This retrospective study examined patients with surgical wound infections after cardiothoracic surgery who were admitted to the Cardiologic Reference Center Hospital (Instituto do Coração ‒ InCor), University of São Paulo, Brazil, between January 2006 and January 2023. Patient selection was conducted prospectively as part of routine surveillance activities by the Hospital Infection Control Department.

This department monitors wound infections in hospitalized patients and actively investigates outpatient infections. Infections were categorized based on definitions established by the Centers for Disease Control and Prevention (CDC)11 and classified as superficial, deep, or organ space (mediastinitis), with osteomyelitis not differentiated in this cohort. The construction of the scoring system adhered to the “Guide for presenting clinical prediction models for use in clinical settings”, and the results were reported following the “Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement”.12

The study included patients diagnosed with surgical wound infections (as per Centers for Disease Control and Prevention guidelines)11 following various types of cardiothoracic surgeries, such as Coronary Artery Bypass Graft Surgery, Coronary Artery Bypass Graft Surgery combined with valve procedures, valve surgery, aorta/vascular surgery, heart transplant, congenital heart surgery, and general cardiac surgery. Patients under 18-years old were excluded from the study.

The primary outcome was the occurrence of 30-day death from all causes after diagnosis of wound infections after cardiothoracic surgery. All participants were followed for at least 30-days following their diagnosis of wound infection.

The predictor variables studied were collected by reviewing medical records on the day of infection diagnosis, such as age, body mass index, presence of fever, hemoglobin, leukocyte, lymphocyte, platelet, C-reactive protein, and creatinine levels. Chest tomography reports performed up to 7-days after the diagnosis of infection were studied, and the following characteristics were evaluated: bilateral pleural effusion, mediastinal collection, bone resorption, bone misalignment, and bone diastase. The primary surgery type and the time elapsed between the diagnosis of infection and the surgical cleaning approach were also collected.

Two models were used in this study. Establishing binary logistic regression allowed researchers to study the marginal impacts of the predictor variables on the phenomenon studied (i.e., ORs) and propose the SWICS-30 score. After this, a Deep Learning (DL) Artificial Neural Network (ANN) was used to validate the results of the SWICS-30 score further.

We stratified the descriptive statistics based on in-hospital mortality. We presented them as means and standard deviations or medians and 25th/75th percentiles for continuous variables, as appropriate for their distribution, and as counts and percentages for categorical variables. The analysis involved t-tests or Wilcoxon rank-sum tests for continuous variables, as appropriate, and the Chi-Square test for categorical variables.

Predictor variables were selected after thorough review of existing literature of surgical wound infections after cardiothoracic surgery. For multivariable analysis, the selected variables included age ≥ 65-years, surgical group such as valve group, transplant group, coronary-valve group, time between surgery and diagnostic of infection, leukocytes > 13.000 mm3, lymphocytes 〈 1.000 mm3, platelets < 150.000 mm3, creatinine 〉 1.5 mg/dL.

These variables were integrated into an initial model, and a multivariable logistic regression was performed, with a purposeful backward selection of variables to create the final model based on their contribution to the model explainability by likelihood ratio tests and information criteria. Our objective was to develop the most parsimonious model. Afterwards, we developed a scoring system, assigning points based on logistic regression coefficients for each variable. Finally, patients were categorized into three distinct risk levels of mortality by analyzing the numerical and graphical distribution of the risk of this outcome and testing the optimal cutoff point for group differentiation based on the clinical decision-making required at the bedside. The score was given using the logarithmic coefficients from the logistic regression following the Guide for presenting clinical prediction models for use in clinical settings.

Apparent validation was conducted by analyzing the Area Under the Receiver Operating Characteristic (AUROC) curve and a calibration belt. Bootstrap internal validation was performed with 500 bootstrap resamples for AUROC, calibration-in-the-large, calibration slope and a calibration plot.

All tests were 2-sided, and a p-value < 0.05 was considered statistically significant. No adjustment for multiplicity was performed. The analysis was performed using StataSE® software (StataCorp, College Station, Texas), version 18.0.

