Patients and setting
This study was a secondary analysis of a prospective observational study in three tertiary care, surgical-medical ICUs conducted to evaluate a novel diagnostic marker for sepsis [14]. Patients at least 18 years of age were enrolled within 24 hours of admission to the ICU. Patients admitted for elective surgery, those admitted with overdoses, and patients who were expected to stay less than 24 hours were excluded. Given the nature of this observational study, no attempt was made to standardize care, including nutritional practices, across participating ICUs. The clinical management of patients was determined by the clinical team looking after the patient and the clinical protocols operational in participating ICUs at that time. Patients were started on enteral nutrition within 24 to 48 hours of ICU admission (on average), according to local practice. Feeds were advanced or continued at goal hourly rate if the gastric residuals (checked every four hours) were less than 200 to 250 ml. The local dietitians determined the goal rates using standard formulae. Gastrointestinal prokinetic agents and, eventually, small bowel feeding tubes were prescribed in the event of problems with persistent high gastric residual volumes. When clinically indicated, parenteral nutrition was prescribed by the clinical team. Arterial or venous blood glucose levels were assessed daily in the morning and frequently through the day and a glycemic control protocol was used to prescribe the dose of insulin to titrate blood sugars between 4.0 and 9.0 mmol/L.
This study was approved by the Queen's Research Ethics Board and informed consent was obtained from all participating patients or their substitute decision-makers.
Clinical data collection
All data were collected prospectively. Research coordinators interviewed family members, where available, to obtain historical nutrition variables (recent reduction in intake by mouth (% in last week) and history of weight loss in the past six months). Data on baseline demographics, past medical history including a detailed list of comorbidities and medications were abstracted from patients' charts. Acute Physiology and Chronic Health Evaluation Scores (APACHE II) [15] and Sequential Organ Failure Assessment (SOFA) scores [16] variables were recorded on admission to ICU. Data pertaining to nutrition prescriptions and intake were collected daily until death or discharge from the ICU, or to a maximum of 14 days. Percentage adequacy of nutrition was calculated as energy or protein actually received divided by total energy or protein prescribed. Outcomes were collected until day 28 and these included ventilator free days in 28 days; ICU length of stay, and 28-day mortality.
Laboratory measurements
Upon enrollment and daily thereafter until ICU discharge, death or a maximum of 10 days, morning blood samples were collected. Plasma was analyzed for inflammatory markers using the following assays: C-reactive protein (CRP) by the CRPh reagent Beckman Coulter Unicel DxC 600/800 Synchron Clinical System (Hoffman - La Roche Ltd, Basel, Switzerland), procalcitonin (PCT) using the BRAHMS PCT LIA, B·R·A·H·M·S (Berlin, Germany); IL-6 using the Bender MedSystems ELISA Kit-Cat. BMS213 (Bender MedSystems Inc, Burlingame, CA, USA).
The conceptual model
To develop this nutrition risk score, we first started with a conceptual model that linked starvation, inflammation, nutrition status, and clinical outcomes (Figure 1). Potential variables to represent these constructs were chosen based on their fit with the conceptual model and their ease of routine use. With respect to starvation and inflammation, we considered that there would be two forms, both acute and chronic. We considered recent decreased oral intake [4–9] and pre-ICU stay in hospital [17] as candidate variables for acute starvation and a history of recent weight loss [4–9] (within past three months) and a low BMI (current BMI < 20) [4, 6, 7] as measures for chronic starvation. To represent inflammatory markers, we were limited by the measurements available to us from the original study. We chose PCT, IL-6, and CRP to be representative markers of acute inflammation and the presence of comorbid illnesses to reflect a measure of chronic inflammation. All of the variables selected based on the conceptual model were candidates for the inclusion in the NUTRIC score algorithm. We expected this model to explain additional mortality risk, above and beyond what would be derived from use of traditional measures of severity of illness (APACHE II score and baseline SOFA).
Statistical approach
Our first step was to validate our choice of candidate variables, derived from our conceptual model, by describing their association with 28-day outcomes. Candidate variables were compared between 28-day survivors and non-survivors. Categorical variables were described as counts and percentages and compared by the Chi-Square test whereas continuous variables were described as medians and inter-quartile range (IQR) and compared by the Wilcoxon rank-sum test. The Spearman correlation coefficient was used to assess the association between patient characteristics and candidate variables and ventilator-free days within 28 days.
