Outcomes after surgical repair of primary parastomal hernia

This is a retrospective, descriptive study on surgical treatment of PSH surgery in Horsens Regional Hospital, which is one of the five specialized hernia centers in Denmark. All surgical procedures are registered in the electronic patient journal system (EPJ) and coded according to the International Statistical Classification of Disease (ICD-10). All patients undergoing PSH repair during January 2017 till March 2021 were identified by retrieving procedure codes regarding PSH correcting surgery from the electronic patient system, EPJ. As this is a quality study, our application only allowed us to go back five years in the medical records, thus limiting us in retrieving data on all the patients located in the register. We identified 119 patients, and their data were reviewed by one of the authors. The study included patients who underwent both urgent and elective parastomal hernia repair. Inclusion criteria were age over 18 and no former parastomal hernia repair. Sex, age, body mass index (BMI), smoking status and comorbidities were obtained. Details regarding the PSH and the PSH correcting surgery was registered along with postoperative events including Clavien-Dindo score, postoperative surgical complications, recurrence, and time of recurrence (see Table 1 for further details). PSH recurrence is defined as PSH detected following primary repair through either clinical assessment by a surgeon or imaging. If no event regarding PSH were to be found in the records, this was regarded as no recurrence. Follow up time was a mean of 72 months and SD of 15.2 months (see Table 1).

In addition to univariate chi-square analyses (Monte Carlo simulations for small sample sizes), we also performed a multivariate analysis that shows a fuller picture by accounting for several variables at once. The univariate analysis, which is subject to disproportionate influence, is thus corrected for by using the multivariate approach.

We distinguish between modifiable and non-modifiable risk factors. Non-modifiable risk factors refer to variables that are pre-existing to a patient’s health or condition before any medical intervention. These are pre-operative factors inherent to the patient, such as sex, age, smoking status, co-morbidity, BMI, ASA classification, type of stoma, EHS classification, urgency of repair, reason for stoma and reason for PSH repair (see Table 2). However, modifiable risk factors refer to factors that are related to, or directly caused by, the surgical intervention. These are post-operative surgical complications, type of repair, Clavien Dindo index score, and time of recurrence (see Table 3). In separating modifiable from non-modifiable risk factors, we further decided to compute two separate generalized linear models (GLMs) using R Studio (V. 2023.03.0). Following statistical 2-step recommendations in variable selection [17], we first identified candidate variables based on existing literature and expert knowledge. We then subjected the selected variables to Least Absolute Shrinkage and Selection Operator (LASSO from “glmnet”) as recommended by Heinze et al. [18]. LASSO is a form of linear regression that selects variables and regularizes by adding an L1 penalty. This is to minimize residuals, which effectively simplify models by driving some coefficients to zero, for automatic feature selection and better interpretability. The selected candidate variables were further fitted to a GLM, which we report on below (Table 4). The same approach was used for modifiable risk factors. Reasonable outliers were excluded based on visual inspection of a normal Q-Q plot.

The overview in Table 1 provides detailed patient characteristics and composition. Patients ranged in age from 31 to 94 years, with an average of 68.1 years, where the majority (~ 83%) were non-smokers. Gender distribution was relatively equal, with approximately 41% male and 59% female. While about ~ 62% had no comorbidity, about 13.5% had either diabetes mellitus (DM) type 1 or 2. Additionally, the average BMI was 27.14 indicating an over-weight average patient group. Colostomy (~ 60%) and ileostomy (~ 30%) were the most frequent types of stomas, where colorectal cancer (~ 42%) and inflammatory bowel disease (IBD) (~ 19%) where the most frequent causes. The two most frequent reasons for hernia correcting surgery were intestinal obstruction (~ 27%) and pain related to ostomy (~ 26%). The most frequent type of repair was the Sugarbaker procedure (~ 61%). Our data contains 2/3 elective surgeries, leaving 1/3 for acute surgery. We have an over-representation of patients scoring Grade 1 in Clavien Dindo (~ 82%) compared to the other grades. While 57% experienced no post-operative complication, ~ 18% reported seroma post-surgery. Overall, we have a recurrence rate of 17.64%.

Table 2 provides a detailed overview of the univariate analyses (Chi-squared, unless otherwise indicated) for non-modifiable risk factors against the recurrence rate, whereas Table 3 provides an overview of the univariate analysis for the modifiable factors. We found that younger patients, ≤55 years old, have a significantly higher recurrence rate compared to older patients (p = 0.0019). We also found that ileostomy (p = 0.0436) and loop ileostomy (p = 0.0254) reach significance, indicating that these types of stomas are more susceptible to recurrence. Additionally, reasons for stoma, including colorectal cancer (p = 0.0187), IBD (p = 0.0314), constipation (p = 0.0184), and EHS Classification 1 (p = 0.0474) reached significance.

However, in terms of modifiable factors, re-operation (p = 0.0554) and Clavien Dindo grade 3b (p = 0.0704) both show a tendency.

We employed LASSO regularization in linear regression using a dataset of 115 observations (4 were excluded due to being outliers) with all factors in Table 2 as predictor variables. Regression results are summarized in Table 4. The selected model yielded 6 non-zero coefficients. The chosen lambda value for regularization was 0.0487 (lambda 1 standard error from the minimum lambda: 0.1235). Performance evaluation indicated a mean squared error of 0.1389. The LASSO procedure effectively identified significant predictors, achieving sparsity by shrinking several coefficients to zero. The candidate variables were selected as predictors for our GLM model below.

We computed a GLM with recurrence rate as dependent variable and the LASSO-qualified variables as independent variables. These were: age, diabetes, loop ileostomy, inflammatory bowel disease, fecal incontinence, EHS classification, and constipation (reason for stoma). The first model resulted in extreme estimates for the loop ileostomy, so we suspected that multicollinearity was interfering with the quality of the fitted model. By removing loop ileostomy as a factor, we ran a second model and compared it to the first one. As the second model (without loop ileostomy) proved to perform better (BIC: 112.13 < 114.59. Log likelihood: -39.46 < 38.31), we went with that model. The GLM revealed, similar to the univariate analysis, that younger patients had a significantly higher chance for recurrence (OR = -0.052, SE = 0.02, p = 0.0292), and that only patients with DM (3 levels: yes, no, unknown) had a significantly greater chance for recurrence (OR = 1.727, SE = 0.83, p = 0.0386). We also observed that IBD (OR = 1.604, SE = 0.72, p = 0.0260), fecal incontinence (OR = 2.455, SE = 0.92, p = 0.0078) and constipation (OR = 2.719, SE = 0.99, p = 0.0060) all significantly affected the recurrence rate. In other words, we observe an age-related decrease in recurrence rates and that patients with either fecal incontinence or constipation prior to the operation have a 23.29% and 30.33% chance to experience recurrence, respectively. See Table 4 for confidence intervals.

To be sure, this model suggests that for a 35-year-old patient with fecal incontinence has 70.22% chance for recurrence, compared to one who has both fecal incontinence and diabetes where chances are 83.98%.

That’s an increase of 13% simply due to DM. However, this should be interpreted with caution given the low number of observations.

In a similar fashion, we employed the LASSO technique using the same dataset, however, with all factors in Table 3 as predictor variables. The LASSO regression analysis minimized all coefficient estimates to zero, indicating that none of the variables demonstrated sufficient predictive strength to be retained. As a result, no variables were selected for further analysis or consideration in the model.