It is estimated that the prevalence of malnutrition in patients undergoing elective abdominal surgery reached 50 %(Reference Sungurtekin, Sungurtekin and Balci1). This is particularly true for Crohn’s disease (CD) patients who, in a recent prospective study, were reported to be malnourished in up to 38 % of cases(Reference Sun, Kong and Jing2). Moreover, it has been demonstrated that the risk of malnutrition is associated with increased postoperative complications, mortality, longer hospital stays and higher costs(Reference Correia and Waitzberg3). On the other hand, preoperative improvement of the patient’s nutritional status and early postoperative nutritional support significantly decrease these risks(Reference Jie, Jiang and Nolan4). In CD patients, the risk of malnutrition is related to inflammatory activity, decreased oral intake or impaired digestive capacity or adsorption. Perioperative artificial nutrition could be useful for decreasing inflammatory activity and, therefore, for decreasing morbidity. To limit this, some approaches, including a temporary diverting stoma, have been proposed; however, this could lead to an additional risk of specific complications, such as elevated output and dehydration. Despite the international guidelines regarding nutritional management in CD patients, the recommendations on preoperative nutritional support are principally based on small-sample retrospective studies that have, for the most part, used total parenteral nutrition(Reference Jacobson5). The role of malnutrition in patients with CD has previously been reported; however, the sole cause of the malnourishment might have been misinterpreted based on multiple confounding factors. The present study was designed to shed light on this conundrum, using a novel methodology called entropy balancing (EB) (Reference Hainmueller6). EB is a novel pre-processing technique for researchers to achieve covariate balance in observational studies using a binary treatment, which is more accurate than any propensity score matching system(Reference Hainmueller6).
Most commonly, researchers ‘manually’ iterate between propensity score modelling, matching and balance checking until they attain a satisfactory balancing solution. The hope is that propensity score matching would then stochastically balance the covariates; however, this requires finding the correct model specification and often on fairly large samples. As a consequence, low balance levels prevail in many studies, and the user experience can be tedious. Instead, EB involves a reweighting scheme that directly incorporates covariate balance into the weight function applied to the sample units in order to generate a better-performing matching system(Reference Hainmueller6).
The goal of this study was to demonstrate exclusively the role of malnutrition on postoperative outcomes in patients with CD, using a novel, more accurate processing technique called entropy balancing. This method could determine the potential role of pre-habilitation programmes on reducing surgical morbidity in patients with CD and how much the risk associated with surgery could be mitigated by treating malnutrition.
Methods
This was a retrospective cohort study of a prospectively maintained database. This dedicated database is designed to prospectively monitor clinical outcomes in our unit and identify potential associations with preoperative, intraoperative and postoperative factors. The study was conducted according to the guidelines laid down in the Declaration of Helsinki. The study was approved by the Institutional Review Board as part of an internal surgical auditing programme (Ravenna Surgical Quality Improvement Program – RaSQIP) with the following code: 214/2016/O/OssN. Written consent was obtained from patients for the use of their treatment-related data for research purposes upon initial consultation. All consecutive patients who underwent elective surgery for CD between January 2017 and October 2022 were included. Patients who underwent surgery without bowel resection (i.e. loop ileostomy for severe perianal disease and/or exploration under anaesthesia), as well as those with unavailable nutritional data, were excluded from the study. All patients were preoperatively screened for malnutrition with both the nutritional risk screening (NRS-2002), derived from the European Society for Clinical Nutrition guidelines(Reference Kondrup, Allison and Elia7), and the BMI (kg/m2). Other demographic data, routinely recorded for all patients, included serum levels of lymphocytes (10^9/L), Hb (g/dl), creatinine (mg/dl), albumin (g/L), age-adjusted Charlson Comorbidity Index (CACI)(Reference Chang, Yin and Wei8), Eastern Cooperative Oncology Group – Performance status (ECOG-PS) (which is also utilised in the authors’ practice for non-cancer patients as a frailty tool)(Reference Ethun, Bilen and Jani9,Reference Kochar, Orkaby and Ananthakrishnan10) , American Society of Anaesthesiologists (ASA) score, smoking habit and Harvey-Bradshaw score(Reference Hanzel, Bossuyt and Pittet11) for CD activity. To reduce potential selection biases, the study included only patients with at least a 4-week wash-out from biological or steroidal therapy before surgery. For study purposes, the ASA score was dichotomised into 1–2 (mild systemic disease) v. 3–4 (severely ill). Other disease-specific factors included in the analysis were type and number of previous bowel surgeries, type of disease (fistulising v. stenosing), presence of an enterocutaneous fistula and need for preoperative nutritional support. The surgical factors were type of surgery (only ileal, only colonic, ileo-colonic), surgical approach (open v. laparoscopic), number of resections during the same procedure (single or multiple) and need for a diverting ileostomy. All patients were treated according to a comprehensive enhanced recovery pathway. The pathway included avoiding preoperative fasting, favouring opioid-sparing analgesia, prophylaxis of postoperative nausea and vomiting, early resumption of a normal diet and ambulation. The protocol included the early removal of the Foley catheter and the avoidance of both drain placement and the use of a nasogastric tube whenever possible.
