Construction and application of candida patient death risk prediction model

Abstract

The invention relates to the technical field of biological medicine, in particular to construction and application of a candida patient death risk prediction model. The model is based on traditional SOFA scoring, core variables are screened through single-factor and multi-factor logistic analysis and LASSO regression analysis, respiratory system, blood coagulation function and circulatory system SOFA scoring and key indexes such as the maximum lactic acid value, the minimum albumin value and the maximum blood urea nitrogen (BUN) are integrated, and a prediction model is constructed through logistic regression. Through multi-center data verification, the AUC of an internal verification queue reaches 0.826, the AUC of an external verification queue is 0.813, and a calibration curve shows that the predicted death rate is highly consistent with the actual death rate. Patients can be divided into a high risk group and a low risk group through a 13-score threshold value, and the 28-day survival rate difference is significant (plt; 0.05) of the method. An efficient and convenient early risk assessment tool is provided for critical candida patients, and clinical precise intervention and prognosis improvement are assisted.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine technology, and in particular to the construction and application of a mortality risk prediction model for candidemia patients. Background Art

[0002] Candidemia is a serious, life-threatening bloodstream infection caused by Candida species, primarily affecting critically ill patients in the intensive care unit (ICU). Despite advances in antifungal therapy and supportive care, its mortality rate remains high (30%-60%). Early diagnosis and timely initiation of appropriate antifungal therapy are crucial to improving prognosis. However, nonspecific clinical manifestations, delayed laboratory confirmation, and the growing problem of antifungal resistance make this goal challenging.

[0003] In clinical practice, the Acute Physiology and Chronic Health Evaluation II (APACHE II) score is widely used to predict mortality risk in critically ill patients, including those with candidemia. However, its numerous variables and computational complexity limit its routine use in the ICU. In contrast, the Sequential Organ Failure Assessment (SOFA) score, designed to assess organ dysfunction in patients with sepsis, is more concise. However, it fails to capture the metabolic disturbances and inflammatory responses specific to candidemia, resulting in poor predictive efficacy (the SOFA score has an area under the curve (AUC) for predicting Candida albicans only of 0.77, see PMC7811217). Previous studies have attempted to improve the SOFA score by incorporating the Charlson Comorbidity Index (CCI) or introducing other biomarkers. However, these approaches, while increasing complexity, have failed to effectively address the systemic inflammation, tissue hypoperfusion, and renal stress caused by fungal burden that are specific to candidemia. Therefore, the development of reliable prognostic tools to accurately assess patient outcomes, guide clinical decision-making, and optimize treatment strategies is urgent. Summary of the Invention

[0004] The present invention first provides a method for constructing a mortality risk prediction model for candidemia patients, comprising the following steps:

[0005] (1) Data acquisition and preprocessing: Clinical data of patients with candidemia, including demographic information, SOFA score sub-items, laboratory indicators, and survival outcomes, were obtained from a multicenter database. Cases with more than 50% missing values ​​were eliminated, and the remaining missing values ​​were processed through multiple imputation.

[0006] (2) Key variable screening: Univariate and multivariate logistic analysis and LASSO regression were used to screen core variables from traditional SOFA score sub-items and candidate biomarkers. LASSO regression parameters were optimized through five-fold cross-validation to determine respiratory system SOFA score, coagulation function SOFA score, circulatory system SOFA score, maximum lactate value, minimum albumin value, maximum blood urea nitrogen value, and minimum bicarbonate value as predictors;

[0007] (3) Model construction and optimization: Based on the logistic regression algorithm, the screened variables were integrated according to the 0-4 score standard to construct the mSOFA_3 model;

[0008] (4) Model validation: The model performance was evaluated using the internal validation set and the external test set, and the predictive efficacy and clinical practicality were verified through the ROC curve, AUC value, calibration curve, and decision curve.

[0009] The multicenter database includes the MIMIC-III, MIMIC-IV and ICU-JN databases, covering data on patients with candidemia from June 2001 to January 2024.

[0010] The SOFA score is the maximum SOFA score recorded within 24 hours after diagnosis, including respiratory system score, coagulation function score, liver function score, cardiovascular score, Glasgow Coma Scale (GCS) and renal function score; the laboratory indicators include maximum white blood cell count, minimum hemoglobin, minimum platelet, maximum lactate, minimum PaO2 / FiO2 ratio, minimum bicarbonate, maximum partial thromboplastin time, maximum international normalized ratio, maximum prothrombin time, maximum blood urea nitrogen, maximum creatinine, maximum alanine aminotransferase, maximum aspartate aminotransferase, and maximum bilirubin.

