Risk Stratification for Superficial Surgical Site Infection after Emergency Trauma Laparotomy

Abstract

Background: Superficial surgical site infections (S-SSIs) are common after trauma laparotomy, leading to morbidity, increased costs, and prolonged length of stay (LOS). Opportunities to mitigate S-SSI risks are limited to the intra-operative and post-operative periods. Accurate S-SSI risk stratification is paramount at the time of operation to inform immediate management. We aimed to develop a risk calculator to aid in surgical decision-making at the time of emergency laparotomy.

Methods: A retrospective cohort study of patients requiring emergency trauma laparotomy between 2011 and 2017 at a single, level 1 trauma center was performed. Operative factors, skin management strategy, and outcomes were determined by chart review. Bayesian multilevel logistic regression was utilized to create a risk calculator with variables available upon closure of the laparotomy. Models were validated on a 30% test cohort and discrimination reported as an area under the receiver operating characteristics curve (AUROC).

Results: Of 1,322 patients, the majority were male (77%) with median age of 33 years, injured by blunt mechanism (54%), and median injury severity score of 19. Eighty-eight (7%) patients developed an S-SSI. Patients who developed S-SSI had higher final lactate, blood loss, transfusion requirements, and wound classification. Patients with S-SSI more frequently had mesenteric or large bowel injury than those without S-SSI. Superficial SSI was associated with increased complications and prolonged length of stay (LOS). The S-SSI predictive model demonstrated moderate discrimination with an AUROC of 0.69 (95% confidence interval [CI], 0.56–0.81). Parameters contributing the most to the model were damage control laparotomy, full-thickness large bowel injury, and large bowel resection.

Conclusion: A predictive model for S-SSI was built using factors available to the surgeon upon index emergency trauma laparotomy closure. This calculator may be used to standardize intra- and post-operative care and to identify high-risk patients in whom to test novel preventative strategies and improve overall outcomes for patients requiring emergency trauma laparotomy.

Superficial surgical site infections (S-SSIs) are common after emergency trauma laparotomy, with reported rates as high as 25% [1]. Superficial SSIs can lead to fascial dehiscence, ventral hernia formation, and sepsis and contribute more than $1.6 billion annually in additional healthcare costs [2,3]. Multiple evidence-based guidelines exist regarding protocols to reduce the risk of S-SSIs [4–6]. However, they pertain primarily to elective procedures and target pre-existing risk factors, such as a smoking history, glucose control, steroid use, and obesity [7–9]. Unfortunately, modifications of these pre-existing risk factors are often not possible in emergency trauma laparotomy patients, and attention is instead focused on intra- and post-operative measures to reduce S-SSI risk.

Traditionally, patients deemed to be at high risk of developing S-SSI are managed with an open-skin strategy with or without delayed primary closure in order to mitigate these risks [1,10]. More recent studies have challenged this dogma, suggesting that primary closure is safe [11]. Additionally, adjuncts such as negative pressure wound therapy over closed wounds may further reduce the risk of S-SSI after emergency laparotomy [12–14]. Given the increased emphasis on patient satisfaction, other outcomes such as cosmesis [15] and resource utilization [16] are also being considered when comparing skin management strategies after laparotomy.

The surgeon must balance the risks and benefits in choosing between open versus closed skin management of the laparotomy wound. Accurate risk assessment at the conclusion of an emergency trauma laparotomy is necessary to inform such decision-making. However, current methods for risk assessment after abdominal trauma are limited in their ability to predict S-SSI using parameters available at the end of the operation. The aim of this study was to develop a risk calculator to aid in surgical decision-making at the time of emergency laparotomy.