To enhance the robustness of the study, a DL model was suggested that made use of the SWICS-30 scores recommended by the logistic estimation discussed earlier. The established DL model was a multi-layer feed-forward artificial neural network, estimated using the h2o package, version 3.44.0.3, of the R computer language core team.

The DL model was trained with 70 % of the sample using a cross-validation technique with 10k-folds and misclassification as a stopping metric. The algorithm was trained using stochastic gradient descent and backpropagation criteria.13

For the input layer of the deep learning estimation, the same predictor variables utilized in the binary logistic model were taken into account. However, the variables age, time between surgery and diagnosis of infection, leukocytes, lymphocytes, platelets, and creatinine were considered in their original metric forms. The hyperparameters of the DL model were established using grid simulations following a random discrete strategy.14,15

The aforementioned grid simulations were also used to establish an ideal number of three hidden layers for the ANN, with 600 neurons each, and helped to choose the activation functions for each layer of the model. The rectifier function was used to calculate the interaction between neurons in the hidden layers.

For both the hidden and output layers, we opted for lasso regression (L1) and ridge regression (L2) regularizations16,17 with weights of 0.00001 and 0.001, respectively. The aim was to add a penalty term to the selected loss function,18 namely, cross entropy. All the choices made helped minimize the log loss metric, both for the training and the validation samples, to mitigate the possibility of overfitting the model.19

Finally, given that, the proposed SWICS-30 scores are polytomous categorical variables (i.e., Low, middle and High), a SoftMax activation function was considered for the output layer.

This study was approved by the Ethics and Research Committee of the Instituto do Coração (InCor), University of São Paulo, Brazil. Heart Institute of the Faculty of Medicine of the University of São Paulo and registered under SDC 3695|11|113 and CAAE 31,593,814.8.0000.0068. The requirement for informed consent was waived due to the retrospective nature of the study.

In the present study, there was a female predominance of 55 %. The mean age of participants was 60 years (interquartile range: 18‒89). The baseline characteristics of the participants, including demographics, risk factors, signs and symptoms, laboratory tests, microorganisms, and surgical group, are described in Table 1.

The 30-days overall mortality rate was 8.3 % (143/1713). The comorbidity highlighted was diabetes mellitus in 517 (30 %) cases. Coronary surgery was the largest surgical group with 919 patients (53 %), followed by the valve group with 389 patients (23 %) and the aorta/vascular group with 137 patients (7.9 %).

The classification of surgical wound infections was as follows: 660 patients (38.5 %) had superficial infections, 770 patients (44.4 %) had deep infections, and 179 cases (10.4 %) had mediastinitis.

Staphylococcus spp. coagulase-negative and Staphylococcus aureus were the most commonly observed microorganisms, with 455 cases (26 %) and 343 cases (20 %) respectively. Gram-Negative Bacilli (BGN) were also frequently identified, accounting for 465 cases (27 %), while fungi were detected in 133 cases (7.7 %). Enterococcus spp. were identified in 77 cases (4.5 %), and Streptococcus spp. were found in 13 cases (0.7 %). It is important to highlight that 14 % of cases had negative culture results or were not collected due to the presence of superficial infections with poor secretion. Surgical cleaning of the wound was indicated in 934 instances (55 %) and occurred on average 21 days (± 35-days) after diagnosis of the infection. Among the 934 individuals who underwent surgical intervention, 696 (75 %) of them utilized negative pressure wound therapy.

Results of the final logistic regression model are described in Table 2. The variables and their respective scores are detailed in Table 3.

The risk groups were established as follows: low risk (1 to 5-points), moderate risk (6 to 8-points) and high risk (≥ 8-points) of death, as shown in Table 4. The incidence of hospital death was 2.7 %, 14.2 % and 47.1 % for each risk group, respectively.

The SWICS-30 score presented an area under the ROC curve of 0.82 (95 % CI 0.78‒0.86), as shown in Fig. 2A. Model calibration, evaluating the alignment between the predicted probabilities and the actual observed results, was considered adequate. The calibration belt in Fig. 2B visually illustrates this agreement.

Fig. 2.