In our second step, we developed the NUTRIC score using the candidate predictor variables. Percent oral intake in the week prior to enrolment was dichotomized into patients who reported less than 100% versus everyone else including those without this variable reported. This dichotomization did not result in substantial information loss because only 10% of patients report less than 100% but more than 10% oral intake. Similarly, as 76% of patients reported less than 1% weight loss, weight loss was dichotomized as patients who reported any weight loss versus everyone else. A sensitivity analysis was performed where we assumed that weight loss or less than 100% oral intake did occur when not reported. BMI was dichotomized as less than 20 versus other as the data were too sparse to have multiple BMI categories. The number of comorbidities was left as integer values (range 0 to 5). The remaining candidate variables were categorized into five equal sized groups (quintiles). The categorized candidate variables were then each fit as categorical predictors in separate single predictor logistic regression models predicting 28-day mortality. The parameters for each logistic regression model estimate the log of the odds ratio (logit) for each category (usually quintile) of the variable compared with the lowest risk (reference) category. These parameters were rounded to whole numbers to provide the points used in the NUTRIC risk score. Equal point categories were collapsed, and the exact quintile ranges were subsequently rounded to convenient values. The total NUTRIC score was simply the sum of the points across all included variables. Variables with an overall significance of more than 0.2 or with all categories assigned 0 points were excluded from the scoring algorithm. Furthermore, variables were excluded if their inclusion in the NUTRIC score did not improve the score's ability to predict 28-day mortality. The resulting total scores ranged from 0 (lowest risk category for all included variables) to 10 (highest risk category for all included variables). We choose this simple approach to model building over more sophisticated and data-dependent methods because it was intuitive and (given our relatively limited sample size) less susceptible to overfitting and optimism bias [18, 19]. Nevertheless, we did confirm that the multivariable fractional polynomial approach proposed by Saurerbrei and Royston yielded a similar model with no improvement in performance (data not shown) [20, 21].
In step 3 we evaluated the quality of the NUTRIC score model for predicting 28-day mortality. Model discrimination was assessed by the C-statistic derived from calculating the area under the receiving operating characteristic curve (interpretation of c-index: excellent ≥0.90, adequate 0.70 to 0.89, poor < 0.70) and the generalized max-rescaled R-squared statistic [18, 22]. These statistics were also used to compare the discriminative capacity of the NUTRIC score model with logistic model with only age, APACHE II score and baseline SOFA and logistic model excluding any measure of acute inflammation. Model calibration (i.e. goodness of fit) was assessed descriptively by visually comparing the predicted (model based) and actual (observed) mortality rates for each score value, and formally by the Hosmer-Lemeshow goodness of fit test [23]. The performance of the scoring algorithm was cross-validated by independently deriving the NUTRIC score using a random split half of the sample and then evaluating its discriminative ability on the other half sample.
To further validate the NUTRIC score, we examined the association between the NUTRIC score and mechanical ventilation (MV) duration among 28-day survivors. The agreement between observed and model-based estimates of MV duration for each NUTRIC score value was also examined. Our a priori hypothesis was that patients with a higher NUTRIC score would have a longer duration of MV.
In our final step, we examined if the risk score modified the association between nutritional intake and 28-day mortality in a subset of patients who started MV within 48 hours after ICU admission and stayed in ICU for three days or longer (n = 211). A priori, we hypothesized that among patients who remained in ICU more than three days, those with high risk scores would benefit more from more nutritional intake whereas nutrition intake would not be as important in patients with low NUTRIC scores. Nutrition intake was expressed as the total amount of energy received from either enteral nutrition (EN) or parenteral nutrition (PN) over the number of ICU days divided by the amount prescribed as per the baseline assessment and expressed as percentage. Data from the last day in the ICU (unless day 14) were excluded in the calculation of the nutrition intake as these are only partial days on which we would not expect patients to receive their entire prescriptions. The association between nutritional intake and 28-day mortality was plotted by risk score. Logistic regression with nutritional intake, risk score and their product as continuous independent variables was used to generate a plot of the association between nutritional intake and 28-day mortality by risk score and to perform a likelihood ratio test for an interaction (effect modification) between NUTRIC score and nutritional intake among this subgroup of patients. However, for clarity the figure groups risk scores as 0 to 5 and 6 to 10.