Methods

A retrospective cohort study was conducted of adult (≥16 years) trauma patients requiring emergency trauma laparotomy from 2011 to 2017 at a high-volume, level 1 trauma center. Demographic, injury, and outcome data were abstracted from a prospectively maintained database and supplemented by medical record review. The surgeons have consensus regarding absolute and relative indications for damage control laparotomy and for standardized post-operative care [17,18]. All patients undergoing damage control laparotomy receive hypertonic saline until first take-back at 24 to 48 hours. Patients with packs in place receive cefoxitin. For patients undergoing definitive laparotomy, post-operative antibiotic agents are only administered if there is evidence of pre-existing infection such as an abscess, and antibiotic agents are limited to four days after source control [19]. The primary outcome was an S-SSI that was determined by agreement of two independent reviewers of clinical documentation (G.H. and S.W.). The U.S. Centers for Disease Control and Prevention definition of S-SSI was applied: an infection that occurs within 30 days from the index surgery, involves only skin and subcutaneous tissue of the incision, and has either purulent drainage from the incision, an organism identified from an incision culture, signs or symptoms of infection followed by incision opening, or if otherwise diagnosed by a physician [20]. Peri-operative antibiotic agent administration related to trauma laparotomy was limited to 24 hours post-operatively and was standardized across all patients. Skin closure was not standardized and was performed according to physician preference. Open skin entailed use of gauze packing or negative pressure wound therapy and closed skin entailed closured with staples (with or without wicks) or sutures. The earliest case of S-SSI occurred two days after the index laparotomy. Therefore, deaths prior to post-operative day two were excluded to avoid death as a competing outcome. Additionally, patients with missing data were excluded from the analyses. The Institutional Review Board at the McGovern Medical School approved this study.

Statistical analysis

Median values with interquartile ranges were used to describe continuous data and discrete data were reported as frequency and percentage. Parameters of interest were compared using the χ² test or a Wilcoxon rank-sum test for categorical and continuous parameters, respectively.

Bayesian multilevel logistic regression using regularization was conducted to develop an S-SSI predictive model based on a 70% training sample. Evaluation of model fit using area under the receiver operating curve (AUROC) was then carried out on a 30% test sample. Clinically sound covariables available to a surgeon upon the conclusion of the index laparotomy were selected a priori because of a known or suspected relation with S-SSI development and included: patient age, body mass index (BMI), year of surgery, mechanism of injury, duration of laparotomy, final intra-operative patient temperature, intra-operative packed red blood cell transfusion requirement, damage control laparotomy, full-thickness small bowel injury, small bowel resection, full-thickness large bowel injury, large bowel resection, skin closure, and wound classification.

The selected approach for developing the predictive model has been described previously [21]. Regularization is the machine learning technique that optimizes the bias-variance trade-off, rendering models with likely superior out-of-sample performance because of avoidance of overfitting in the training dataset [22,23]. Regularization is particularly in need in the context of sparsity. In our case, we have 62 events in the training set to develop a predictive model with 14 continuous and binary predictors. From a Bayesian perspective, assigning regularized prior distributions to model parameters allows shrinkage towards zero, limiting the likelihood of selecting predictors that are not truly associated with outcomes but owing to idiosyncrasies of the training sample [22–27]. We adopted Bayesian approaches so that we could calculate predictive probabilities along with their credible intervals. Bayesian approaches also allow the clustering within primary surgeons to be incorporated in the computation of predictive probabilities and update the prediction model as we have more data.

We checked the data for zero- and near zero-variance parameters and highly correlated parameters. We then standardized continuous parameters before splitting the dataset into a training and a test set. We performed Bayesian multilevel logistic regression with random intercepts to account for clustering by primary surgeons. We applied two shrinkage priors namely Horseshoe and Laplace distributions on the population-level coefficients; in addition, we implemented a vague neutral normal prior with mean 0 and standard deviation 1,000 on the coefficients for comparison. A normal prior with mean 0 and standard deviation 100 was set for the intercept and a half Student-t prior with three degrees of freedom and standard deviation of 10 for the standard deviation of random effects. We examined the regularized models with different degrees of freedom one, three, five, and seven for different levels of shrinkage. We determined the best model based on predictive accuracy estimated by leave-one-out cross-validation as well as the widely applicable information criterion. The best model was then evaluated in terms of discriminative power on both training and test sets using the AUROC [26]. All data analyses were run using R version 3.53 (R Core Team. 2013. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria). Data preprocessing was done using the caret package [28]. Bayesian multilevel logistic regression models with and without regularization, and cross-validation was performed using the brms package [29,30]. The ROC curves were plotted and the AUROC was computed using the pROC package [31].