[0011] The screened variables were combined in different ways to construct different models. Four machine learning algorithms, including random forest, logistic regression, support vector machine, and extreme gradient boosting, were used to predict the mortality risk of patients with candidemia. The optimal prediction model was determined based on the obtained AUC.

[0012] The present invention also provides a mortality risk prediction model for candidemia patients constructed using the above method, wherein the model comprises the following predictive factors: respiratory system SOFA score, coagulation function SOFA score, circulatory system SOFA score, maximum lactate value, minimum albumin value, and maximum blood urea nitrogen value.

[0013] The scoring criteria for each predictor are:

[0014] Respiratory system (PaO2 / FiO2, mmHg): <100 = 4 points, <200 = 3 points, <300 = 2 points, <400 = 1 point, ≥400 = 0 points;

[0015] Coagulation function (platelet 10 9 / L): <20 counted as 4 points, <50 counted as 3 points, <100 counted as 2 points, <150 counted as 1 point, ≥150 counted as 0 points;

[0016] Circulatory system: dopamine (μg / Kg / min) > 15 or epinephrine > 0.1 or norepinephrine > 0.1 is scored as 4 points; dopamine (μg / Kg / min) > 5 or epinephrine ≤ 0.1 or norepinephrine ≤ 0.1 is scored as 3 points; dopamine (μg / Kg / min) ≤ 5 or any dose of dobutamine (μg / Kg / min) is scored as 2 points; mean arterial pressure (mmHg) < 70 is scored as 1 point; mean arterial pressure (mmHg) ≥ 70 is scored as 0 points;

[0017] Maximum lactate score: >7.0mmol / L = 4 points, 4.0-7.0mmol / L = 3 points, 2.0-4.0mmol / L = 2 points, 1.2-2.0mmol / L = 1 point, ≤1.2mmol / L = 0 points;

[0018] Minimum albumin value score: <2.0 g / dL is 4 points, 2.0-2.5 g / dL is 3 points, 2.5-3.0 g / dL is 2 points, 3.0-3.5 g / dL is 1 point, and ≥3.5 g / dL is 0 points;

[0019] Maximum blood urea nitrogen score: >80 mg / dL is 4 points, 60-80 mg / dL is 3 points, 40-60 mg / dL is 2 points, 20-40 mg / dL is 1 point, and <20 mg / dL is 0 points.

[0020] The mortality risk prediction model for patients with candidemia was determined based on the AUC value obtained, and the best prediction model was named mSOFA_3 (Note: mSOFA_3 only represents one name in the application).

[0021] mSOFA_3=0.5×Resp_SOFA+0.7×Coag_SOFA+0.6×Circ_SOFA+0.9×Lactate_score+0.8×BUN_score+0.6×Albumin_score;

[0022] Patients were divided into high-risk and low-risk groups using the mSOFA_3 score with a cutoff value of 13 points;

[0023] The present invention also provides an application of the above-mentioned candidemia patient death risk prediction model in predicting the death risk of candidemia patients, comprising the following steps:

[0024] (1) Collect predictor data within 24 hours of admission;

[0025] (2) Calculate the scores of each indicator according to the scoring criteria in claim 4 and add them up to obtain the total score of mSOFA_3;

[0026] (3) When the total score is >13 points, the patient is judged as a high-risk patient and the predicted risk of death within 28 days is higher; when the total score is ≤13 points, the patient is judged as a low-risk patient and the predicted risk of death within 28 days is lower.

[0027] The predictive factors include respiratory system SOFA score, coagulation function SOFA score, circulatory system SOFA score, maximum lactate value, minimum albumin value, and maximum blood urea nitrogen value.

[0028] The predictor data include the worst index value within 24 hours after the diagnosis of candidemia, and the worst index value is the maximum value or the minimum value.

[0029] The beneficial effects of the present invention are:

[0030] 1. This study developed a modified SOFA (mSOFA) score for candidemia. By replacing the original SOFA renal score with BUN and integrating markers such as lactate and albumin, as well as key SOFA sub-scores such as respiratory, circulatory, and coagulation, this score overcomes the limitations of the traditional score. Model training and internal validation were performed using data from the MIMIC-III and MIMIC-IV databases, and external testing and validation was performed with the ICU-JN database. Advanced statistical and machine learning techniques were employed to construct and evaluate the model, enabling the SOFA score to more accurately reflect the unique pathophysiological characteristics of ICU patients with candidemia. Multicenter validation demonstrated an AUC of 0.826 in the internal validation cohort and 0.813 in the external validation cohort. Calibration curves showed high agreement between predicted and actual mortality rates. This represents a significant improvement over the reported AUC of 0.77 for the SOFA score combined with other indicators for predicting Candida albicans. A threshold of 13 points allowed patients to be stratified into high- and low-risk groups, with significant differences in 28-day survival (p < 0.05). The present invention provides an efficient and convenient early risk assessment tool for patients with critical candidemia, and provides a more reliable and clinically applicable tool for early risk assessment and intervention.

