Discrimination between acute coronary syndrome cases and controls was assessed in the validation set (n = 450) with receiver operating characteristic curves. Areas under the curves (c-statistics) were compared for the standard risk factors alone (c = 0.77; broken gray line), the RBC-FA model alone (c = 0.85; solid black line), and the combined model (c = 0.88; dashed black line). C-statistics for both models including FAs were significantly greater than the standard model but were not different from each other (Table 3; abbreviations as in Table 1). doi:10.1371/journal.pone.0005444.g002 

Context in source publication

Context 1

... regard to matching or case status) as a training dataset for model building, while the remaining one-third was used later as a validation dataset to estimate prediction capabilities. Although disregarding case- control matching sacrifices power, it does not introduce bias, and since we were developing prediction (as opposed to inference) models, we chose the more conservative approach. The training and validation datasets contained 445 and 223 cases, and 453 and 227 controls, respectively. We evaluated the predictive value of RBC-FA profiles alone, the standard risk factors alone, and then the combination. We also performed a secondary analysis including only those individuals who were not taking statin drugs. We used total cholesterol instead of LDL-C for two reasons. First, since both provide equivalent predictive value in the Framingham Risk calculation [7], they are essentially interchangeable (as would be expected for values with a Spearman correlation of 0.91, p , 0.0001). Secondly, 3% of subjects had triglyceride levels greater than 400 mg/dL (making LDL-C incalculable), and thus using LDL-C would have reduced the number of subjects available for our analysis. Stepwise unconditional multivariable logistic regression was used to develop prediction models with p = 0.01 used to enter and remain in the model. One model was developed using RBC- FAs(FA), another with the 7 standard risk factors (SRF), and another using the FAs selected in the FA model combined with the standard risk factors (SRF + FA). Natural log transformations were used for HDL-C and total-C to improve normality. Robust, nonparametric 95% confidence intervals (CI) of the parameter estimates were obtained using bootstrapping method with 10,000 replicates from the training data set for both FA models. In addition to using the stepwise selected FAs, four pre-specified FA metrics were also tested for their ability to add to the SRF model: the omega-3 index (eicosapentaenoic acid (EPA) + docosahexaenoic acid (DHA)) [21], the n-6:n-3 ratio [22], the total long-chain n-3 FAs (EPA + DHA + docosapentaenoic acid), and the proportion of all long-chain polyunsaturated FAs that were of the n-3 class [23]. For each MLR model a single continuous variable, a risk score, was calculated (equation 1) as the linear combination of the parameter estimates ( b i , i = 0 to p) multiplied by each subject’s FA levels (expressed as a percent of total FAs) or by the standard risk factors ( x , j = 1 to n) as follows: The risk score was then used in the logit function (equation 2) to determine the probability of case status, Pr(case). A Pr(case) . 0.5 was classified as a case, otherwise as a control. Several metrics were examined to compare the performance of the various models using the validation set [9,10,24,25,26]. Discrimination was assessed with the c-statistic (concordance index). Positive likelihood ratios combine in one number the sensitivity and specificity at the cut-point threshold by dividing the proportion of true positives by the proportion of false positives. This statistic indicates how likely it is that a case will have an abnormal test compared to a control. Calibration was examined using the Hosmer-Lemeshow statistic, a goodness-of-fit measure- ment that compares predicted to observed counts of subjects by risk score deciles. Misclassification rates were also determined. The area under the ROC curve (c-statistic) was determined for each model and the difference compared to the SRF alone. The standard error (SE) for the c-statistic was computed as described by Hanley and McNeil [27] taking into account the fact that the areas were correlated since the same patient data were used in each method [28]. Due to matching on age, sex and race, there were no differences in these attributes (Table 1). As expected, classic CHD risk factors were generally more common among cases than controls. Twelve of the 18 FAs differed between groups, with cases having lower levels in 6 and higher levels in the other 6 FAs (Table 1). Odds ratios for the 7 standard risk factors alone, FAs alone and the combination are presented in Table 2. The only factors that were significantly related to ACS case status were HDL-C (OR = 0.56, 95% CI 0.43 to 0.71) and smoking status (OR = 2.86, 95% CI 1.79 to 5.07; age and sex were not predictive because they were matched variables). Stepwise selection identified ten FAs significantly related to ACS case status comprising the final model. Two FAs (eicosadienoic acid and trans oleic acid) were directly related to case status, whereas the other eight were inversely related. On a per-standard deviation basis, the 3 strongest contributors to case status prediction among the latter were linoleic acid, stearic acid, and docosahexaenoic acid. Using the standard risk factors, and the parameter estimates for blood cell FAs, the ability of MLR models to discriminate cases from controls were compared, both alone and in combination (Table 3 and Figure 2). The FA performed better than the SRF, with a c-statistic 8 percentage points higher (p = 0.003). Adding the FA profile to the standard risk factors significantly increased the c- statistic of the latter by 11 percentage points (p , 0.0001), whereas the FA-profile derived c-statistic was not significantly improved by including the standard risk factors (0.85 to 0.88, p = 0.16). Although the 10-FA profile added significantly to the standard model, none of the simpler, pre-defined FA metrics (the omega-3 index, the total n6:n3 ratio, the long-chain n-6:n-3 ratio, and total n-3) added significantly to SRF discrimination (c-statistics were 0.77–0.78 for all, compared to 0.77 to SRF alone). In the ...