Results

Throughout the study duration, 1,492 emergent trauma laparotomies were performed. Of those, 166 were excluded for death prior to the second postoperative day, and four were excluded for missing outcome data (Fig. 1). The final cohort contained 1,322 patients. Included patients were primarily male (78%), injured by blunt mechanism (53%), and were a median age of 33 years (interquartile range [IQR], 24–46; Table 1) Patients were severely injured with a median injury severity score of 18 (IQR, 10–29). Of 1,322 patients, 88 (7%) patients developed an S-SSI. Those who developed an S-SSI were more likely to have a higher body mass index (BMI), history of diabetes, and higher injury severity score. Other demographic data were similar between groups.

FIG. 1.

Flow chart of patients included in analysis.

Table 1.

Demographic and Arrival Data

All patients (n = 1,322)	No S-SSI (n = 1,234)	S-SSI (n = 88)	p
Age, y	32 (23–46)	37 (27–50)	0.05
Year			0.08
2011	166 (13%)	14 (16%)
2012	170 (14%)	18 (20%)
2013	167 (14%)	16 (18%)
2014	185 (15%)	12 (14%)
2015	180 (15%)	15 (17%)
2016	209 (17%)	9 (10%)
2017	157 (13%)	4 (5%)
Male gender	967 (78%)	65 (74%)	0.39
BMI	26 (23–30)	28 (24–33)	0.006
Race/ethnicity			0.42
White	466 (38%)	32 (36%)
Black	314 (25%)	17 (19%)
Hispanic	408 (33%)	34 (39%)
Asian	46 (4%)	5 (6%)
History of diabetes	76 (6%)	10 (11%)	0.06
History of smoking	390 (32%)	29 (33%)	0.89
Blunt mechanism of injury	653 (53%)	53 (60%)	0.22
Injury severity score	18 (10–29)	22 (14–34)	0.03
Arrival systolic blood pressure, mm Hg	115 (95–132)	112 (90–127)	0.13
Arrival heart rate	98 (92–115)	100 (84–117)	0.24
Arrival Glasgow Coma Scale	15 (13–15)	15 (7–15)	0.10
Arrival hemoglobin	13.2 (11.9–14.4)	12.8 (11.6–14.2)	0.16
ED crystalloid, mL	0 (0–0)	0 (0–0)	1
ED PRBC, units	0 (0–1)	0 (0–2)	0.14
ED FFP, units	0 (0–1)	0 (0–2)	0.13

Continuous data presented as: median (interquartile range [IQR]).

S-SSI = superficial surgical site infection; BMI = body mass index; ED = emergency department; PRBC = packed red blood cell; FFP = fresh frozen plasma.

During index laparotomy, patients who developed an S-SSI had higher final lactate values, transfusion requirements, blood loss, and wound classification scores (Table 2). More patients who suffered an S-SSI had an initial damage-control laparotomy. Patients who developed an S-SSI more frequently had mesenteric or large bowel injury than those without S-SSI (Table 3). Other injury patterns were similar between groups. Notably, 98 patients underwent laparotomy without injuries found. Two of these patients developed an S-SSI. Among patients with an initial definitive laparotomy (76%), defined as fascial closure at the conclusion of the index operation, S-SSI rates were similar between those who did and did not have their skin closed (6% vs. 4%, respectively, p = 0.29).

Table 2.