[0031] 2. Accurate risk stratification: Patients were divided into high- and low-risk groups using the mSOFA_3 score (cutoff 13 points), with significant differences in 28-day survival (Kaplan-Meier analysis, p<0.05). Validated in a multicenter database, the model demonstrated robustness across diverse healthcare settings and is suitable for early risk assessment and personalized treatment decisions for ICU patients with candidemia.

[0032] 3. Clinical decision support: The model provides early warning (such as increased lactate and BUN indicating an intervention window) to guide antifungal treatment and resource allocation.

[0033] 4. Simplified operation process: only routine detection indicators are required (no special equipment is required), and the scoring calculation time is less than 5 minutes. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Figure 1 is an overview of the study design of the present invention: (a) univariate and multivariate analysis of clinical data for variable selection; (b) LASSO analysis uses selected clinical variables in combination with SOFA score for further variable selection to construct an improved SOFA score model; (c) model training using logistic regression (LR), random forest (RF), support vector machine (SVM) and extreme gradient boosting (XGBoost) algorithms; (d) model validation in internal and test cohorts; (e) model evaluation and clinical application.

[0035] Figure 2 Flowchart for patient selection.

[0036] Figure 3 Figure 1 shows the results of using LASSO regression analysis to identify variables in the modified SOFA model.

[0037] Figure 4 Figure 2. Comparison of variable combinations in different mSOFA models with the original SOFA score. The x-axis represents the number of variables, and the y-axis represents the model.

[0038] Figure 5 Figure 3 is the receiver operating characteristic (ROC) curve and performance index results of the mSOFA model.

[0039] Figure 6 Figure 3 Calibration curves of the mSOFA_3 model using logistic regression in the internal validation (A) and test (B) cohorts: the dashed line indicates ideal calibration, in which the predicted probabilities match the observed outcomes; in each figure, the blue line indicates apparent calibration, while the red line indicates bias-corrected calibration after bootstrapping (B = 1000 replicates); for the internal validation cohort (A), the mean absolute error (MAE) was 0.017, n = 162; for the test cohort (B), the MAE was 0.03, n = 145.

[0040] Figure 7 Figure 3 shows the decision curve analysis (DCA) results of the mSOFA_3 model: A is the DCA for the internal validation cohort; B is the DCA for the external validation cohort.

[0041] Figure 8 Figure 3 Kaplan-Meier survival curves stratified by mSOFA_3 score in the internal validation and external validation cohorts: patients were divided into low-risk (mSOFA ≤ 13) and high-risk (mSOF A > 13) groups, where a is the survival curve of the internal validation cohort and b is the survival curve of the test cohort. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. It should be noted that if no specific conditions are specified in the examples, they are carried out under normal conditions. The drawings described below are only some embodiments of the present application and do not limit the present invention in any form. For those skilled in the art, several adjustments and improvements can be made without departing from the concept of the present invention. All technologies implemented based on the above content of the present invention belong to the scope of protection of the present invention.

[0043] The present invention relates to abbreviations as follows: CKD: chronic kidney disease; HF: heart failure; COPD: chronic obstructive pulmonary disease; SOFA: sequential organ failure assessment; GCS: Glasgow Coma Scale; HR: heart rate; MAP: mean arterial pressure; SBP: systolic blood pressure; RR: respiratory rate; WBC: white blood cell count; PT: prothrombin time; PTT: partial thromboplastin time; INR: international normalized ratio; BUN: blood urea nitrogen; ALT: alanine aminotransferase; AST: aminotransferase; RR: respiratory rate; BUN: blood urea nitrogen; GCS: Glasgow Coma Scale.

[0044] Example 1: Construction and evaluation of a mortality risk prediction model for patients with candidemia

[0045] 1. Data Source

[0046] The data for this study came from two large open-source databases: MIMIC-III (v1.4) and MIMIC-IV (v1.0), and a local dataset: the Intensive Care Unit of Jining First People's Hospital (ICU-JN) database.

