A retrospective cohort study was conducted at 2 surgery facilities (A and B) of the Department of Surgical Gastroenterology (NDGD Hospital), collecting data from July 1st, 2015, to December 31st, 2015. All medical records with the following patient characteristics were collected: (1) aged 18 or above; (2) admission date within the data collection timeframe; and (3) had a clean-contaminated surgery at the Department of Surgical Gastroenterology of the hospital. Records were excluded if the patients: (1) were using treatment antibiotics; (2) were immunocompromised (due to immunosuppressive medications or immunodeficiency disorders); or (3) had signs or symptoms of infections prior to the surgery (evaluated by physicians). We reported this study following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement (Supplementary Table S1).

As this was a relatively new practice in Vietnam, the hospital only implemented the ASPi at Facility A as a pilot program while maintaining the standard of care (SoC) at Facility B. Thus, the 2 cohorts in this study were patients treated at Facility B with SoC or at Facility A with SoC + ASPi. All the protocols, guidelines, medical equipment/device, medication supplies, medical support/care, and infection control measures were similar between the 2 facilities. SoC included all relevant medical support and care for patients undergoing gastrointestinal surgeries. Antimicrobial prophylaxis in the SoC cohort was primarily empirical therapy. The SoC + ASPi cohort received guideline-directed or expert-consensus prophylactic antibiotics. Details of the ASPi were presented in Supplementary Table S2.

The primary outcome was 30-day SSI, measured as cumulative incidence. We followed all patients up to 30 days after the surgery for SSI diagnoses that were given by either the surgeons (during hospitalization) or treating physicians (during outpatient visits or rehospitalization). Postoperative mortality within 30 days, unless ruled out by other causes, was also considered an SSI case. The secondary outcomes were length of stay (LoS) after surgery (in days, excluding in-hospital mortality), cost of antibiotics (for both prophylaxis and treatment), and cost of treatment (in USD). We only collected direct medical cost data based on the insurer’s perspective.

As the board of directors of our hospital required the full launch of ASPi at all departments in 2016, we had to conduct this study within a short timeframe and limited population size. Thus, we decided to collect all eligible medical records within the last 6 months of 2015. We then performed a power analysis (significance level of 5%) to determine if this sample size were adequate to detect any differences in the outcomes. With 227 patients in the SoC and 168 patients in the SoC + ASPi group, we estimated that the ASPi could reduce the 30-day SSI cumulative incidence from 19.8% (2014 data) to around 10.0%. This resulted in a power of 0.77, which was deemed comparable to the common value of 0.80, considering the short timeframe of this study.

All statistical calculations and analyses were performed using R software (version 3.2.3, R Foundation for Statistical Computing, Vienna, Austria). Single imputation was used to address missing data. We presented categorical variables as frequencies with percentages and continuous variables as mean with standard deviation (SD) or median with interquartile range (IQR).

We considered the following covariates as potential confounders: sex (female or male), age (years, <60 or ≥60), body mass index (kg/m²), chronic comorbidities (cardiovascular diseases, endocrine diseases, respiratory diseases, gastrointestinal diseases, or cancers), and risks of acquiring multidrug-resistant pathogens (yes or no). The difference in primary outcome was compared using beta-binomial regression, while linear regression was used to analyze the secondary outcomes. Due to the non-normality of LoS after surgery and cost of antibiotics and treatment, we applied bootstrapping with 10,000 replications to estimate the effect differences in these cases. We used Holm method to control for multiple comparisons of outcomes. All statistical tests were performed with a family-wise error rate of 5% and reported with 95% confidence interval (95% CI).

