A prediction model of surgical site infection after instrumented thoracolumbar spine surgery in adults

This was a retrospective cohort study of all instrumented spinal surgery procedures of the thoracic, lumbar and thoracolumbar spine that have been performed in adult patients (≥ 18 year) in an academic referral center for spinal pathology from January 1, 1999, up to January 1, 2016. Patients diagnosed with an infection after instrumented spinal surgery elsewhere were excluded as well as patients for whom the medical files for at least up to 1 year after surgery were not available.

All operations were performed by 3 experienced orthopedic surgeons specialized in spine surgery. In select cases when neurological decompression was needed, neurosurgeons participated in the operation. All patients underwent an instrumented posterior (posterolateral or interbody) spinal fusion of the thoracic, lumbar and thoracolumbar spine, with or without an additional procedure (anterior fusion or release, spinal decompression, the removal of instrumentation, tumor resection or corpectomy/osteotomy).

Patients were followed for a minimum of 1 year after the index operation to monitor all complications and incidences of revision surgery. All complications, extensive demographics, comorbidity and surgical details were recorded by collecting data from all available electronic and paper records of the patients. The primary outcome of interest was the occurrence of SSI. The diagnosis of SSI was based on the CDC criteria (Centre for Disease Control and prevention) [16] and the Dutch national PREZIES network (prevention of hospital infections through surveillance) [17]. An SSI was considered to be deep if it presented at the site of the operation with involvement of the subfascial tissues.

An often-used rule of thumb states that at least 10 events (i.e., occurrences of SSI) are needed per predictor variable that is tested in the prediction model development step [18]. When more predictor variables are added to the model, the probability of overfitting (i.e., the model predicts exceedingly worse for patients not comprised in the derivation cohort) increases. As a result, we needed to perform a pre-selection of all baseline characteristics of those that we thought would be most likely to result in an accurate prediction model. The pre-selection was based on what was already known from other studies, the distribution of the predictor in our sample, and experience in our own hospital. Using this method, we were able to reduce the initial set of potential predictor variables to 8, i.e., age, body mass index (BMI: kg/m²), smoking status, diagnosis, revision surgery, ASA (American Society of Anesthesiologists) physical status, surgical invasiveness index (SII) [19] and the use of nonsteroidal anti-inflammatory drugs (NSAIDs) preoperatively.

Smoking status was dichotomized into currently smoking yes or no, independent of the volume and tobacco product used. All passive smokers and ex-smokers were regarded as non-smoker. ASA physical status, a classification to assess the fitness of the patient before surgery, was coded according to the five-category physical status classification system of the American Society of Anesthesiologists in 1963 (1 = healthy person, 2 = mild systemic disease, 3 = severe systemic disease, 4 = severe systemic disease that is a constant threat to life, 5 = a moribund person who is not expected to survive without the operation) [20]. The surgical invasiveness index is a validated instrument with a range from 0 to 48 points, containing the sum of the following six weighted surgical components: the number of levels anterior decompressed, the number of levels anterior fused, the number of levels anterior instrumented, the number of levels posterior decompressed, the number of levels posterior fused and the number of levels and posterior instrumented. The weight of each component represents the number of vertebral levels at which each respective component has been performed [19]. A higher score means higher invasiveness. For example, in an L4–L5 posterior fusion and decompression with the use of an intervertebral cage, and posterior instrumentation, the score would be 9 (anterior fusion = 2, anterior instrumentation = 2, posterior decompression = 1, posterior fusion = 2, posterior instrumentation = 2).

Diagnosis of the included patients was divided into 4 subgroups, i.e., one- or two-level degenerative disorders (with or without neurologic compromise), failed back syndrome (patients that had already undergone previous spine surgery on the same level), trauma (unstable vertebral fractures with or without neurological compression) and other (adult spinal deformity, spinal metastases/malignancy, spondylodiscitis). Nonsteroidal anti-inflammatory drugs use was defined as the daily use of NSAIDs before surgery for more than 1 week and still in use at the time of surgery.

Incomplete patient records were imputed using stochastic regression imputation, to prevent a potentially considerable loss of statistical precision and to decrease the probability of biased results when compared to using only complete patient records (see Table 1). We used predictive mean matching to generate the imputed values. After imputation, we included all potential predictor variables in a logistic regression analysis. Using stepwise backward elimination on the hypothesized predictor variables, we excluded nonsignificant predictors from this category to arrive at a more parsimonious model. As suggested by prediction modeling guidelines, we used a less strict alpha for eliminating variables from the model to prevent too early deletion of potentially important predictor variables [21]. We chose to use an alpha of 0.10 compared to the conventional 0.05.

For continuous variables, the association is assumed to be linear. Nonlinear effects were visualized using plots and formally tested using restricted cubic splines, a regression technique that can be used to test for deviations from linearity [21]. In case of significant evidence of a nonlinear relation, the continuous variable was categorized into clinically meaningful categories.

The model’s performance was quantified using measures of discriminative ability and measures of calibration. We assessed the model’s ability to discriminate between those who developed SSI and those who did not by computing the area under the receiver operating characteristic (ROC) curve (AUC). This AUC can range from 0.5 (no discriminative ability) to 1.0 (perfect discriminative ability). Model calibration (i.e., agreement between predicted and observed probabilities) was evaluated by visual inspection of the calibration plot and by computing the Hosmer and Lemeshow goodness-of-fit test (HL test). A significant HL test indicates evidence against good model fit.

We internally validated the initial prediction model using standard bootstrapping techniques. Using results from the bootstrap procedure, we penalized the model’s regression coefficients, so future predictions will be less extreme (to counter the effect of overfitting) by multiplying them with a shrinkage factor, and re-estimating the model intercept. Also, we computed the estimated optimism in the AUC. This is a measure of the likely difference in AUC when the model is applied to future patients. All analyses were performed using R version 3.3.3.