Index Laparotomy Intra-Operative Data

All patients (n = 1,322)	No S-SSI (n = 1,234)	S-SSI (n = 88)	p
Initial operative measurements
Temperature, °F	96.8 (95.7–97.7)	97.0 (96.0–98.1)	0.09
SBP, mm Hg	120 (102–140)	127 (100–142)	0.43
HR, bpm	97 (84–110)	100 (90–113)	0.22
pH	7.32 (7.25–7.37)	7.30 (7.24–7.36)	0.45
Base excess	–4 (−7 to −1)	−4 (−7 to −1)	0.64
Lactic acid	2.7 (1.6–4.0)	3.0 (1.8–4.6)	0.20
Final operative measurements
Temperature, °F	97.0 (96.1–98.2)	97.0 (96.0–97.9)	0.73
SBP, mm Hg	128 (115–140)	126 (115–140)	0.74
HR, bpm	90 (80–100)	92 (84–100)	0.26
pH	7.34 (7.30–7.38)	7.34 (7.30–7.39)	0.68
Base excess	−4 (−6 to −2)	−4 (−6 to −1)	0.90
Lactic acid	2.9 (1.8–4.2)	3.4 (2–5)	0.02
OR crystalloids, mL	1,400 (1,000–2,000)	1,500 (1,000–2,000)	0.84
OR colloids, mL	500 (0–1,000)	500 (0–1,000)	0.07
OR PRBC, units	0 (0–4)	2 (0–4.25)	0.05
OR FFP, units	0 (0–3)	1 (0–4.25)	0.08
OR platelets, units	0 (0–1)	0 (0–6)	0.17
Estimated blood loss, mL	300 (100–1,000)	400 (150–1,325)	0.05
Time from ED to OR, min	71 (42–142)	66 (45–109)	<0.001
Index laparotomy OR time, min	127 (89–184)	130 (100–200)	0.28
Initial damage control laparotomy	269 (22%)	50 (57%)	<0.001
Skin closed at initial laparotomy	1,072 (87%)	69 (78%)	0.04
Wound class			<0.001
1	248 (20%)	7 (8%)
2	335 (27%)	41 (47%)
3	617 (50%)	36 (41%)
4	34 (3%)	4 (5%)

Continuous data presented as: median (interquartile range [IQR]).

S-SSI = superficial surgical site infection; SBP = systolic blood pressure; HR = heart rate; bpm = beats per minute; OR = operating room; PRBC = packed red blood cell; FFP = fresh frozen plasma; ED = emergency department.

Table 3.

Intra-Operative Injury Findings

All patients (n = 1,322)	No S-SSI (n = 1,234)	S-SSI (n = 88)	p
No injuries found	96 (8%)	2 (2%)	0.09
Traumatic hernia	90 (7%)	7 (8%)	0.99
Diaphragm	200 (16%)	14 (16%)	1
Stomach	93 (8%)	9 (10%)	0.48
Liver	364 (29%)	32 (36%)	0.38
Gallbladder or biliary	25 (2%)	3 (3%)	0.65
Spleen	357 (29%)	28 (32%)	0.65
Pancreas	117 (9%)	9 (10%)	0.97
Adrenal	48 (4%)	7 (8%)	0.12
Kidney	184 (15%)	20 (23%)	0.07
Omentum	40 (3%)	4 (5%)	0.73
Mesentery	398 (32%)	41 (47%)	0.008
Small bowel	334 (27%)	21 (24%)	0.60
Large bowel	313 (25%)	34 (39%)	0.009
Rectum	34 (3%)	3 (3%)	0.98
Bladder or ureter	94 (8%)	7 (8%)	1
Vascular	164 (13%)	18 (20%)	0.08

Continuous data presented as: median (interquartile range [IQR]).

S-SSI = superficial surgical site infection.

Post-operatively, patients who suffered an S-SSI more frequently suffered an incisional hernia, fascial dehiscence, entero-atmospheric fistula, pneumonia, organ/space SSI, and sepsis than those without S-SSI (Table 4). Furthermore, S-SSI was associated with prolonged intensive care unit and hospital lengths of stay but was not associated with increased in-hospital mortality.

Table 4.

Outcome Measures

All patients (n = 1,322)	No S-SSI (n = 1,234)	S-SSI (n = 88)	p
Deep vein thrombosis	35 (3%)	6 (7%)	0.08
Incisional hernia	73 (6%)	32 (36%)	<0.001
Fascial dehiscence	46 (4%)	21 (24%)	<0.001
Entero-atmospheric fistula	17 (1%)	11 (13%)	<0.001
Evisceration	10 (1%)	2 (2%)	0.42
Pneumonia	178 (14%)	32 (36%)	0.005
Organ/space SSI	19 (2%)	17 (19%)	<0.001
Sepsis	218 (18%)	31 (35%)	<0.001
Multiple organ failure	70 (6%)	8 (9%)	0.28
Intensive care unit length of stay	2 (0–7)	7 (2–16)	<0.001
Hospital length of stay	10 (6–19)	25 (14–38)	<0.001
In-hospital mortality	56 (5%)	1 (1%)	0.21

Continuous data presented as: median (interquartile range [IQR]).

S-SSI = superficial surgical site infection.

