Identifying critically ill patients who benefit the most from nutrition therapy: the development and initial validation of a novel risk assessment tool

Identifying patients who are at risk of adverse events because of their nutrition status is a core competency of nutrition practitioners, recommended by clinical practice guidelines, and mandated by accreditation agencies [1–3]. Inherent in this discussion of nutrition risk is that patients at high risk are more likely to benefit from nutritional therapeutic interventions than those at low risk, as nicely demonstrated by Kondrup and colleagues [4]. Many scores or assessment tools exist to enable the quantification of nutrition risk [5–10]. For the most part, these tools were developed and validated in outpatient or inpatient settings but not specifically for the ICU setting [11]. In fact, most scores consider that all critically ill patients are at a high risk in terms of their scoring or risk assessment [4, 5].

We posit that this is not the case, and that not all critically ill patients are the same in terms of their nutritional risk. The evidence for this assertion comes from studies that demonstrate a differential treatment effect of artificial nutrition in different subgroups of ICU patients. In a recent analysis, we observed a significant inverse linear relation between the odds of mortality and total daily calories received [12]. An increase of 1,000 calories per day was associated with an overall reduction in mortality (odds ratio for 60 day mortality 0.76, 95% confidence intervals (CI), 0.61-0.95, P = 0.014). However, the beneficial treatment effect of increased calories on mortality was observed in patients with a body mass index (BMI) below 25 or 35 and above with no benefit for patients with a BMI of between 25 or less than 35. Similar results were obtained when comparing increasing protein intake and its effect on mortality in different BMI groups. One of the main inferences from this work is that not all ICU patients are the same with respect to their response to artificial nutrition.

So how do we begin to approach discriminating 'nutritional risk' in the critical care setting? In a recent International Consensus Guideline statement, Jensen and colleagues offer some ground breaking definitions of malnutrition relating it to both acute and chronic malnutrition and inflammation [13]. Consistent with this definition, in Figure 1, we present our conceptual model of how measures of acute and chronic starvation and inflammation may influence nutrition status at ICU admission and ultimately impact on patient outcomes. Our ultimate goal was to develop a score using the variables presented in the model that would quantify the risk of an individual patient developing adverse events and that may be potentially modifiable by aggressive or adequate nutritional intervention. In fact, to validate our score, we not only had to demonstrate that it discriminated risk among a heterogenous group of ICU patients, but also that the association between the risk score and outcome was modified by nutritional invention.

Thus, the purpose of the study was to develop and validate a novel method for quantifying risk of adverse outcomes that may be modified by nutrition therapy in the critical care setting, the NUTrition Risk in the Critically ill (NUTRIC score).

Figure 2 displays the observed and logistic modeled mortality by NUTRIC score. It may be seen that mortality clearly increases with the NUTRIC score, and that the logistic model (with NUTRIC score as a continuous predictor) appears to adequately fit the observed data. Moreover, the Hosmer-Lemeshow test did not suggest a significant lack of fit (P = 0.28). Figure 3 shows that NUTRIC risk score is strongly associated with observed days on mechanical ventilation among 28-day survivors (P < 0.0001). Although it appears that the linear regression model somewhat attenuates this trend at the extreme ranges of the score, which may suggest a non-linear relation, it should be noted that there are few patients with NUTRIC scores of 0, 9, and 10 and the lack-of-fit test does not suggest the linear model is inadequate (P = 0.67).

Figure 4 visualizes the association between nutritional intake and mortality by risk score. It appears that the association between risk score and mortality is attenuated in patients who meet their caloric targets and that increased nutritional risk is associated with reduced mortality in high-risk patients only. This apparent effect modification is statistically significant (P = 0.01).

We found that patients with missing values for weight loss in the previous three months or percentage of oral intake in the prior week had similar elevated mortality rates as those that reported some weight loss or less than 100% oral intake. Grouping patients missing these variables with those reporting weight loss or less than 100% oral intake resulted in these variables being significantly associated with 28-day mortality. Nevertheless, the performance of the overall risk score did not improve (and in fact worsened slightly) by the inclusion of these re-defined nutrition variables.

Recognizing that not all ICU patients will respond the same to aggressive nutritional interventions, we set out to develop and conduct some preliminary validation work on a score that helps practitioners discriminate who might benefit the most (or least) from nutritional therapy. Borrowing heavily from Jensen and colleagues elaboration on the associations between nutritional state and inflammation [13], we developed a conceptual model that linked starvation, inflammation, and outcome. We chose candidate variables to represent acute and chronic starvation and inflammation and collected these data in a multicenter observational study of 597 patients. We demonstrated that, individually, these proposed variables did have a statistically significant relation with 28-day mortality and ventilator-free days, with the exception of BMI. These observations support the validity of our conceptual model linking measures of acute and chronic starvation and inflammation to outcome. BMI still may be an important determinant to outcome; however, we have very few patients with a low BMI (< 20) in this study. In other databases with larger numbers, BMI may still contribute to risk of adverse events and may be influenced by nutritional intake [12].