The postoperative course was classified prospectively at the time of discharge, and postoperative complications were recorded and considered up to 90 d after discharge. They were classified according to the Clavien-Dindo Classification(Reference Dindo, Demartines and Clavien12). The Comprehensive Complication Index (CCI®)(Reference Slankamenac, Graf and Barkun13) scores were calculated from all the single postoperative complications using an online calculator (www.assessurgery.com/about_cci-calculator/). The length of stay (LOS) was defined as the time from the day of surgery to the day of discharge. Based on nutritional risk screening (NRS), the patients were divided into two groups: patients with a low or null risk of malnutrition (NRS < 2) and patients with moderate or severe risk (NRS ≥ 2) in agreement with the previous literature(Reference Huisman, Veronese and Audisio14). The postoperative course was compared between the two groups regarding ileus and anastomotic leak rates, readmission, overall and major complications (Clavien-Dindo Classification ≥ 3a), CCI score and LOS before and after balancing.
Statistical analysis – entropy balancing methodology
The data were reported as percentages or means and sd. The two groups were compared for both the confounding factors and the nutrition-related parameters.
Hainmueller’s entropy balance analysis(Reference Hainmueller6) is used to achieve a balance between treatment and control groups in observational studies, aiming to mimic the conditions of a randomised controlled trial. This method adjusts the weights of the observations to ensure that the distributions of the covariates are as similar as possible between the groups, thereby reducing bias and making the groups comparable.
The process of entropy balance works as follows: authors identified the covariates or confounding factors that need to be balanced between the treatment and control groups. These are characteristics that could influence the outcome of interest and should be equally distributed across the groups. Initial weights are assigned and calculated for all observations. In an unweighted analysis, all observations would ideally have equal weight. However, in real life, two groups are typically different for many covariates, bringing different weights. The first step is then to optimise weights via entropy balancing. The goal is to find a set of weights for the observations that minimises the difference in the covariate distributions between the treatment and control groups. The optimisation process involves an entropy function, which measures the divergence between the initial and adjusted weights. The function ensures that the adjusted weights are as close as possible to the initial weights, maintaining the original sample structure, which is why categorical variables did not differ after the weighting. On the contrary, continuous variables may vary slightly, as they inherently represent the weight of the covariate. Weights were computed and obtained using the calliper matching method in Stata, with the default calliper set to satisfy the balance constraints (mean and sd or variance) between the groups. Constraints are added to ensure that the weighted means of the covariates are equal across the groups. This means that after weighting, the average value of each covariate should be the same in both the treatment and control groups. The algorithm adjusts the weights to minimise the entropy function while satisfying the balancing constraints. Once the optimal weights are found, they are applied to the observations. The resulting weighted sample should have balanced covariates, meaning that the treatment and control groups are comparable with respect to the measured covariates. As noted by Heinmüller et al. (Reference Hainmueller6), classical statistical analyses can be performed with weighted data. Although EB may violate the independence assumption, its impact on statistical analyses, such as regression, is minimised due to its rigorous balancing methodology. After each step, a logistic regression multivariate analysis and linear regression were carried out to evaluate the risk factors related to the postoperative results and to calculate the OR or mean difference (MD).The balanced sample allows for a more accurate estimation of the treatment effect, as it reduces the bias from confounding variables. In summary, entropy balance is a powerful tool for reducing bias in observational studies by creating weighted samples with balanced covariate distributions, thus facilitating more reliable causal inferences(Reference Hainmueller6).
In an ideal world, to measure the exact role of malnutrition alone over its absence, we should know the impact of both conditions (good nutrition and malnutrition) on the surgical outcomes of every single patient, but this is obviously impossible since a patient can be only well or malnourished. A possible way to solve the problem is to identify the causal effect of preoperative malnutrition over its absence. The causal effect defines how the malnourished patients would fare if they were well-nourished. The results obtained are called counterfactual outcomes. The analysis was made in two steps: in the first one, only non-nutritional confounding factors were balanced to estimate the true effect of preoperative status on postoperative outcomes. In the second step, also the clinical and biochemical indicators of malnutrition were balanced.