[0047] MIMIC-III version 1.4 contains detailed medical records for 46,520 patients hospitalized at Beth Israel Deaconess Medical Center in Boston, Massachusetts, between June 2001 and October 2012. MIMIC-IV version 2.0 covers data on nearly 300,000 hospitalized patients between 2008 and 2019. Because there is some overlap between the MIMIC-III and MIMIC-IV databases, to achieve comprehensive coverage, we extracted data from MIMIC-IV between 2012 and 2019 and from MIMIC-III between 2001 and 2008. Relevant clinical data included patient demographics, vital signs, laboratory results, microbiological culture results, medication records, procedure records, and survival outcomes. Because both MIMIC-III and MIMIC-IV data are publicly available and anonymized, no additional ethical approval was required. The ICU-JN database was obtained from the Donghua electronic medical record system of Jining First People's Hospital, with data covering the period from December 1, 2020, to January 31, 2024. The use of the ICU-JN data was approved by the Ethics Committee of Jining First People's Hospital (approval number: KYLL-202308-134). Given the retrospective nature of this study, the Ethics Committee agreed to waive informed consent from patients. All research methods adhered to relevant guidelines and regulations, including the Declaration of Helsinki.

[0048] 2. Study Population

[0049] This study focused on critically ill adult patients diagnosed with candidemia, defined as the presence of Candida species detected in at least one blood culture, accompanied by clinical symptoms and signs of infection. Inclusion criteria included: (1) age ≥18 years at ICU admission; (2) complete SOFA score data. Exclusion criteria included: (1) ICU stay <24 hours; (2) patients with concurrent bacteremia within 7 days before or after candidemia detection; and (3) pregnant or breastfeeding patients. For patients with multiple ICU admissions and candidemia, only the first admission record was included.

[0050] 3. Data extraction and study design Figure 1-2 )

[0051] Data extraction was completed using Navicat software (v15.1), and record integration was completed using STATA software (v18.0), merging records based on codes such as hadm_id, stay_id, or icustay_id. The data extraction process is detailed in Figure 2 , we extracted the following variables within 24 hours after candidemia diagnosis:

[0052] (1) Demographic information at admission: age, gender;

[0053] (2) Vital signs within 24 hours: maximum heart rate (HR_max), minimum heart rate (HR_min), minimum systolic blood pressure (SBP_min), minimum diastolic blood pressure (DBP_min), minimum mean arterial pressure (MAP_min), maximum body temperature (temperature_max), maximum and minimum respiratory rates (RR_max, RR_min), and minimum blood oxygen saturation (SPO2_min);

[0054] (3) Laboratory examination indicators: maximum white blood cell count (WBC_max), minimum hemoglobin (hemoglobin_min), minimum platelet (platelet_min), maximum lactate (lactate_max), minimum PaO2 / FiO2 ratio (PaO2 / FiO2_min), minimum bicarbonate (bicarbonate_min), maximum partial thromboplastin time (PTT_max), maximum international normalized ratio (INR_max), maximum prothrombin time (PT_max), maximum blood urea nitrogen (BUN_max), maximum creatinine (creatinine_max), maximum alanine aminotransferase (ALT_max), maximum aspartate aminotransferase (AST_max), maximum bilirubin (bilirubin_max);

[0055] (4) Vasoactive drugs used within 24 hours: dopamine, epinephrine, norepinephrine, dobutamine, vasopressin;

[0056] (5) Initiation of antifungal therapy: voriconazole, amphotericin B, caspofungin, posaconazole;

[0057] (6) comorbidities, including hypertension, chronic obstructive pulmonary disease, diabetes, myocardial infarction, and solid tumors, identified based on ICD-9 and ICD-10 codes at discharge;

[0058] (7) The maximum SOFA score recorded within 24 hours after diagnosis, including respiratory system score, coagulation function score, liver function score, cardiovascular score, Glasgow Coma Scale (GCS) and renal function score.

[0059] The primary outcome of this study was death within 28 days after the diagnosis of candidemia. Patients were considered alive if there was no record of death within 28 days.

[0060] Clinical data were integrated from the MIMIC-III and MIMIC-IV databases (see Figure 1First, univariate logistic regression was used to analyze and screen out statistically significant prognostic variables (p < 0.05), and these variables were included in multivariate logistic regression to determine their independent prognostic value ( Figure 1 a) in the above.

[0061] The screened significant variables were combined with the SOFA score components to construct a modified SOFA score (mSOFA) to improve its accuracy in predicting the prognosis of patients with candidemia. The clinical data were then standardized, while the components of the SOFA score remained unchanged, and the LASSO regression method was used to screen the key variables for constructing the mSOFA score ( Figure 1 b) in the above example.

[0062] The mSOFA dataset was randomly divided into a training set (70%) and an internal validation set (30%). The ICU-JN data set was used as an external validation set to assess the generalization ability of the model. All variables were then scored, and four different combinations of mSOFA models were constructed, combining the variables selected by the LASSO and various SOFA indices, to determine the optimal prediction model.

