THE SYMPTOM INDEX: Have I been lying to you?

Linda Joyce Beck, queen of the desert

Actually, this was on Catalina Island

 

 

Okay, let’s start with some fascinating stuff: math.

Suppose you plan to run a clinical trial to ascertain whether a particular “test” – blood marker, urine chemicals, symptoms, etc – is “useful” in detecting the presence of a medical condition.  You administer the test to a large number of people, and it either comes back “positive” (it says that the subject has the condition) or “negative” (it says he/she does not have the condition).  Of course, in a real world the test will make mistakes.  Let

TP = # of subjects  who tested positive and who do have the condition.  (True Positives)

FP = # of subjects who tested positive but do NOT have the condition.  (False Positives)

TN = # of subjects who tested negative and do not have the condition.  (True Negatives)

FN = # of subjects who tested negative but in fact DO have the condition.  (You can  guess this one)

Pretty obviously, you want your test to have as few FPs and FNs as possible.  Medical researchers calculate several  useful ratios from these statistics, viz

PPV (Positive Predictive Value).  PPV = TP/(TP  + FP)

SE (Sensitivity).  SE = TP/(TP + FN)

SP (Specificity).  SP = TN/(TN + FP)

The meaning of these things is fairly obvious.  SE tells you how likely you are to have the condition if you test positive. SP, on the other hand, tells you how likely you are to be free of the condition if you get a negative test.  Obviously, a good test would have SE and SP values close to 1.  PPV measures how ultimately useful the test will be:   you want PPV to be as close to 1 as possible. 

There:  wasn’t that fun!

I am writing this blog because I have run on some articles that report that the symptom index (SI) for ovarian cancer – which I have been pushing for nearly a year – doesn’t work.   For instance,  the PPV (for a particular study) was only .006 to .011 – meaning roughly that out of every 1000 women who tested positive using the SI, only about 10 would actually have ovarian cancer.  Thus SE and SP would both be near zero.  So: don’t use the test.

Well, nuts to that, say I.  First:  nobody intends that the SI be used unsupported by additional lines of evidence.  For instance, the studies I am familiar with merely use the SI as a trigger for more medical attention: go see your doctor, maybe get an ultrasound, etc.  Second: who cares?  Maybe it would lower overall national medical expenses if women having the symptoms ignored them instead of bothering a doctor.  From my perspective, saving (or prolonging) the lives of those ten women  is worth far more than any costs associated with those  990 unneeded visits to a doc.  Hell, the doctors probably feel good about telling someone they are healthy!  And they get paid for doing it.

One other thing comes forth from these papers: when the SI does work, the cancer probably is fairly far advanced.  What is needed, everyone agrees, is a test that will catch it when its first ugly little cell begins to divide.  That is what my group at the Hutch is working on.