Given that the association between malnutrition factors and postoperative complications (using regression tests) is already widely demonstrated in the literature, in this study, we wanted to make a different reasoning. The statistical method allowed us to simulate a hypothetical intervention on our patients. In the first step, the intervention was to make them similar in terms of general characteristics, leaving the differences regarding nutrition intact. In the second step, on the other hand, the model homogenised the nutritional values, leaving the basic characteristics of the patients unchanged. In other words, several confounding factors could magnify or mitigate the true impact of preoperative malnutrition on the postoperative course. The entropy balance method is designed to try to solve this problem. Differences between the groups were measured using standardised differences (d-value). In the context of EB, d-values are employed as a metric to ensure that the covariate means in the treatment and control groups are balanced after applying weights. EB transforms categorical variables into binary indicator variables and applies balancing techniques typically used for continuous variables. This allows for the balancing of categorical distributions across treatment and control groups by ensuring that the weighted means and variances of these binary indicators are aligned. Despite treating categorical variables in this transformed binary form, the goal remains to achieve balance in the original categorical distributions across groups. The d-value(Reference Cohen15) essentially reports the differences between groups in units of sd. When two groups have no difference on a certain variable, their populations completely overlap. In statistics, the ‘non-overlap’ population refers to the portion of one group’s population that remains different for such a variable, either at baseline or after an intervention. The d-value provides a standardised measure of effect size, with thresholds defined and benchmarked in the literature by Cohen et al. (Reference Cohen15). A d-value ≤ 0·2 indicates a percentage of the non-overlap population of ≤ 15 % (small between-group difference); a d-value > 0·2 and ≤ 0·5 (average difference) indicates a percentage of the non-overlap population from 15 to 33 %; a d-value from > 0·5 to 1·0 (large difference) indicates a percentage of the non-overlap population of > 33 %(Reference Cook, Cook and Therrien16,Reference Watson and Elliot17) . The d-values serve as a guide for assessing the balance of covariate distributions rather than as a strict measure of distribution shape.
The statistical analysis was carried out using STATA software (StataCorp 2017, StataCorp LP). Entropy balancing was carried out using the ‘ebalance’ module, for which a tutorial is linked in the references(18).
Results
The demographic characteristics of the two groups (Fig. 1) are reported in Table 1. Of the 183 patients surgically treated for CD, 69 patients (37·7 %) were found to be at moderate or high risk of malnutrition based on NRS, while 114 (63·3 %) were considered well-nourished/low risk. As expected, the two groups had average to large differences in BMI (d = 1·000), preoperative Hb (d = 0·715), creatinine (d = 0·553) and serum albumin (d = 0·981), while lymphocyte count (d = 0·124) showed small difference. The malnourished patients had a higher CACI score (d = 0·257), higher ECOG-PS and ASA scores (d = 0·689 and d = 0·327, respectively) and a higher H-B score (d = 0·696). They also required preoperative nutritional support more often (d = 0·766), and the rate of a diverting ileostomy was almost three times higher in this group (d = 0·648). The two groups showed small differences regarding smoking habit (d = 0·148), previous bowel surgery (d = 0·154), disease behaviour (d = 0·104) and rate of enterocutaneous fistulas (d = 0·162). On the contrary, the two groups were almost equivalent with respect to age (d = 0·026), type of surgery (d = 0·060), laparoscopic approach (d = 0·060) and the intraoperative need for multiple resections (d = 0·063).
Flow chart of the included patients.
Characteristics of the patients at baseline and after entropy balancing (Numbers and percentages; mean values and standard deviations)
CACI, age-adjusted Charlson Comorbidity Index; ECOG-PS, Eastern Cooperative Oncology Group – performance status; ASA, American Society of Anaesthesiologists; H-B score, Harvey-Bradshaw score for Crohn’s disease activity. A d-value ≤ 0·2 indicates a percentage of the non-overlap population of ≤ 15 % (small between-group difference); a d-value > 0·2 to ≤ 0·5 (average difference) indicates a percentage of the non-overlap population of between 15 and 33 %; a d-value > 0·5 to 1·0 (large difference) indicated a percentage of the non-overlap population of > 33 %.