[0063] We use four machine learning algorithms, random forest (RF), logistic regression (LR), support vector machine (SVM), and extreme gradient boosting (XGBoost) to evaluate the performance of each model ( Figure 1 c) and draw the receiver operating characteristic curve (ROC), calculate the AUC value, and select the model with the highest AUC as the optimal model ( Figure 1 d) in the above.

[0064] After model training, we performed calibration curve analysis to assess the consistency between the predicted and actual mortality risks. Decision curve analysis (DCA) was also performed to evaluate the clinical utility of the model at different thresholds.

[0065] Finally, we performed survival analysis to further validate the prognostic value of the model in patients with candidemia ( Figure 1 e) in the.

[0066] During data preprocessing, we used the visualization method proposed by Templ (R package "VIM") to identify the distribution of missing values. Cases with more than 50% missing data and incomplete SOFA scores were excluded. Some variables, such as D-dimer, procalcitonin, minimum pH, maximum pH, maximum / minimum blood glucose, and C-reactive protein, were also removed due to missing data exceeding 30%. The missing data for all variables are shown in Table 1. The remaining missing values ​​were handled using multiple imputation using the R package "mice."

[0067] Table 1. Summary of missing data in the training and validation cohorts (n = 561)

[0068]

[0069] 4. Statistical Analysis

[0070] Descriptive statistical analyses were performed for all categorical and continuous variables. To assess comparability among the training, validation, and test sets, we compared the overall distribution of candidate variables across the three data sets. Categorical variables were presented as frequencies and percentages, while continuous variables were summarized as mean ± standard deviation or median (interquartile range), depending on their distribution. Between-group comparisons of categorical variables were performed using the chi-square test or Fisher's exact probability method, while two-group comparisons of continuous variables were performed using the independent sample t-test or Mann–Whitney U test, depending on their distribution. For comparisons of variables between the three groups, one-way analysis of variance (ANOVA) was used for normally distributed variables, and the Kruskal–Wallis test was used for non-normally distributed variables.

[0071] To identify significant predictors of mortality in patients with candidemia, we performed univariate and multivariate logistic regression analyses. The results were used to identify clinically relevant characteristics for inclusion in the modified SOFA score model. All statistical analyses were performed using R software (version 4.1.2), and p < 0.05 was considered statistically significant.

[0072] 5. Model construction and selection

[0073] After completing data preprocessing, we applied LASSO regression analysis with five-fold cross-validation to identify the most predictive features from selected clinical variables and the original SOFA score components. LASSO analysis incorporates an L1 regularization penalty, which compresses the regression coefficients of unimportant variables to zero, thereby enabling feature selection and preventing model overfitting. Using the R package "glmnet," we determined the optimal lambda parameter by minimizing the cross-validation error, striking a balance between model fit and complexity.

[0074] We then combined the selected variables in various ways and constructed multiple predictive models based on various machine learning algorithms, including random forest (RF), logistic regression (LR), support vector machine (SVM), and extreme gradient boosting (XGBoost), to determine the optimal model. The hyperparameters of each model were adjusted using ten-fold cross-validation to ensure robust parameter selection. The models were trained on the training set and evaluated on both internal and external validation sets. We measured model performance using multiple evaluation metrics, including the area under the receiver operating characteristic curve (AUC), F1 score, recall, accuracy, and precision. The model with the highest AUC was selected as the final optimal model.

[0075] 6. Model Evaluation and Application

[0076] To assess the accuracy and clinical utility of the model, calibration curves and decision curve analysis (DCA) were constructed. The calibration curves assessed the consistency between the model's predicted mortality and observed values, and the DCA analyzed the model's net clinical benefit at different risk thresholds. Finally, survival analysis was performed using the Kaplan-Meier method. Patients were divided into high-risk and low-risk groups based on the mSOFA score, and the log-rank test was used to compare survival differences between the two groups. The statistical significance level was set at p < 0.05. All statistical analyses were performed in R software (version 4.1.2).

[0077] 7. Baseline characteristics

[0078] A total of 561 patients with candidemia were included in this study, of which 267 were from the MIMIC-III database and 299 were from the MIMIC-IV database. The data used for testing and validation came from the ICU-JN database of Jining First People's Hospital, which included a total of 145 patients with candidemia. Patients from the MIMIC-III and MIMIC-IV databases were randomly divided into a training set (n = 392) and an internal validation set (n = 169) in a ratio of 7:3. The detailed process of patient screening is shown in Figure 2 shown.

[0079] 8. Baseline Characteristics and Univariate Analysis

[0080] The baseline characteristics and univariate analysis results of the candidemia survival and death groups in the MIMIC-III and MIMIC-IV databases are shown in Table 2 .

