Data analysis
Age was analysed as a continuous variable. To account for the skewed distribution of the coronary calcium scoring, CTCS was transformed by taking the natural logarithm of CTCS + 1. All other variables were dichotomous. Oestrogen status was not available in our dataset. Therefore, we assumed women below the age of 50 to be oestrogen positive, women of 50 years and above to be oestrogen negative and all men to be oestrogen neutral. Obesity was considered in the model by Morise (1997) only. We defined obesity as a BMI >27 kg/m2, corresponding to their definition [17].
The extracted sets of clinical variables were analysed with multivariate logistic regression analysis, fitting new regression coefficients. No attempt was made to validate original regression coefficients, as such coefficients were often not reported. CTCS was subsequently included in each of the models. Models without CTCS were compared with corresponding models including CTCS using the likelihood ratio test. The level of significance was defined at a p value less than 0.05.
Diagnostic performance was assessed by calculating the area under the receiver operating characteristic (ROC) curve, the c-index. The c-index is a measure of discrimination and is interpreted as being the probability that a randomly chosen patient with CAD will have a higher predicted probability of disease than a randomly chosen patient without CAD [18]. An area under the ROC curve (AUC) of 0.5 corresponds to a model that provides no diagnostic information, whereas an AUC of 1.0 corresponds to a perfect diagnostic model.
STATA statistical analysis software v10.0 (StataCorp, Texas, USA) was used for logistic regression analysis.
Next, we quantified the effect of adding CTCS to the model on the classification of patients into probability categories of CAD. Four probability categories were defined: <30%, ≥30–50%, ≥50–70% and ≥70%. Reclassification tables were constructed for the Diamond & Forrester model and the Pryor model (see Tables 4 and 5) [19]. We computed the reclassification calibration statistic (RCS) [20] which is equivalent to the Hosmer–Lemeshow statistic, applied to the cross-classified cells of the reclassification table with at least 20 observations. A significant result indicates a lack of fit.
Furthermore, the following reclassification measures were calculated for each model: the overall (correct) percentage of reclassification, the net reclassification improvement (NRI) [21] and the integrated discrimination improvement (IDI) [20]. The NRI is the difference in proportions reclassifying to higher and lower probability categories among cases and non-cases. It is interpreted as the percentage reclassified, adjusted for the reclassification direction. A significant NRI indicates that classification improves when CTCS is included. The IDI compares the difference in the average regression slope of cases and non-cases among the models with and without CTCS. A significant IDI indicates that the new model performs better in discriminating cases and non-cases.
Reclassification computations were executed by using syntax made available by Cook and Ridker [20] in SAS Enterprise Guide v3 (SAS Inc, North Carolina, USA).