* STEP 1: To reweight the malnourished patient group for all the preoperative, intraoperative and disease-specific confounding factors. The first step corrects the clinical and operative parameters, which are not directly related to malnutrition or which are unmodifiable (i.e. number of surgeries, fistulising disease).
** STEP 2: To reweight malnourished patients for all those parameters directly related to malnutrition, thus potentially modifiable.
Postoperative results (Table 2) before balancing showed similar rates for postoperative ileus, anastomotic leaks and Clavien-Dindo Classification ≥ 3a complications (P = 0·757, P = 0·532 and P = 0·670, respectively). Malnourished patients experienced a 3-fold higher readmission rate (15·9 % v. 5·3 %; OR 3·41; 95 % CI 1·20, 9·70; P = 0·021) and a prolonged LOS (6·5 v. 5 d; MD = 1·47; 95 % CI 0·33, 2·6; P = 0·012). They also showed a higher rate of overall complications (59·4 % v. 35·1 %; OR 2·71; 95 % CI 1·46, 5·01; P = 0·002) and a higher CCI score (13·1 v. 7·2; MD = 5·90; 95 % CI 2·29, 9·45; P = 0·001).
Postoperative outcomes of the two groups at baseline (Numbers and percentages; odds ratios or mean difference and 95 % confidence intervals)
MD, mean difference; LOS, length of stay; CDC, Clavien-Dindo Classification score; CCI, Comprehensive Complication Index.
After first reweighting (STEP 1)
As previously described, in the first reweighting, the two groups were equalised for all the confounding and non-modifiable factors. This created a ‘virtual malnourished group’ that was optimally balanced with the ‘normal nutrition’ group for all disease-specific and intraoperative characteristics (d = 0), thus enhancing the differences in the nutritional/modifiable parameters which remained the same as before balancing. The first balance included sex, age, CACI, ECOG-PS, ASA score, previous bowel surgery, type of disease (fistulising v. stenosing), Harvey-Bradshaw index, presence of enterocutaneous fistulas, preoperative nutritional support, type of surgery and surgical approach, number of resections and the need to create a diverting ileostomy. Postoperative results after the first reweighting are shown in Table 3. As in the baseline analysis, the malnourished patients experienced prolonged LOS (MD = 1·9; 95 % CI 0·11, 3·71; P = 0·038), higher overall complication rate (OR 4·42; 95 % CI 1·39, 13·97; P = 0·012) and a higher CCI score (MD = 8·9; 95 % CI 2·2, 15·7; P = 0·010). Moreover, they showed a trend towards a higher anastomotic leak rate (OR 6·69; 95 % CI 0·98, 46·4; P = 0·052), even if it was not statistically significant.
Postoperative outcomes of the two groups after step 1 and step 2 of entropy balancing (Odds ratios or mean difference and 95 % confidence intervals)
MD, mean difference; LOS, length of stay; CDC, Clavien-Dindo Classification score; CCI, Comprehensive Complication Index.
After second reweighting (STEP 2)
In the second reweighting, the two groups were equalised only for nutritional or preoperatively modifiable parameters, such as smoking habit, BMI, lymphocyte count, Hb, serum albumin and creatinine (d = 0). The other confounding factors were left as the baseline in order to simulate a nutritional intervention, which was effective in bringing the patient to an optimal nutritional level, while disease characteristics and intraoperative strategies were not modified. This statistical technique enables the researcher to highlight the role of nutritional optimisation. Postoperative results after the second reweighting are summarised in Table 3. The postoperative course of the normal nutrition group and the ‘virtual, malnourished’ patients balanced for nutritional parameters showed no differences.
Discussion and conclusions
The present study focused attention on some crucial aspects of nutritional status, which are highlighted in patients with CD by the use of entropy balance. Using innovative mathematical statistical systems was not a mere academic exercise in this case as it could help clinicians to prioritise preoperative interventions in order to mitigate substantial risk factors, such as malnutrition. First of all, as previously reported in the literature(Reference Kassin, Owen and Perez19), the preliminary analysis in this study showed that malnourished patients had a higher risk of readmission and a higher complications rate (including a higher CCI index) as well as a longer LOS. These results should be treated with great caution as they reflected the huge differences between the two groups. Malnourished patients are more fragile at baseline (worse ASA and ECOG scores) have more comorbidities and a less favourable disease history (forming fistulas). They more frequently had surgery in the past and had a higher Harvey-Bradshaw index. These differences could increase surgical and postoperative risks independently or by association, making the evaluation of their individual burden challenging. The EB analysis, as previously reported in the literature(Reference Ricci, Serbassi and Ingaldi20), allowed a more focused evaluation of the individual risk factors by weighting and thus equalising each factor between two very diverse groups.