[0081] Table 2. Baseline characteristics and univariate analysis of 28-day all-cause mortality in patients with candidemia

[0082]

[0083]

[0084]

[0085]

[0086] Results showed that the all-cause 28-day mortality rate reached 62.21% (349 / 561 patients). The mean age of the nonsurvivors was significantly higher than that of the survivors (64.16±15.55 years vs. 58.09±15.64 years). The nonsurvivors had significantly lower peak body temperature (37.78°C vs. 38.15°C) and lowest mean arterial pressure (45.98 mmHg vs. 51.43 mmHg), and a higher rate of vasopressor use (75.6% vs. 55.7%), suggesting hemodynamic instability. In terms of coagulation function, the partial thromboplastin time (PTT: 63.20s vs. 40.55s) and prothrombin time (PT: 17.30s vs. 14.85s) in the death group were significantly prolonged, the international normalized ratio (INR: 1.60 vs. 1.30) was increased, and the platelet count (137.12×10^9 / L vs. 198.96×10^9 / L) was decreased. Comorbidities had limited impact on mortality, while organ dysfunction was more significant in the death group: alanine aminotransferase (ALT: 758.37 IU / L vs. 304.65 IU / L), aspartate aminotransferase (AST: 2041.05 IU / L vs. 611.09 IU / L), lactate (3.70 mmol / L vs. 2.20 mmol / L), and blood urea nitrogen (BUN: 61.64 mg / dL vs. 36.72 mg / dL) were significantly increased, and albumin (2.31 g / dL vs. 2.66 g / dL) was significantly decreased (all p < 0.05).

[0087] 9. Prediction model construction and evaluation

[0088] To identify predictors of mortality in patients with candidemia, this study first identified clinically significant variables through univariate analysis. Multivariate logistic regression models were then used for further analysis. Results showed that seven variables—age, maximum respiratory rate (RR_max), maximum body temperature (temperature_max), maximum lactate (lactate_max), minimum bicarbonate (bicarbonate_min), minimum albumin (albumin_min), and maximum urea nitrogen (BUN_max)—were significantly associated with the risk of mortality (Table 3). Data were collected from the training, internal, and external validation cohorts. Table 4 shows the baseline characteristics of patients in each cohort. In addition to comparing the survival and mortality groups, we systematically assessed the distribution of baseline characteristics across the three cohorts. Key clinical variables (such as age, vital signs, laboratory parameters, and organ function scores) were well-consistent across the cohorts: the mean age ranged from 61.6 to 62.7 years, and lactate, albumin, urea nitrogen levels, and the SOFA score showed similar distribution trends (Table 5). This result shows that the clinical characteristics of each cohort are evenly distributed, which provides support for the comparability and extrapolation of the model. Based on the above findings, we combined the six dimensional indicators of the original SOFA score and further optimized the variable selection through LASSO regression to construct a modified SOFA (mSOFA) model for predicting death in patients with candidemia. When the λ value takes the optimal solution (lambda.min=0.03710371), LASSO regression screened out three original SOFA indicators (respiratory system, coagulation system and cardiovascular system scores) and five clinical variables (highest urea nitrogen, lowest bicarbonate, lowest albumin, highest lactate value and age) as the factors with the greatest predictive value ( Figure 3 On this basis, we constructed four mSOFA derivative models by combining different variables and determined mSOFA_3 as the best model. The specific model variable structure is detailed in Figure 4. The mSOFA_3 score replaces the renal function indicators in the original SOFA score with urea nitrogen. Although serum creatinine and urine volume are still the basic indicators for diagnosing acute kidney injury (AKI) according to standard guidelines, their predictive efficacy for mortality in critically ill patients may be affected by clinical interventions. Diuretics, fluid resuscitation, or renal replacement therapy may interfere with urine volume assessment, thereby limiting its reliability in predicting mortality. Although serum creatinine is an excellent marker for assessing renal function, blood urea nitrogen (BUN) can more comprehensively reflect the patient's inflammatory state, protein catabolism, and nutritional status, making it a more sensitive predictor of mortality risk in critically ill patients. Multicenter clinical studies have confirmed that elevated BUN levels can independently predict poor prognosis in ICU patients and are superior to serum creatinine and estimated glomerular filtration rate (eGFR) in predicting mortality. At the same time, this study removed liver and GCS_SOFA and introduced lactate and albumin. Albumin, the most abundant protein in human serum, can neutralize candidalysin, the cytolytic toxin of Candida albicans, thereby mitigating tissue damage. Low albumin levels reflect both inadequate nutrient reserves and an increased inflammatory response, both of which are closely associated with adverse clinical outcomes in critically ill patients. Lactate is a classic marker of tissue hypoxia and metabolic dysfunction; elevated lactate levels indicate insufficient tissue oxygenation and cellular metabolic stress. Our modified model, which additionally incorporates metabolic (lactate), nutritional / inflammatory (albumin), and toxin load (BUN), provides a more comprehensive reflection of the critically ill patient's overall status and improves the sensitivity of assessing adverse outcomes in candidemia. Furthermore, these markers can be obtained earlier, enabling earlier prediction of mortality risk and prompting of treatment decisions, potentially reducing mortality risk. Finally, excluding GCS_SOFA avoids assessment bias caused by sedation or non-neurological disorders, making it particularly suitable for patients with infectious diseases and atypical altered mental status.

