This retrospective study utilized the Medical Information Intensive Care (MIMIC-IV) database (v3.1), an iterative version following MIMIC-III. The database complies with HIPAA security regulations and ensures anonymization of the data. MIMIC-IV contains a large amount of clinical data from 70,000 adult intensive care unit (ICU) patients at the Boston Diabetes Research Institute (BIDMC) between 2008 and 2019 (12).

All patient data within the database is anonymized, obviating the need for informed consent. In adherence to the ethical standards articulated in the 1964 Declaration of Helsinki and its subsequent amendments, the study was conducted. Access to the database was secured following the completion of the National Institutes of Health Web-based training course and the Protecting Human Research Participants examination (No. 43258214).

Patients fulfilling specific criteria were screened through the MIMIC-IV database (version 3.1) for this study. We identified individuals in the database meeting the following criteria:

(1) patients were diagnosed with CAUTI according to the International Classification of Diseases, as indicated by ICD-9 codes, or ICD-10 codes;

(2) only the initial ICU admission date was considered for patients with multiple ICU admissions;

(3) patients were aged 18 years or older.

Patients who had more than 30% missing values were excluded (13). Ultimately, 545 patients were enrolled in this study (Figure 1).

The predicted outcome was the probability of death during a stay in the intensive care unit (ICU). Baseline demographic variables, comorbidities, vital signs, length of hospital stay, severity scores, and laboratory data were extracted from the MIMIC database based on previous studies as well as expert input and implemented using SQL (Structured Query Language) programming. With the exception of length of stay, vital signs were collected within the first 24 h after each ICU admission, whereas other variables were measured at admission. In addition, in order to avoid overfitting, the least absolute shrinkage and selection operator (LASSO) method was used for variable selection and screening, and the LASSO regression was used to select the optimal regularization parameter λ by 10-fold cross-validation (14).

Missing data often occur in the MIMIC-IV database. However, if these missing values are ignored during the analysis, the results may be biased. Therefore, we used chained equation multiple imputation (MICE) to deal with missing values, and the number of imputation was set to 5 times to deal with missing data (15). The proportion of missing values in each of the selected variables was less than 30%.

The prediction model is interpreted by SHAPmethod, which is a comprehensive method that can accurately evaluate the contribution and influence of each feature on the final prediction result. SHAP analysis is implemented based on the SHAP 0.44.0 library of Python 3.8 (16). SHAP values indicate the extent to which each predictor variable affects the target variable, either positively or negatively. Furthermore, each data point can be understood by its specific set of SHAP values.

DecisionLinnc1.0 software is used for data analysis, DecisionLinnc1.0 is a platform that integrates multiple programming language environments and realizes data processing, data analysis and machine learning through a visual interface (17). Categorical variables were presented as total amounts and percentages, and the chi-square test or Fisher exact probability method was used to compare the differences between different groups. Continuous variables were expressed as medians and interquartile ranges (IQR), and comparisons between the two groups were performed with the use of the Wilcoxon rank-sum test.

Four machine learning models -XGBoost, DecisionTree (DT), Logistic Regression (LR), and Random Forest (RF) -were used to construct the prediction model. The predictive performance of each model was evaluated by the area under the receiver operating characteristic curve. In addition, we calculated accuracy, precision, and F1 scores. In addition, in order to evaluate the practicality of the model in decision making by quantifying the net benefits under different threshold probabilities, decision curve analysis (DCA) was performed (18).

In this study, 545 adult patients diagnosed with CAUTI were included out of a total of 1344 patients with CAUTI in the MIMIC-IV database. The depiction of the patient screening process can be observed in Figure 1.

Table 1 presents the baseline characteristics of 545 patients who fulfilled the inclusion criteria, categorized into the ICU survival group and non-survival group. The mortality rate of ICU patients diagnosed with CAUTI was 7.89% (43/545). Among these patients, there were 261 females (47.89%) and 284 males (52.11%), with a median age of 74 (21–99) years, and the age difference was not significant (P = 0.803). In terms of length of hospital stay, the median length of hospital stay was 11.94 days for survivors and 15.86 days for non-survivors (P = 0.117), and there was no significant difference in length of ICU stay (P = 0.065). Severity of illness score showed that SOFA score, APSIII score, APSII score, OASIS score and shock index of non-survivors were significantly higher than those of survivors (P < 0.05). The duration of mechanical ventilation of non-survivors was significantly longer (P = 0.004), and vital signs such as heart rate, respiratory rate, and blood oxygen saturation showed significant differences. Among the laboratory indicators, lactic acid, PH value, international normalized ratio of prothrombin time and creatinine were significantly poor in non-survivors (P < 0.05). In terms of complications, the incidence of acute renal failure was significantly higher in non-survivors, as was the incidence of malignancies. In terms of drug use, the use of sedatives, analgesics and vasoactive drugs in non-survivors was significantly higher than that in survivors (P < 0.05). The LASSO regularization method was used to select 13 potential predictors from the training dataset, and these factors were used for model development.

The dataset was divided in a random fashion into two parts: 70% of the data was used to train the model, while 30% was used to validate the model. In the training dataset, we built four models: XGBoost, Logistic Regression (LR), Random Forest (RF), and Decision Tree (DT). The AUC values obtained from the test dataset are shown in Figure 2 and Table 2, respectively. Among these models, LR showed superior predictive performance with an AUC of 0.871, while DT had the lowest generalization ability with an AUC of 0.542. The net benefit of the best-performing model was compared with an alternative approach to clinical Decision making using Decision Curve Analysis (DCA) on the test dataset.

We evaluated the overall payoffs at different probability thresholds. The assumptions in Figure 3, represented by the black line, assume that all patients received the intervention. On the other hand, the dashed line represents the case where no patient received any intervention. Given the diverse nature of the study population, developing a treatment strategy based on any of the four machine-learning models would be preferable to treating all or none of the patients by default.

The SHAP algorithm was used to determine the importance of each predictor variable in the prediction results of the LR model. The variable importance map presents a list of variables ranked from highest to lowest according to their level of importance.

The use of vasoactive agents was considered to have the highest predictive value of all prediction periods, followed by the shock index, coexisting malignancies and APSIII score (Figure 4).

In addition, SHAP values were used to identify predictor variables that had a significant effect on mortality risk and to determine their positive or negative association with the target outcome. As shown in Figure 5, the horizontal position indicates whether the effect of the value is associated with an increase or decrease in the predicted value, while the color indicates the high or low state of the variable in a particular observation.

Figure 6 shows the heat force plots for patients who did not survive and survived. The SHAP values provide insights into the predictive factors of individual patients and quantify the impact of each factor on mortality prediction. The numbers highlighted in bold represent the probabilistic predictions (f(x)), while the base values indicate the predictions made by the model without any input. The log odds ratio of each observation is represented by the function f(x). The left side displays red features that are associated with an elevated risk of mortality, while the blue features represent factors linked to a reduced risk of mortality. The magnitude of the effect on the prediction can be easily visualized by observing the length of the arrows.