Introduction

Medical diagnostics is an information processing endeavor based on probabilities. Two types of medical information can be distinguished, patient-specific and knowledge-based [1]. Diagnostics is about connecting these two types of information [2]. The production of patient-specific information is the main objective of the clinical laboratory. Laboratory tests can significantly contribute to this effort by modifying the probabilities of a diagnosis. Many modern laboratory techniques provide quantitative test results and it would be important to know how much a particular test result would increase or decrease the odds for a specific diagnosis. For example, how much does a D-dimer result of 1000 μg/L, double the recommended cut-off, increase or decrease the clinical suspicion of thrombosis. The answer to that question lies in applying Bayes’ theorem [3–5]: pretest odds multiplied by the likelihood ratio (LR) of the laboratory test result give the posttest odds (S1 Appendix). LRs are defined by the ratio of the probability of the test result in the population carrying the disease versus the probability in the non-diseased population.

The question is how the LR of a measured quantitative test result can be determined.

The analysis of Receiver Operating Characteristics (ROC) of diagnostic tests is common in establishing the merits of laboratory tests and in determination of cut-off values [6]. ROC curves are defined by the relation between the true positive rates (TP, sensitivity) and the false positive rates (FP, 1-specificity) for various cut-offs in dichotomous test interpretation. The area under the curve (AUC) serves to compare the diagnostic value of different tests [7]. The higher the AUC is, the better the diagnostic value of the test. However, ROC curves contain much more information in that for a specific quantitative test result, the LR of this result is equal to the slope of the tangent to the ROC curve at the point on the ROC curve corresponding to the measured test result [8]. Unfortunately, publications of ROC curves of diagnostic tests or test information given by the test producers do not contain this information. Particularly, the quantitative test results underlying the ROC curves are usually not published in detail, so it is impossible to calculate the LR of a particular quantitative test result. In addition, no simple method is currently available to reach this goal.

Knowing the distribution of a test parameter in the diseased and non-diseased population would allow calculation of the slopes of the tangents, i.e. the LRs, with statistical methods. However, for many tests these distributions are not exactly known and differ for various test parameters. Consequently, a distribution-free estimation of the slopes on the ROC curves is required. Here it is demonstrated that the approximation of the empirical points of a ROC curve by a cubic Bézier curve directly leads to the desired slopes of the tangents and thereby to the LRs of a measured quantitative test result. To illustrate the method, some examples of ROC data are used where raw data of the ROC curves have been published, which is rarely the case in the literature.

Methods

Pierre Bézier (1910–1999) was a French engineer who developed a method of producing computer-driven curves to be used in the design of automobiles at Renault, which came to be known as Bézier curves. The algorithm to calculate these curves was developed by mathematician Paul de Casteljau at Citroën. The mathematical basis for Bézier curves are the Bernstein polynomials (for review see [9]). Bernstein polynomials of degree n are defined by (1)

For the purpose here, we make use of cubic Bézier curves defined by (2)

The cubic Bézier curve is determined by the four control points P₀, P₁, P₂, and P₃ (Fig 1). The variable, relative position of the points T₁, T₂, T,₃ T₄, and T₅ between the control points P_(0,1,2,3) is equal to the ratio t. The Bézier curve is given by the tangents defined by T₄ and T₅ for all t from 0 to 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Principle of constructing cubic Bézier curves.

First, the lines between the control points P0, P1, P2, and P3 are divided by the ratio t leading to T1, T2, and T3. Second, the lines between T1, T2, and T3 are again divided by the ratio t leading to T4, and T5. Third, the line between T4, and T5 is again divided by the ratio t leading to B(t) on the Bézier curve. The line between T4, and T5 is the tangent to B(t).

https://doi.org/10.1371/journal.pone.0192420.g001

Given a particular empirical ROC curve, a Bézier curve can be fitted to the points of the ROC curve by adjusting the control points P_(0,1,2,3) with the following least square methods.

Results

The three steps described above are exemplified by their application to a simple example of a ROC curve with raw data available from the literature [12].

The starting data are given in Table 1 and Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. HbA1c test as a tool in the diagnosis of gestational diabetes mellitus.

t values are calculated according to step 1 in methods.

https://doi.org/10.1371/journal.pone.0192420.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. HbA1c test as a tool in the diagnosis of gestational diabetes mellitus.

ROC curve of the original data [12].

https://doi.org/10.1371/journal.pone.0192420.g002

Step 1 and 2

Cubic Bernstein polynomials are fitted to the data points by establishing cubic polynomials for Se(t) and 1-Sp(t) (Fig 3A). The x and y coordinates of the control points P_(0,1,2,3) for the Bézier curve are calculated from the coefficients of the cubic polynomials according to (5). The Bézier curve is constructed as described in Fig 1 and shown in Fig 3B.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Bernstein polynomials (A) for Se and 1-Sp for calculating the control points P_0,1,2,3 of the Bézier curve (B).