We reparameterized each variable and used regression techniques to determine the strength of the association between each derived variable and 28-day outcome. This process served to inform our scoring system. Unfortunately, we had difficulty in obtaining an accurate history of recent decreased oral intake and recent weight loss, a finding also noted in other studies [4, 9, 24], which seriously limits the clinical utility of these measurements. Despite the large amount of missing data, we attempted to include oral intake and recent weight loss by treating missing data both as normal and abnormal. Using this imperfect approach, these variables did not improve the fit of the regression models and were dropped from the score. However, it is possible that these variables would have come into the model had complete information been available. Further work with databases with more complete dietary history information may yield different results. Time in hospital prior to ICU stay was included in the NUTRIC score and can be considered as a marker for recent reduced oral intake as we know that iatrogenic malnutrition occurs commonly in hospitalized patients [25]. The various measures of acute inflammation did not increase the fit of the model by much either. CRP and PCT did not increase the fit of the model and were not included in the final score. IL-6 increased the c-index by only 0.007 points, which is neither clinically nor statistically different from the other scores. It was included in the final score because of the improvement in model fit but given the logistical problems and cost considerations in obtaining IL-6 levels in critically ill patients, we suggest that in settings where IL-6 is not readily available, it could be dropped from the score. There would be no added advantage to substituting CRP or PCT for IL-6 levels as the discriminative ability of the score did not improve with these variables included (data not shown). It should be acknowledged that the overall performance of the final model, as judged by the c-index, was still only 'adequate.' Further work needs to be conducted to enrich the signal and develop better scoring systems that enable us to determine which patients will benefit the most of aggressive nutrition therapy.

In the meantime, we think the NUTRIC score has some merit. In this paper, we have conducted several analyses to validate the derived scoring system and have demonstrated that patients with a higher score have worse clinical outcomes (high mortality and amongst survivors, longer duration of mechanical ventilation). We found that a logistic model with NUTRIC score as the sole continuous independent variable predicted mortality as well as a multivariable logistic model including all of the component variables. This substantial reduction in model degrees of freedom is especially important in small and medium size samples where multivariable models can be unstable and biased. Most importantly, in a subgroup of patients who stayed in ICU more than three days, we have demonstrated that patients with a high NUTRIC score benefit the most from aggressive provision of protein-energy requirements, towards meeting their estimated requirements. On the other hand, patients with a low score may even be harmed by such an approach. This adds strength to our prior observations that patients who receive more protein and energy have a better outcome [12] and further refines the observation by localizing the treatment effect in the sickest critically ill patients. Furthermore, this finding counters the argument that the sickest patients are the most difficult to feed and that is why we observe improve outcome in the better fed patient. In practical terms, the NUTRIC score (or some future derivative of it) may be used to help determine which patients receive supplemental parenteral nutrition or strategies to enhance EN delivery (such as motility agents, small bowel feeding tubes, and aggressive feeding protocols, such as the PEP uP protocol [26]).

The NUTRIC score, or the concepts contained therein, may have utility in the design and interpretation of clinical trials of nutrition therapies in the ICU setting. Studies that include heterogeneous ICU patients, some at high nutritional risk, some at low nutritional risk, are more likely to be negative than those who focus on treating only high-risk patients. Future clinical trials need to use the NUTRIC score or some other measurement tool validated in the ICU setting to describe 'nutrition risk' to enable an adequate interpretation of their findings.

One of the limitations of the NUTRIC score is that it only applies to the provision of macronutrients, protein, and energy. We do not expect this score to identify patients who may benefit more or less from pharmaconutrient supplementation (arginine, glutamine, antioxidants, for example). Likely, other biochemical measure of inflammation, immunity or nutrient values will have to provide clinicians with a sense of who will benefit the most from these specialized nutrients [27]. We further acknowledge that perhaps because the primary purpose of this study was to evaluate a sepsis marker, that standardization of nutrition practices and compliance with nutritional history variables was suboptimal. A major limitation of this dataset is that the nutrition history and intake information was only available in a minority of patients despite attempts to collect this data directly from families. Increased abstraction of these key nutrition variables in future studies may lead to further refinement of the NUTRIC score where these data variables may be included. Finally, it may be a limitation to the current derivation of the NUTRIC score that it is based on 28-day mortality, which was the only mortality data available to use in this dataset. Choosing longer term outcomes, such as 90-day mortality or some measure of functional status at hospital discharge may yield different but important results.

We propose a novel scoring tool, the NUTRIC score, to help discriminate which ICU patients will benefit more (or less) from aggressive protein-energy provision. This scoring tool represents the first nutritional risk assessment tool developed and validated specifically for ICU patients. Whereas most methods of nutritional screening in hospitalized patients are reported to be cumbersome and time-consuming and hence are not routinely done [28], the NUTRIC score is a practical, easy-to-use tool based on variables that are easy to obtain in the critical care setting. The importance of this work rests not just in the statistical analyses presented, but at a conceptual level. We assert that not all ICU patients are the same, that there are some that benefit more (or less) from aggressive protein-energy provision in the critical care setting. Further refinement of this tool or others like it will ensure that the right ICU patient gets the right treatment and has implications for both clinical practice and designing future clinical trials.