Figure 3.

Information graphs for comparing rule-in and rule-out test potentials for predicting a low and high risk of prediabetes post-GDM

Information graphs provide means to distinguish between diagnostic test performance. We compared the diagnostic information obtained from T_out, T_in−out, and T_in defined by the cut-points 0.140, 0.260, 0.381. A positive diagnosis made by the ‘rule-in-specific-test’ and a negative diagnosis made by the ‘rule-out-sensitive-test’ gives us the most information, as expected.

(A–C) Maximum information from a positive test diagnosis (blue) is obtained at a lower pre-test probability than the maximum information from a negative test diagnosis (red). The diagnostic test with a lower cut-point gives maximum information when the diagnosis is negative (i.e., the test is very sensitive and we can rule out the negative cases safely) and the diagnostic test with a higher cut-point gives maximum information when the diagnosis is positive (i.e., the test is very specific to the disease and we can rule in the positive cases safely). I_E is the expected information from the diagnostic test (x × I₊ + (1 − x) × I₋, where x is the probability of a positive test diagnosis).

(D) The sum of the distances between the tangents to the negative Shannon entropy function at p = g₁(c) and p = 1 − g₂(c) is the discrete Bregman divergence, which represents total K-L divergence.