[0089] Table 3. Multivariate logistic regression analysis of potential factors associated with 28-day mortality in patients with candidemia (n = 561, 160 survived, 331 died)

[0090]

[0091]

[0092] Table 4. mSOFA variables in the training set, internal validation set, and external validation set

[0093]

[0094]

[0095] Table 5. Comparison of the overall distribution of mSOFA-related variables in the training set, internal validation set, and external validation set

[0096]

[0097]

[0098] Logistic regression (LR), random forest (RF), support vector machine (SVM), and XGBoost algorithms were used to evaluate model performance. The LR-based mSOFA_3 model performed best: AUC 0.826 on the internal validation set.

[0099] (95% CI: 0.759-0.892), F1 score 0.825, recall 0.758, accuracy 0.763, precision 0.904 ( Figure 5 ad in the external validation set); AUC 0.813 (95% CI: 0.740-0.886), F1 score 0.788, recall 0.744, accuracy 0.752, precision 0.838 ( Figure 5 The performance of other models is detailed in Table 6.

[0100] Table 6. Predictive performance of each model in the internal and external validation cohorts.

[0101]

[0102] The calibration curve of the mSOFA_3 model showed a high degree of agreement between the predicted and observed mortality rates (mean absolute error 0.017 for the internal validation set and 0.03 for the external validation set; see Figure 6 Decision curve analysis (DCA) showed that the model had a net clinical benefit within the threshold probability range of 15%-85% (inside) and 20%-90% (outside) (see Figure 7 ), confirming its clinical applicability.

[0103] Example 2: Application of a mortality risk prediction model for candidemia patients

[0104] To improve clinical practicality, lactate, BUN, and albumin were converted into a 0-4 score scale according to severity (see Table 7), and the original SOFA respiratory, coagulation, and circulation scores were retained. Based on the median cutoff value (13 points), patients were divided into a low-risk group (≤13 points) and a high-risk group (>13 points). The Kaplan-Meier survival curve showed that the 28-day survival rate of the low-risk group was significantly better than that of the high-risk group (see Figure 8 , where a is the survival curve of the internal validation cohort and b is the survival line of the external validation cohort).

[0105] Table 7. Modified Sequential Organ Failure Assessment (mSOFA) score

[0106]

Claims

1. A method for constructing a mortality risk prediction model for patients with candidemia, characterized in that: The following steps are involved: (1) Data acquisition and preprocessing: Clinical data of patients with candidemia were obtained, including demographic information, vital signs within 24 hours, SOFA score, laboratory indicators, and survival outcomes. Cases with more than 50% missing values ​​were eliminated, and the remaining missing values ​​were processed through multiple imputation. The SOFA score is the Sequential Organ Failure Assessment score. (2) Key variable screening: Univariate and multivariate logistic regression analysis was performed on the data described in (1) to obtain significant predictive variables for the death of patients with candidemia. LASSO regression analysis was used to screen the core variables among the predictive variables and determine the final variables for constructing the model. (3) Model construction and optimization: Based on the clinical significance of the variables and literature reports, the selected variables were scored, and the sub-items with poor prediction effects in the original SOFA score were replaced. The selected variables were combined in different ways to construct mSOFA models with different combinations. ROC analysis was performed using multiple machine learning algorithms, and the model with the maximum AUC was determined as the best prediction model; (4) Model validation: The model performance was comprehensively evaluated using internal and external validation sets. The model's predictive efficacy, calibration, and clinical practicality were systematically evaluated using a multi-dimensional validation method using the area under the ROC curve (AUC), calibration curve analysis, survival analysis, and clinical decision curve analysis (DCA).

2. The construction method according to claim 1, characterized in that The patient data came from multicenter databases, including MIMIC-III, MIMIC-IV and ICU-JN databases, covering data on patients with candidemia from June 2001 to January 2024.