The S-SSI predictive model demonstrated moderate discrimination with an AUROC of 0.78 (95% confidence interval [CI], 0.72–0.85) on the model development cohort and an AUROC of 0.69 (95% CI, 0.56–0.81) on the test cohort (Fig. 2) [32, 33]. The parameters that contributed the most to the model were damage control laparotomy, full-thickness large bowel injury, and skin closure (Table 5).

FIG. 2.

Receiver operating characteristic curve for superficial surgical site infection (S-SSI) predictive model on the (A) training set and (B) the test sample set.

Table 5.

Predictive Model for S-SSI Development

Parameters	Odds ratio (95% CI)
Damage control laparotomy	5.88 (2.93–12.09)
Full-thickness large bowel injury	1.91 (0.94–4.72)
Skin closure	1.62 (0.92–3.66)
BMI	1.34 (1.02–1.74)
Large bowel resection	1.32 (0.78–3.00)
OR duration	1.19 (0.95–1.58)
Age	1.18 (0.94–1.57)
Small bowel resection	1.12 (0.71–2.17)
Wound class	1.05 (0.82–1.42)
Penetrating mechanism	0.95 (0.55–1.48)
Last OR temperature	0.94 (0.65–1.14)
Year	0.84 (0.61–1.07)
Full-thickness small bowel injury	0.83 (0.36–1.36)
Intra-operative PRBCs	0.60 (0.22–1.04)
Intercept	0.02 (0.01–0.04)

S-SSI = superficial surgical site infection; CI = confidence interval; BMI = body mass index; OR = operating room; PRBCs = PRBC = packed red blood cells.

Discussion

A model using patient parameters available to the surgeon during index emergency trauma laparotomy can predict risk of S-SSI at the completion of the operation. The most important variables are damage control laparotomy, full-thickness large bowel injury, and skin closure. Although only 7% of patients who underwent an emergency trauma laparotomy developed an S-SSI, the present study found that patients with S-SSI suffer substantial morbidities including infectious complications, fascial dehiscence, and prolonged hospitalizations.

Prior work from our institution demonstrated that there was a lack of consistency in the decision for skin closure after emergency trauma laparotomy between trauma surgeons [34]. One reason may be that surgeons, regardless of years of experience, exhibit substantial variability in their estimates of S-SSI risk (K.D. Isbell, unpublished data). Although other scores and models exist for predicting S-SSI risk, they have multiple limitations. Simple risk stratification scores such as the National Nosocomial Infections Surveillance index are not trauma-specific [35]. On the other hand, trauma-specific models often include many parameters and not all may be available at the end of a laparotomy [36,37]. For example, Durbin et al. [37] performed a multivariable analysis of data contained within the Trauma Quality Improvement Project (TQIP) database to predict S-SSIs. Similar to our model, the TQIP model included age, penetrating mechanism, and small and large bowel injuries, and their model had an AUROC of 0.67. However, their model also included patient comorbidities; complete patient histories may not be able to be ascertained prior to emergency laparotomy. Additionally, injury severity score was included; complete diagnostic work-up required to calculate the score is often deferred in emergency laparotomies.

Other trauma-specific models may include parameters that may not be universal such as admission to the intensive care unit, the criteria for which may differ between hospitals [35]. Although addition of other variables may improve the prediction of S-SSIs, being able to predict the probability of S-SSI at the point of care (i.e., at the time of completion of the index laparotomy) allows potential modification of treatment strategies starting with the decision of whether or not to close the skin.

Several published models predicted SSIs after trauma or emergency laparotomy combine superficial with deep and organ/space infections as the outcome of interest [35,36,38,39] However, risk factors have been shown to differ in magnitude and significance between the levels of SSIs [40]. Models for both organ/space SSIs and S-SSIs have been developed by our group using Bayesian regularization [21]. Although both models share common elements, those parameters do not exert the same influence on the probability of SSI. Additionally, there are parameters, such as skin closure, that only apply to S-SSIs and not to organ/space infections.

