platelet count, albumin, lactate dehydrogenase, activated partial thromboplastin time, prothrombin time, antithrombin-III, and D-dimer, and the values ​​of the laboratory test indicators were all converted into categorical variables according to predetermined rules; S2. Input multidimensional clinical data into a pre-trained logistic regression model for sepsis prediction to obtain the probability P of the patient developing sepsis. S3. Output the probability P of the patient developing sepsis, and compare the probability P with the clinical risk threshold. When the probability P is higher than the clinical risk threshold, issue a high-risk warning.

2. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 1, characterized in that, In S2, the logistic regression model for sepsis prediction is shown below: ； in, For the intercept term, , , ..., The regression coefficients are the corresponding values ​​for each of the multidimensional clinical data. , , ..., This refers to multidimensional clinical data after preprocessing and variable selection.

3. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 2, characterized in that, The method for establishing a logistic regression model for sepsis prediction includes the following steps: S2.1, collect clinical datasets of historical burn patients to determine whether sepsis has occurred, and establish a training dataset based on the outcome labels; S2.

2. Group the continuous laboratory test indicators into quartiles and convert them into ordered categorical variables; Then, a variable selection strategy combined with clinical assessment criteria was used to select the combination of variables with the greatest predictive power from the training dataset; S2.

3. Based on the selected variable combinations, use the logistic regression algorithm to train the final prediction model and determine the regression coefficients of each variable.

4. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 3, characterized in that, In S1 to S3, the applicable patients are adults over 18 years of age who were hospitalized within 24 hours of being burned and did not receive antibiotic treatment.

5. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 4, characterized in that, In S2.2, the method for selecting the most predictive combination of variables from the training dataset is as follows: the training dataset is subjected to binary Logit regression, stepwise regression, Lasso regression screening, and Boruta feature variable screening, and independent risk factors for the occurrence of sepsis after burns in adult critically ill patients are screened in combination with clinical experience.

6. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 5, characterized in that, In S2.2, the method for screening independent risk factors for the occurrence of burn sepsis in critically ill adult patients based on clinical experience is as follows: a nomogram prediction model of the risk of burn sepsis in critically ill adult patients is constructed using R software, the discriminant power of the model is evaluated using the area under the ROC curve, and the calibration curve and clinical decision curve are used to evaluate the calibration and clinical applicability of the model, respectively.

7. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 6, characterized in that, In S2.2 to S2.3, both internal and external validation methods will be used to verify the model's predictive performance and clinical applicability.

8. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 7, characterized in that, Internal validation was performed using Bootstrap self-sampling 1000 times.

9. The method for predicting the risk of sepsis in burn patients based on multiple indicators according to claim 8, characterized in that, In S2.2, considering that the SOFA score covers liver and kidney function in clinical practice, and the Baux score is also an important score in clinical practice, the variables of the final prediction model include SOFA score, D-dimer, and Baux score. Logistic regression is then performed and a nomogram is plotted.

10. A multi-indicator-based sepsis risk prediction system for burn patients, characterized in that, include: The data acquisition module is used to input and acquire multi-dimensional clinical data; The model building module is used to input and train the sepsis prediction logistic regression model in the multi-indicator-based method for predicting the risk of sepsis in burn patients as described in any one of claims 1-9. The early warning and reminder module is used to compare the probability P with the clinical risk threshold. When the probability P is higher than the clinical risk threshold, a reminder will be displayed.

11. The multi-indicator-based sepsis risk prediction system for burn patients according to claim 10, characterized in that, The data acquisition module also includes a data filtering module, which is used to convert laboratory test indicators into categorical variables according to predetermined rules in the multi-indicator-based method for predicting the risk of sepsis in burn patients as described in any one of claims 1-9.

12. The multi-indicator-based sepsis risk prediction system for burn patients according to claim 11, characterized in that, The data acquisition module is also used to establish data associations based on the sepsis prediction logistic regression model, compare the data associations with multi-dimensional clinical data, and send a missing data reminder instruction to the early warning module when multi-dimensional clinical data is missing.

13. The multi-indicator-based sepsis risk prediction system for burn patients according to claim 12, characterized in that, The warning and alert module displays alerts in two ways: text and chart.

Images