We also built a predictive model to identify patients at high risk for post-discharge SSIs. The variables for modeling include age, gender, days in hospital before/after the surgery, blood transfusion, number and type of comorbidities, serum level of aspartate aminotransferase/alanine aminotransferase/urea/creatinine/glucose, surgery site, surgery type, duration of surgery, American Society of Anesthesiologists (ASA) score, presence of cancer, antimicrobial agent, rational choice/dosage/timing of administration of antibiotics. Four types of models were included and compared, of which the most accurate one would be used for further predictions and estimations. The performances of the models were assessed using: (1) F1 score; (2) accuracy; (3) sensitivity; (4) specificity; (5) positive predictive value (PPV); (6) negative predictive value (NPV) (13). For this predictive modeling, we randomly split the dataset by a ratio of 8:2 into training and testing sets, with the former for model building and the latter for cross-validation. For the data pre-processing, each variable was standardized to a mean of 0 and a variance of 1. Principle component analysis was then used to reduce the dimensionality of the training and testing datasets, of which low-variance (confounding) dimensions were excluded, leaving a total of 95% sum of variances for the remained dimensions. After that, we built and trained our models based on the following methods: (1) logistic regression; (2) random forest; (3) support vector machines (SVM); (4) kernel Fisher discriminant analysis (KFDA) (14). For random forest-based model, we adjusted the cut-off threshold for the labels positive:negative (P:N) by a ratio of 2:8 to prioritize prediction for the P label. With the SVM-based model, we applied the radial basis function kernel, and to avoid overfitting, we set the cost parameter to 10⁻⁶. Due to the highly unbalanced distribution of P:N labels, instead of using the original F1_1,1 score, we used the customized F1_1,4 to improve the P label prediction based on the following equation:

This study was approved by the Institutional Review Board (IRB) of NDGD Hospital, under approval number 108/CN-HDDD. The IRB did not require informed consent from patients to conduct this study, as we only used the medical records to collect retrospective data without revealing patient identity.

There were no missing data in this study sample (Figure 1). Overall, 395 patients (48.1% being female) had an average age of 49.4 ± 15.8 years. No patients had an ASA score higher than 3, showing a relatively low risk of SSI in this study setting (15). The majority of patients did not have or have only 1 comorbidity, primarily hypertension (72 out of 395) and diabetes (25 out of 395). Most patients underwent open surgeries (67.1%) in the large intestine (50.4%). Surgery duration varied widely from 15 to 430 min, with a median of 60 min (IQR: 40–100). β-lactam antibiotics were given in most gastrointestinal surgeries, with the predominance of amoxicillin-clavulanic acid in the SoC group and cefazolin in the SoC + ASPi group. Metronidazole was rarely prescribed and was only combined with other antibiotics, including ampicillin/sulbactam, ceftazidime, and ceftriaxone. Among 395 patients, only 3 self-declared to have an antibiotic allergy (2 with penicillin and 1 with cefuroxime). Further details between the SoC and SoC + ASPi groups were reported in Table 1.

Table 2 reported the effect estimates for the outcomes of this study. No in-hospital mortality was recorded. We found a significantly lower risk of 30-day SSI (primary outcome) in patients receiving SoC + ASPi (11.9%) than those with SoC (20.7%), with an effect difference of −8.8, 95% CI: −16.0 to −1.6, adjusted p = 0.042. We also observed similar results for the secondary outcomes. LoS after surgery was 24.4% shorter following the implementation of ASPi (mean difference of −1.1 days, 95% CI: −1.8 to −0.4, adjusted p = 0.004). Healthcare costs were also lower in the SoC + ASPi group (mean difference in cost of antibiotics: −37.3 USD, 95% CI: −59.8 to −14.8, adjusted p = 0.012; mean difference in cost of treatment: −191.1 USD, 95% CI: −348.4 to −33.8, adjusted p = 0.042).

Table 3 summarizes the internal and cross-validated results of SSI predictive modeling by assessing four types of models. Random forest model underwent severe overfitting, as implicated by the significant declines in cross-validity. The KFDA model seemed to outweigh others in terms of F1 score, accuracy, PPV, and NPV. These results suggest the most appropriate model for implementation could be based on the KFDA method.

Considering the sensitivity (80.6%) and PPV (44.9%) of the KFDA-based model, we estimated the total reduced cost through the implementation of ASPi and SSI predictive modeling. Patients with a prediction of high risk for SSI would be provided post-discharge prophylaxis (7-day surgical wound care kit, which included sterile gauze pads, antiseptic wipes and swabs, antibiotic ointment, sterile saline solution, medical tape, bandages, surgical gloves, and specialized dressings). In our setting, the expense for this preventive measure was 10.5 USD per patient, while the average cost of SSI treatment was approximately 550 USD per patient. Thus, the implementation of ASPi and SSI predictive modeling could reduce healthcare costs by 398120.7 USD per 1,000 cases of SSI (Supplementary Figure S1).