Prediction models derived using traditional multivariable analytic strategies focus on identifying independent risk factors for SSIs [36]. These models provide odds ratios for each identified risk factor. In contrast, the current Bayesian model provides a probability of S-SSI using predictor variables all of which influence but not all are independently associated with the outcome. This modeling strategy can increase the number of variables in the model, but this complexity can be overcome through the creation of a user-friendly web interface (Fig. 3). Real-time data entry will allow trauma surgeons to calculate the probability of an S-SSI, and to vary that probability based on whether or not the skin is closed. That is, at the time of closure at the index laparotomy, the probabilities of S-SSI if the surgeon closes the skin or leaves it open can be calculated. Although leaving the skin open reduces the risk of S-SSI, the strategy does not completely preclude a subsequent S-SSI. Further trials are necessary to determine if use of our model versus usual care to determine skin closure strategy improves outcome.

FIG. 3.

Superficial surgical site infection (S-SSI) risk calculator web application with data representative of the average trauma patient from the multicenter cohort.

Despite implementation of multiple interventions for prevention of S-SSI, including programs such as the Surgical Care Improvement Project (SCIP) [41], optimal prevention for S-SSI after emergency laparotomy remains unknown. There is a substantial knowledge gap regarding accurate risk stratification of trauma laparotomy patients for S-SSI and to the risks versus benefits of promising interventions that range in treatment intensity, risks, resource utilization, and costs. Additionally, there is no consensus among surgeons for thresholds for classifying patients as low, moderate, and high risk. Therefore, surgeons vary in their decision-making regarding skin management, including use of adjuncts such as negative pressure wound therapy, after emergency laparotomy. Perhaps most importantly, patient preferences are unknown with regards to the risk they are willing to accept of a S-SSI with a closed skin incision. Unfortunately, because of the emergency nature of trauma laparotomies, shared decision-making is not always feasible. Thus, further research is necessary to better understand patient-reported outcomes both with open wounds and with S-SSIs to inform cutoffs for changing care. Additionally, the calculator can be used to identify risks at which novel or expensive treatments are cost-effective.

There are several limitations of this study. First, the study was performed at a single center and was retrospective in design. To evaluate our predictive model and assess generalizability, a multicenter, prospective observational study is currently planned. Second, given that not all patients follow-up routinely and that S-SSIs may develop after the post-operative visit, the rate of S-SSIs may be underestimated. Additionally, death may be a competing outcome to S-SSI, and patients who died were not excluded from the calculator. More rigorous follow-up and sensitivity analyses are planned for the prospective validation study. Third, the use of adjuncts such as negative pressure wound therapy devices was not standardized across all patients which may decrease the accuracy of the calculator. Going forward, the calculator is intended to be used to test the use of skin management strategies and other interventions such as extended duration antibiotics in patients at moderate to high risk for S-SSI. Fourth, the AUROC of 0.69 only demonstrates moderate discrimination. However, the model is not designed to be 100% accurate at identifying patients who will develop S-SSI, but rather to stratify patients based on risk of S-SSI at a time point at which management can be altered. Additionally, as more data becomes available from the multicenter study, the Bayesian models may be iteratively updated and improved.

Conclusion

A predictive model for S-SSI was built using factors available to the surgeon upon index emergency trauma laparotomy closure. A web-based calculator has been established to allow real-time estimation of the probability of S-SSI. This calculator may be used to standardize intraoperative and postoperative care and to identify high-risk patients in whom to test novel preventative strategies and improve overall outcomes for patients who require emergency trauma laparotomy.

Authors' Contributions

Study design: Wei, Hatton, Green, Truong, Pedroza, Harvin, Wade, Kao. Data collection: Wei, Hatton, Woloski, Harvin. Data analysis: Wei, Hatton, Green, Truong, Pedroza. Data interpretation: Isbell, Wei, Hatton, Green, Truong, Pedroza, Kao, Wade, Harvin. Writing: KDI, SW, GEH, CG, VT, CP, JW, JAH, CEW, LSK. Critical revision: KDI, SW, GEH, CG, VT, CP, JW, JAH, CEW, LSK.

Financial Information

Dr. Wade receives funding from the William Stamps Farish Fund, the Howell Family Foundation, and the James H. “Red” Duke Professorship. Dr. Harvin receives support from the Arsenal Medical Grant for ResQFoam. Dr. Wei and Dr. Hatton are supported by a T32 fellowship (grant no. 5T32GM008792) from National Institute of General Medical Studies of the National Institutes of Health. Dr. Kao is supported by a Surgical Infection Society Innovation Grant.

Author Disclosure Statement

The authors have no conflicts of interest.