Youden indices (Y) with their maximum (Ymax) are indicated. The slope of the tangent to the ROC curve at Ymax equals 1.

https://doi.org/10.1371/journal.pone.0192420.g003

The HbA1c value where the LR = 1 i.e. where the slope of the tangent to the curve equals 1 is 34.5 mmol/mol Hb. This corresponds to the point where the Youden index (Y = Se+Sp-1) [10] reaches its maximum, i.e. where the cut-off is optimal for maximizing the number of correctly classified individuals (Fig 3B).

Step 3

For all known data points the LR(t) are calculated using the formulas in S3 Appendix, and a general relation between Hba1c values and corresponding LRs is established by a least square approximation (Fig 4). In this way, for each quantitative test result the LR can be calculated, independent of the parameter distribution and independent of any cut-offs. At Hba1c = 38 mmol/mol, e.g., the LR reaches 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Calculating LRs from test results with three different methods.

Curve fitting with cubic polynomials of test results of known data points to LRs (1), to λ or to t (3) for calculating the desired LR of a given test result, directly or indirectly from λ or t (A). At Hba1c = 38 mmol/mol, e.g., the LR reaches 2 (B). λ-weighted Youden indices, Y(λ), with their maximum, Y(λ)max, are indicated. The slope of the tangent to the ROC curve at Y(λ)max equals LR.

https://doi.org/10.1371/journal.pone.0192420.g004

A remarkable observation is that when λ-weighted Youden indices [10, 11] are used, as defined by (6) the point on the ROC curve with the corresponding LR = (1 –λ)/λ always is at the maximum of the λ-weighted Youden index. The case of the “optimal” cut-off where the non-weighted Youden index is at its maximum, as shown in Fig 3B, is then just a special case of the more general λ-weighted Youden index, when λ = 0.5, i.e. LR = 1 and Y = Se + Sp − 1.

In the following (Fig 5), some examples from the literatur are calculated by using the above procedure. The source of data are published ROC curves with detailed laboratory test results:

1. A. Prostate-specific antigen in men with an initial PSA level of 3.0 ng/ml or lower (data from [13])
2. B. Fasting capillary blood glucose as a screening test for impared glucose tolerance (data from [14])
3. C. D-dimer testing for suspected pulmonary embolism in outpatients (data from [15])
4. D. Heart-type fatty acid-binding protein in suspected acute myocardial infarction (data from [16] see supporting information S1 File)

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Examples of Bézier curve approximation and likelihood ratio calculation.

https://doi.org/10.1371/journal.pone.0192420.g005

Discussion

In order to calculate LR values for quantitative test results, a distribution-free algorithm based on Bézier curves is proposed. The necessary calculations can easily be done, e.g. by using a Microsoft Excel tool. The advantage of the method is that it is generally applicable, independently and without knowledge of the test parameter distribution in the population. The accuracy of the method is independent on the clinical situation but only depends on the accuracy of the empirical ROC points. The clinical examples chosen here to demonstrate the method are selected on ground of detailed ROC data availability in the literature.

Bézier curves are mathematically well defined and widely used in computer graphics. Here, we make use of cubic Bézier curves defined by Bernstein polynomials of degree 3. Approximation of the cubic Bernstein polynomials B(t) to empirical points on the ROC curve is done by a least square method for the B_x(t) and B_y(t) coordinates separately, t being a variable between 0 and 1. The crucial advantage of this procedure is that Bézier curves are constructed by tangents to the curve, whose slopes immediately provide the LR of a specific point on the curve. It remains to relate all quantitative results of a test to positions on the Bézier ROC curve and thereby to their LRs. Three methods to do so are proposed.

The example question posed in the introduction as to how much a D-dimer result of 1000 ug/L increases or decreases the clinical suspicion of thrombosis can be answered after this analysis with LR = 0.7, i.e. the result with LR being smaller than 1 lowers the initial clinical suspicion of thrombosis, although 1000 ug/L is double the cut-off of 500 ug/L. A LR = 1 is only reached with 1300 ug/L. In fact, it is a frequently observed mistake in test interpretation that results that lie closely to the point on the ROC curve where LR = 1 are wrongly interpreted or overestimated in their diagnostic significance.

The merit of using LRs in addition to or even instead of quantities like mg/L or nmol/L as test results lies in the comparability of tests that are using different methods and that are produced by different test suppliers, which is an unsolved problem of standardization in laboratory medicine, particularly in immune serology [17]. Of course, LRs are always related to a specific diagnosis and ROC curves must be established for each diagnosis separately. This is reasonable, since it is good clinical practice to base the choice of a laboratory test and its corresponding ROC curve on a tentative diagnosis. However, this requirement requests that the test producers cannot just calibrate their products by comparing them to other products, but have to do clinical studies with the test.

In conclusion, ROC curves of diagnostic tests that are approximated by Bézier curves provide likelihood ratios for quantitative test results, independent on test methods. These likelihood ratios allow to estimate the probabilities of diagnosis based on pretest probabilities according to Bayes’ theorem. Such inferences based on quantitative test results have otherwise not been possible so far.

Supporting information

Acknowledgments

The author would like to extend many thanks to Brigitte Walz for providing hFABP data (Fig 5D) [16] and Pietro Vernazza and Xavier Bossuyt for helpful discussions and suggestions in review of this manuscript. Leslie Bisset’s professional editorial help is also thankfully appreciated.