The entire data set was segregated into class intervals according to their total cholesterol (TC) results, and alternately, according to their triglycerides (TG) results, and arranged into the six groups viz. AU5800, Alinity, Cobas, AU Diluted, Alinity Diluted and Cobas Diluted. Mean Absolute Percent Variation (MAPV), defined as the average of the absolute values of percentage deviation between C-HDL and D-HDL results, were determined for each class interval. These MAPV values were finally plotted against their corresponding class intervals (Fig. 1). As is evident from the figure, though MAPV increased with increasing TC and TG concentrations for all the series, it increased significantly after 350 mg/dL of TC and after 2000 mg/dL of TG for the AU5800 series. However, when the same specimens were diluted with saline and run on the same platform (AU5800), the MAPV values improved significantly (orange data series on Fig. 1).
While it was evident that the variation in D-HDL results was dependent upon the corresponding TC and TG concentrations, most prominently for the AU5800 platform, it was nevertheless imperative to determine if this variation was related to the dyslipidaemia phenotypes. Hence, in the next level of data analysis, the entire data set was segregated according to the dyslipidaemia phenotypes. For each type, the relationship between D-HDL and C-HDL was examined by PB regression and their corresponding slope and intercept enumerated. Due to the non-parametric nature of the data sets, Spearman’s correlation analysis was done for all the five types and their coefficients (ρ) determined. Regression analysis revealed that variation in D-LDL results indeed depended on the type of dyslipidaemia phenotype. It was, however, important to examine if any of the three platforms was more responsible than the others in contributing to this variation within each dyslipidaemia phenotype. Hence Bland-Altman Plots (BA plots) were constructed with (D-HDL + C-HDL)/2 as the abscissa and (D-HDL – C-HDL) as ordinate for the five dyslipidaemia phenotypes separately. The points were colour coded according to the six data series as described in the ‘Materials and Methods’ section. BA plots comprehensively highlighted the role of each platform in contributing to the variation in D-HDL results in each of the dyslipidaemia phenotypes. However, quantification of the variation caused by each instrument platform within each dyslipidaemia phenotype needed to be enumerated. Herein lay the importance of ROC curves. Using the NCEP (1995) guidelines (14), all D-HDL results with < 22% of variation vis-á-vis C-HDL estimates were designated as ‘Concordant’ and those with variation of 22% and above were designated as ‘Discordant’. ROC curves were constructed for each of the five dyslipidaemia phenotypes with multiple comparator functions for the six data series as described in the ‘Materials and Methods’ section with 1-Specificity (False positive rate) as the abscissa and Sensitivity (True positive rate) as the ordinate. Findings of all the data analysis are summarized in Table 2.
Table 2
| | Type I | Type II | Type III | Type IV | Type V | Total |
N | 23 | 41 | 66 | 101 | 96 | 327 |
Median Age (Years) | 27 | 38 | 41 | 46 | 39 | 43 |
Male:Female Ratio | 1.1 | 1.48 | 1.92 | 1.03 | 0.96 | 1.3 |
Total Cholesterol (mg/dL) | Range | 74–552 | 132–1189 | 297–715 | 88–310 | 140–496 | 74-1189 |
Median | 390 | 390 | 421 | 192 | 266 | 263 |
Triglycerides (mg/dL) | Range | 625–5690 | 53–386 | 100–782 | 223–539 | 642–2003 | 53-5690 |
Median | 2379 | 173 | 327 | 448 | 859 | 543 |
D-HDL (AU5800), in mg/dL | Range | 17–220 | 59–169 | 8-115 | 11–79 | 21–72 | 8-220 |
Median | 35 | 78 | 66.5 | 39 | 40.5 | 45 |
D-HDL (Alinity ci), in mg/dL | Range | 11–32 | 22–130 | 6–97 | 5–71 | 19–52 | 6-130 |
Median | 21 | 53 | 45.5 | 31 | 31.5 | 34 |
D-HDL (Cobas Pure), in mg/dL | Range | 9–35 | 20–163 | 6-107 | 5–81 | 16–57 | 6-163 |
Median | 20 | 57 | 45 | 34 | 32 | 35 |
C-HDL (mg/dL) | Range | 13–44 | 16–130 | 9-103 | 8–82 | 23–60 | 8-130 |
Median | 25 | 62 | 52 | 36 | 36 | 40 |
Passing-Bablok Regression Analysis | Slope | 1.17 | 1.06 | 1.08 | 1.05 | 1.15 | |
Intercept (mg/dL) | -4.83 | -4.79 | -5.30 | -2.76 | -6.97 | |
Spearman Coefficient, ρ | 0.755 | 0.817 | 0.876 | 0.910 | 0.844 | |
Bland Altman Plots | Bias (mg/dL) | 5.34 | 1.99 | 0.21 | -0.98 | -1.17 | |
Limits of Agreement (mg/dL) | ± 60.41 | ± 29.63 | ± 16.69 | ± 9.09 | ± 11.09 | |
Receiver Operating Characteris-tic Curves, Area Under Curve (AUC) | AU5800 | 0.576 | 0.816 | 0.574 | 0.448 | 0.526 | |
Alinity ci | 0.522 | 0.814 | 0.666 | 0.585 | 0.428 | |
Cobas Pure | 0.523 | 0.399 | 0.702 | 0.5 | 0.611 | |
AU Diluted | 0.383 | 0.667 | 0.984 | 0.978 | 0.639 | |
Alinity Diluted | 0.383 | 0.667 | 0.969 | 0.975 | 0.59 | |
Cobas Diluted | 0.381 | 0.667 | 0.984 | 0.98 | 0.872 | |
* The Types denote dyslipidaemia phenotypes as described by Fredrickson et al.