Patient characteristics in the two groups were then divided into unmodifiable (e.g. type of disease, Harvey-Bradshaw index, CACI, ECOG, ASA, etc.) and potentially modifiable with a preoperative intervention (e.g. smoking habit, BMI, albumin, etc.). The first reweighting was carried out to correct the unmodifiable confounding factors in order to evaluate the true role of nutritional status alone. Once the two groups were equalised for those characteristics, the negative impact of malnutrition on the postoperative course was confirmed and highlighted, being an independent risk factor for complications and prolonged LOS. Interestingly, this first step also balanced the malnourished cohort for ‘diverting ileostomy creation’ by reducing its rate. By doing so, the EB allows to show a trend towards a higher anastomotic leak rate in patients with malnutrition alone, even if this was not statistically significant. The second step confirmed the burden of malnutrition in an even more spectacular way. Entropy balancing step 2 aimed at creating a virtual cohort of malnourished patients in which a hypothetical optimal preoperative nutritional intervention would correct all the nutrition-related characteristics between the two original groups. Interestingly, some non-marginal advantages could have been obtained by correcting the preoperative nutritional status. All the postoperative outcome differences between the two groups disappeared in this step, underscoring what would have happened if the malnutrition had been successfully treated.
This additional evidence reinforced the concept that a multimodal pre-habilitation, also associating the cessation of smoking and initiatives to prevent sarcopenia, could not be only a viable pathway but a necessary pathway in order to improve patients’ outcomes. While it is very well known that systemic inflammation or the inability of tolerating oral food could be reasons for failing to optimise patients, the results of this study should incentivise researchers to identify nutritional treatment, which could be more effective, even in these specific scenarios.
In order to treat malnutrition and reduce inflammation, many authors have already reported their results regarding the role of pre-habilitation before ileocolic resection in CD patients. The main impact on outcomes was obtained by nutritional correction, with a reduction in both anastomotic complications and the re-operation rate(Reference Yamamoto, Nakahigashi and Shimoyama21,Reference Ferrandis, Souche and Bardol22) .
The current study has some obvious limitations: first, the sample size and the retrospective nature of a population that was selected from a prospectively maintained database, and second, the few intrinsic limitations related to the EB methodology. The main weakness of this system is the assumption that preoperative conditions, such as malnutrition, could be completely remedied, generating a virtual group that becomes equivalent to the real study population.
On the one hand, this points out the need for intensive pre-habilitation programmes in order to mitigate the role of poor nutritional status. On the other hand, this is difficult to achieve in real life.
Although this system is purely a statistical tool, it enables clinicians to simulate a randomised control study and establish a foundational basis before deciding to invest in a nutritional pre-habilitation clinical trial.
While many studies have focused on BMI, albumin and other laboratory findings in order to describe the fitness level and the success of nutritional treatments many lack data about sarcopenia and myosteatosis(Reference Santarpia, Alfonsi and Castiglione23). Unfortunately, the present study made no exception for the latter as these parameters were not recorded in this series. Future research should use EB to evaluate the role of sarcopenia and myosteatosis as they have been defined to be key players in inflammatory bowel disease patients(Reference Nishikawa, Nakamura and Miyazaki24).
In conclusion, despite its limitations, the present study highlighted the role of malnutrition on surgical outcomes for CD patients from a different and innovative point of view. The present results, together with other experiences, could be the foundation for paving the way for future studies on preoperative nutritional treatment in patients undergoing major surgery.
Acknowledgements
The authors received no financial support for the research, authorship, and/or publication of this article.
G. T. and I. M.: study concept and study design; G. T., F. G. and D. Z.: data acquisition, analysis and interpretation; G. T., G. F. and F. M.: statistical analysis and manuscript preparation; G. U., I. M. and D. Z.: manuscript editing; G. T., G. U. and F. G.: manuscript review.
Montroni I. is on the speaking bureau of Olympus Europe SE for trans-anal Total Mesorectal Excision (TaTME) training courses and minimally invasive training courses. The remaining authors have no conflicts of interest to declare.
Data described in the manuscript, code book and analytic code will be made available upon reasonable request to the corresponding author.
The study was approved by the Institutional Review Board as part of an internal surgical outcome quality improvement programme (Ravenna Surgical Quality Improvement Program – RaSQIP).
