Pulse - April 2021 - 16

MEDICAL LEEWAY

MYVETZONE@GMAIL.COM

Sense and Sensitivity
Christopher A. Lee, DVM, MPH, DACVPM

op quiz. A new virus infects
pets. The disease begins
asymptomatically, but over
a few years, the disease appears
universally fatal. Despite the low
prevalence, multiple sources report
a rising national incidence. Luckily, a new screening test is 99.99%
sensitive and 99.98% specific. The
point-of-care test turns positive for
one cat in a multi-cat household
but negative for the rest. The disease easily transfers directly between
pets. What do you do?
The brag-worthy characteristics of
this diagnostic test make quick but compassionate euthanasia of the positive pet
tempting. Perhaps knowing the incubation period will help determine when to
retest the other cats? Overall, this appears
straightforward.
To put the question another way, when
staring at a positive screening test, and asking yourself, " Which single statistical value
determines your level of trust for this positive test result? " - what is your answer? Is it
a) sensitivity, b) specificity, c) prevalence, or
d) another value?
Researchers love sensitivity and specificity, epidemiologists love incidence and
prevalence, but clinicians romance an
entirely different statistic. I say " romance "
because you should love biostatistics or
at least how to use them. Do you need to
calculate them? No, but knowing when
and where to use specific ones keeps us
true to the Latin words of " primum non
nocere " or " first, do no harm. "
The statistic you need is not sensitivity, specificity, or prevalence, but rather a
combination of all three.
Let's review. When defining sensitivity
and specificity, researchers know the true
disease status of the subjects in advance. If
you take 100 known-virus-positive animals and your test correctly identifies 95

of the infection, then the diagnostic test's
sensitivity calculates to 95%. Remember,
sensitivity is all about the positives. On
the flip side of this biostatistics coin lives
specificity. If the test correctly identifies
95 healthy out of 100 subjects as negative, the specificity is 95%.
Clinical practice challenges us to determine if the patient is positive for a pathogen. We do not know in advance.
How do you differentiate a true positive from a false positive? If the sensitivity is 99.99%, then why not take our
chances? After all, the risk of a false
positive is only one out of ten thousand, right? Without an epidemiologist's
wonderful prevalence data, you cannot
answer this question. Science defines
prevalence as the total number of diseased or infected individuals divided by
the entire population.
For perspective, recent years have witnessed a concerning rise of primary and
secondary syphilis across the United
States. A recent 2017 peak documented
a rate of just under 0.01% (0.0095%). For
simplicity, let's just use 0.01% prevalence.
Knowing this, we can calculate the value
that clinicians love, Predictive Value Positive (PVP). If you alternatively want to
know your level of trust for a negative
result, then look at its companion Predictive Value Negative (PVN).
When you combine a 99.99% test sensitivity, 99.98% specificity, and a disease
prevalence of 0.01%, you may only trust

Let's review. When
defining sensitivity and
specificity, researchers
know the true disease
status of the subjects
in advance.

a positive test in this scenario 33% of the
time! To be wrong one out of ten thousand
times may represent " acceptable risk, " but
one out of three appears feckless at best.
Just because you need to love PVPs
and PVNs, doesn't mean that you need
to calculate them. You may love your
car. Like biostatistics, you need to know
how to responsibly use it, but there is no
requirement to build one from scratch.
As a result, if " popping the biostatistics "
hood sounds painful, skip the next two
paragraphs.
In scientific studies, data permits both
an easy way to calculate PVP or a hard
way - what biostatisticians call " the fun
way! " As data eases into beautiful 2x2
tables, calculating PVP remains straightforward if it includes prevalence data.
The numerator holds True Positive Tests
and the denominator, all positive tests,
True Positives, and False Positives. When
looking at the 2x2 table, we calculated
sensitivity and specificity vertically and
the predictive values horizontally.

Figure 1- The Easy Way
Sometimes the 2x2 table does not bear
prevalence data, such as case-control studies or randomized clinical trials. While the
basic formula remains the same, prevalence data must be added. In these situations, you will need to borrow known
prevalence numbers from other sources.
Calculate the True Positive number by
multiplying Sensitivity with Prevalence.
The False Positive is one minus Specificity times one minus Prevalence. The PVN
requires a similar derivation.

Figure 2 - The Hard Way or the
Biostatician's Fun Way
In our original example, you likely
noticed that the sensitivity was slightly
higher than the specificity. For screening
purposes, higher sensitivity helps. While
their numbers can be identical, as one
gets close to its upper 100% limit, the
other must decline.

APRIL 2021

PULSE

Pulse - April 2021

Table of Contents for the Digital Edition of Pulse - April 2021