3. The construction method according to claim 1, wherein: The SOFA score is the maximum SOFA score recorded within 24 hours after diagnosis, including respiratory system score, coagulation function score, liver function score, cardiovascular score, Glasgow Coma Scale and renal function score; the laboratory indicators include maximum white blood cell count, minimum hemoglobin, minimum platelet, maximum lactate, minimum PaO2 / FiO2 ratio, minimum bicarbonate, maximum partial thromboplastin time, maximum international normalized ratio, maximum prothrombin time, maximum blood urea nitrogen, maximum creatinine, maximum alanine aminotransferase, maximum aspartate aminotransferase, and maximum bilirubin.

4. The construction method according to claim 1, wherein: The screened variables were combined in different ways to construct different models. Four machine learning algorithms, including random forest, logistic regression, support vector machine, and extreme gradient boosting, were used to predict the mortality risk of patients with candidemia. The optimal prediction model was determined based on the obtained AUC.

5. The mortality risk prediction model for candidemia patients constructed according to any one of claims 1 to 4, characterized in that: The prediction model includes the following predictive factors: respiratory system SOFA score, coagulation function SOFA score, circulatory system SOFA score, maximum lactate value, minimum albumin value, and maximum blood urea nitrogen value.

6. The mortality risk prediction model for candidemia patients according to claim 5, characterized in that: The scoring criteria for each predictor are: Respiratory system (PaO2 / FiO2, mmHg): <100 = 4 points, <200 = 3 points, <300 = 2 points, <400 = 1 point, ≥400 = 0 points; Coagulation function (platelet 10 9 / L): <20 counted as 4 points, <50 counted as 3 points, <100 counted as 2 points, <150 counted as 1 point, ≥150 counted as 0 points; Circulatory system: dopamine (μg / Kg / min) > 15 or epinephrine > 0.1 or norepinephrine > 0.1 is scored as 4 points; dopamine (μg / Kg / min) > 5 or epinephrine ≤ 0.1 or norepinephrine ≤ 0.1 is scored as 3 points; dopamine (μg / Kg / min) ≤ 5 or any dose of dobutamine (μg / Kg / min) is scored as 2 points; mean arterial pressure (mmHg) < 70 is scored as 1 point; mean arterial pressure (mmHg) ≥ 70 is scored as 0 points; Maximum lactate score: >7.0mmol / L = 4 points, 4.0-7.0mmol / L = 3 points, 2.0-4.0mmol / L = 2 points, 1.2-2.0mmol / L = 1 point, ≤1.2mmol / L = 0 points; Minimum albumin value score: <2.0 g / dL is 4 points, 2.0-2.5 g / dL is 3 points, 2.5-3.0 g / dL is 2 points, 3.0-3.5 g / dL is 1 point, and ≥3.5 g / dL is 0 points; Maximum blood urea nitrogen score: >80 mg / dL is 4 points, 60-80 mg / dL is 3 points, 40-60 mg / dL is 2 points, 20-40 mg / dL is 1 point, and <20 mg / dL is 0 points.

7. The mortality risk prediction model for candidemia patients according to claim 5, characterized in that: The mortality risk prediction model for candidemia patients is mSOFA_3, mSOFA_3=0.5×Resp_SOFA+0.7×Coag_SOFA+0.6×Circ_SOFA+0.9×Lactate_score+0.8×BUN_score+0.6×Albumin_score; Patients were divided into high-risk and low-risk groups using the mSOFA_3 score with a cutoff value of 13 points; The Lactate is lactic acid, Albumin is albumin, BUN is urea nitrogen, Resp_SOFA is the respiratory system SOFA score, Coag_SOFA is the coagulation function SOFA score, and Circ_SOFA is the circulatory system SOFA score.

8. Use of the candidemia patient death risk prediction model according to any one of claims 5 to 7 in predicting the death risk of candidemia patients, specifically comprising the following steps: (1) Collect predictor data within 24 hours of admission; (2) calculating the scores of each indicator according to the scoring criteria of claim 6, and accumulating the scores according to the prediction model of claim 7 to obtain the total score of mSOFA_3; (3) When the total score is >13 points, the patient is judged as a high-risk patient and the predicted risk of death within 28 days is higher; when the total score is ≤13 points, the patient is judged as a low-risk patient and the predicted risk of death within 28 days is lower.

9. The use according to claim 8, characterized in that: The predictive factors include respiratory system SOFA score, coagulation function SOFA score, circulatory system SOFA score, maximum lactate value, minimum albumin value, and maximum blood urea nitrogen value.

10. The use according to claim 9, characterized in that The predictor data include the worst index value within 24 hours after the diagnosis of candidemia, and the worst index value is the maximum value or the minimum value.

Images