Apparent validation of the logistic regression model. (A) An Area under the receiving operator characteristic curve. (B) Calibration belt of the model.

(0.23MB).

The bootstrap internal validation with 500 replicates yielded an AUROC of 0.804. Calibration-in-the-large was 0.036 and calibration slope was 0.921. The calibration plot after bootstrapping internal validation is presented in Fig. 3.

Fig. 3.

Calibration plot after bootstrapping internal validation.

(0.24MB).

The DL model was used to validate the SWICS-30 score, considering the Low, Middle and High labels (Table 4). To this end, patients who had missing values among the predictor variables, making it impossible to calculate their SWICS-30 score, were disregarded. It is important to mention that, a priori, all the variables described in Table 1 were considered in the modeling, but dichotomous predictor variables with few events (i.e., <30) were also disregarded. The study variables originally presented as quantitative variables (Table 1) were considered in their observed format, that is, as continuous variables. Therefore, of the 1713 individuals in the initial sample; 1554 were considered for ANN modeling. The validation results are presented in Table 5.

If we consider the sum of the elements on the main diagonal of Table 5 and divide them by the sample considered by the ANN, we see that the accuracy is approximately 90 %.

However, unlike a logistic regression model, ANNs cannot be summarized as a single mathematical equation, making it impossible to calculate the ORs of the predictor variables for the phenomenon studied. Nevertheless, there is the possibility of an approximation given by the so-called Partial Dependence Plots (PDP).20

The PDPs propose visualizing the relationship between a given set of explanatory variables and the dependent variable, considering their average impact on the phenomenon studied.21 It must be said that the PDPs assume independence between a given predictor variable and the rest of them in relation to the phenomenon. Fig. 4 shows the PDP for each predictor variable in this study.

Fig. 4.

Partial dependent plots of the ANN model.

(1.07MB).

Fig. 4 shows the behavior of the variables present in the base in relation to the SWICS-30 score, assuming Low, Middle and High categories, as shown in Table 4.

When analyzing the dichotomous explanatory variables, the results presented in Fig. 4 suggest that patients who underwent coronary-valve surgery, valve surgery, heart transplant, or who presented bilateral pleural effusion, are more likely to be categorized by the SWICS-30 score as middle or high-risk patients. In fact, with the exception of the bilateral pleural effusion variable, the presence of the aforementioned variables leads to a sharp drop in the probability of a patient being considered low risk in the SWICS-30 score.

The outcomes of the study's metric variables, as depicted in Fig. 4, suggest that an increase in the time between surgery and the diagnosis of wound infection, as well as a rise in the age variable, leads to a higher likelihood of a patient being categorized as middle or high-risk in the SWICS-30 score.

The researchers emphasize once again that the interpretations of the PDP are independent of the joint behavior of the other predictors in the phenomenon studied.

We recognize that this study has certain limitations. Firstly, its retrospective design raises the possibility of information bias. Furthermore, the fact that the present study was conducted at a single tertiary cardiac hospital that serves as a referral center for severe and complex cases suggests a relatively higher risk of death. These features might lead to variations in score performance when applied in hospitals with different clinical and microbiological profiles. Although the internal validation of the SWICS-30 score was satisfactory, the absence of a validation cohort warrants caution when interpreting our results. Further validation in other samples is necessary to ensure the reliability of our findings.

Julio Alejandro Cedeno: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Projectadministration, Resources, Writing – original draft, Writing – review & editing. Tania Mara Varejão Strabelli: Fundingacquisition, Writing – review & editing. Bruno Adler Maccagnan Pinheiro Besen: Software, Supervision, Validation,Visualization, Data curation. Rafael de Freitas Souza: Software, Supervision, Validation, Visualization, Data curation. Denise Blini Sierra: Data curation. Leticia Rodrigues Goulart de Souza: Data curation. Samuel Terra Gallafrio: Study design, Data curation. Cely Saad Abboud: Writing – review & editing. Diego Feriani: Writing – review & editing. Rinaldo Focaccia Siciliano: Funding acquisition, Conceptualization, Data curation, Formal analysis,Investigation, Methodology, Project administration, Resources, Writing – original draft, Writing – review & editing.