* AU5800 represents results of undiluted samples on AU5800; Alinity represents results of undiluted samples on Alinity ci; Cobas represents results of undiluted samples on Cobas Pure; AU Diluted represents results of diluted samples on AU5800; Alinity Diluted represents results of diluted samples on Alinity ci and Cobas Diluted represents results of diluted samples on Cobas Pure.
* D-HDL: Directly measured HDL-Cholesterol by homogeneous colorimetry.
* C-HDL: HDL-Cholesterol calculated as an average of the readings obtained by the three platforms on diluted specimens.
* To convert Total cholesterol and HDL-Cholesterol to millimoles per litre, multiply by 0.0259; to convert Triglycerides to millimoles per litre, multiply by 0.0113.
For dyslipaemia phenotype Type I specimens, PB regression revealed moderate deviation but wide dispersion of values (Fig. 2). The corresponding BA plot corroborated the deviation and dispersion of values by means of the bias and the limits of agreement (LoA) statistics respectively. Particularly the AU5800 series reported a predominantly positive bias (overestimation of D-HDL) with several outliers above the upper LoA. This positive bias got corrected on dilution of the specimens (Fig. 2, middle panel, orange data). Surprisingly, when it came to the ROC curve, the AU5800 series outperformed the others with the highest area under curve (AUC) of 0.576. Concordance of AU5800 results decreased upon dilution (AUC = 0.383), contrary to the findings of the corresponding BA plot. In fact, the performance of all three platforms deteriorated after dilution. Overall findings of Type I data are confounding, probably due to the small sample size (N = 23).
For dyslipaemia phenotype Type II specimens, PB regression revealed moderate deviation and dispersion of data (Fig. 3). In comparison with Type I, BA plot of Type II shows diminished bias and narrowed LoA. But like in Type I, Type II data also reveals a positive bias for AU5800 results while the bias turns negative for AU5800 results on diluted specimens. Hence, for Type II specimens, AU5800 appears to overestimate D-HDL on undiluted specimens and underestimate D-HDL on diluted specimens. But this variation in D-HDL estimation on AU5800 does not seem to cross the Total Error Goal of 22%, as the platform outperforms the others on the ROC curve (Fig. 3, Right Panel), with an AUC of 0.816. Cobas Pure performs dismally for Type II specimens, with an AUC of 0.399. On dilution, concordance of results on both AU5800 and Alinity ci decreases, while that of Cobas Pure improves.
For dyslipaemia phenotype Type III specimens, PB regression revealed moderate deviation and dispersion of data (Fig. 4). As in Type II, the BA plot of Type III is also characterized by a negligible bias and narrowed down LoA; and like Type II, Type III specimens also produce a positive bias on the AU5800 platform, which gets corrected to some extent upon dilution. However, in contrast to Type II results, AU5800 underperforms in comparison to the other two platforms on ROC curve for Type III specimens with an AUC of only 0.574. Cobas Pure results for Type III specimens are most accurate, producing an AUC of 0.702. Upon saline dilution and retesting, results of all the three platforms improve significantly (AUC ~ 0.98).
For dyslipaemia phenotype Type IV specimens, PB regression revealed minimal deviation and dispersion of data (Fig. 5). Regression data is reiterated on the BA plot statistics with a negligible bias and a narrow LoA. Outliers on the BA plot are minimal, but when they occur, they outlie widely. As in the other types, AU5800 results show a positive bias which get corrected on dilution. The overestimation of D-HDL results in AU5800 platform is high enough to push up the number of discordant results in the corresponding ROC curve, thereby hampering the performance of the platform: AU5800 has an AUC of only 0.448 on the ROC curve. However, for Type IV specimens, the other two platforms also do not fare any better, AUCs hovering around) 0.5. Upon dilution and retesting, performance of all the three platforms improve (AUC ~ 0.98).
For dyslipaemia phenotype Type V specimens, PB regression revealed moderate deviation and dispersion of data (Fig. 6). The corresponding BA plot depicts minimal bias and a narrow LoA. The overall BA plot is compact with few outliers, but again reveals a preponderance of positive bias for the AU5800 series. The bias diminishes after dilution and retesting, but still remains on the positive side. The ROC curve demonstrates only a modest performance for all the three platforms with AUCs between 0.5 and 0.6. The performance of only Cobas Pure improves appreciably (AUC 0.872) after dilution and retesting.