31.10.2014 Views

The math of decision in Radiology - MIR

The math of decision in Radiology - MIR

The math of decision in Radiology - MIR

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

THE MATH OF DECISION IN RADIOLOGY<br />

Pr<strong>of</strong>. Utku Şenol, MD, PhD <strong>of</strong> medical <strong>in</strong>formatics<br />

Akdeniz University, School <strong>of</strong> Medic<strong>in</strong>e,,<br />

Department <strong>of</strong> <strong>Radiology</strong>, Division <strong>of</strong> Neuroradiology<br />

ANTALYA, TURKEY<br />

utkusenol@gmail.com


Contents<br />

• Decision mak<strong>in</strong>g <strong>in</strong> radiology<br />

• Test characteristics<br />

• Likelihood ratios<br />

• Predictive values<br />

• Calculat<strong>in</strong>g post-test probability<br />

• Bayesian approach<br />

• Probability revision<br />

• Nomograms, Probability graphs


Question 1<br />

A lymphadenopathy <strong>in</strong> mediast<strong>in</strong>um was reported<br />

Sensitivity 90%<br />

Specificity 80%<br />

What is the actual probability <strong>of</strong> the lymphadenopathy?<br />

A) 2-3%<br />

B) 20-25%<br />

C) 50-60%<br />

D) 80-90%<br />

E) Can not be calculated


Question 2<br />

• A Breast cancer screen<strong>in</strong>g program<br />

• Prevalence <strong>of</strong> the Breast Ca is 0.1%<br />

• A positive mammogram was reported.<br />

sensitivity 95%<br />

specificity 95%<br />

What is the actual probability <strong>of</strong> the patient hav<strong>in</strong>g a cancer?<br />

A) 2%<br />

B) 10-20%<br />

C) 50-60%<br />

D) 90-95%<br />

E) Can not be calculated


When Do Radiologists Make Decision ?<br />

• While<br />

– Diagnos<strong>in</strong>g<br />

– Exclud<strong>in</strong>g disease<br />

– Giv<strong>in</strong>g probability for a certa<strong>in</strong> disease<br />

– Choos<strong>in</strong>g the proper exam<br />

– Plann<strong>in</strong>g a survey program<br />

– …


How Do Radiologists Make Decision?<br />

Do <strong>The</strong>y Use Probability or Math?


• Virtually always<br />

• Often<br />

• Not <strong>of</strong>ten<br />

• Virtually never<br />

AJR 174:2000 letter from Kliewer


AJR 174:2000 letter from Kliewer


Sir William Osler<br />

1849 -1919<br />

“Medic<strong>in</strong>e is a science <strong>of</strong><br />

uncerta<strong>in</strong>ty and an art<br />

<strong>of</strong> probability”


Do not treat<br />

Further tests<br />

or follow-up<br />

Treat<br />

0%<br />

Very unlikely<br />

to have disease<br />

25%<br />

Probably doesn’t<br />

have disease<br />

50%<br />

Don’t know<br />

75%<br />

Probably does<br />

have<br />

100 %<br />

Very likely<br />

to have disease


Probability <strong>in</strong> Orders<br />

• Thunderclap headache<br />

• Lower quadrant pa<strong>in</strong>, leukocytosis,<br />

womit<strong>in</strong>g, ribaund<br />

• Lung Ca, recent vertigo, Rull out bra<strong>in</strong><br />

metastasis<br />

• 35 year age woman, pa<strong>in</strong> and visual loss <strong>in</strong><br />

right eye and hemiparesis


Probability <strong>in</strong> Reports<br />

• Hemorrhagic density with sulcal pattern<br />

• Well circumscribed lytic lesion with a<br />

sclerotic rim <strong>in</strong> distal femur<br />

• An ill def<strong>in</strong>ed solitary nodule <strong>in</strong> right apex


Subjective probability <strong>of</strong> estimates are not statistical<br />

and are highly prone to error!!<br />

Those errors may be as;<br />

• Pesudodiagnosticity<br />

• Premature diagnosis<br />

• Errors Caused by Heuristics<br />

– Availability<br />

– Representativeness<br />

– Anchor<strong>in</strong>g and adjustment<br />

– Value-<strong>in</strong>duced bias<br />

– Affect heuristics<br />

– …..


How do we use tests to def<strong>in</strong>e<br />

probability?


Diagnostic tests are not perfect.<br />

False positive and false negative results cannot be elim<strong>in</strong>ated…<br />

Sorry, you can not have baby<br />

But, We had<br />

Tests say<br />

you can’t.<br />

Give<br />

them to<br />

me…


Sensitivity and Specificity<br />

Sensitivity (TPR) = a/a+c<br />

Specificity (TNR) = d/b+d<br />

Test +<br />

Disease<br />

+<br />

a<br />

Disease<br />

-<br />

b<br />

Total<br />

a+b<br />

TP<br />

FP<br />

Test -<br />

c<br />

d<br />

c+d<br />

FN<br />

TN<br />

Total<br />

a+c<br />

b+d<br />

a+b+c+d<br />

TP+FN<br />

TN+FP


Which Test is the Strongest?<br />

Test Sensitivity Specificity<br />

A 80% 75%<br />

B 60% 85%<br />

C 70% 80%<br />

D 78% 78%


Likelihood Ratios<br />

A unique value generated by us<strong>in</strong>g both<br />

specificity and sensitivity<br />

A perfect <strong>in</strong>dicator for def<strong>in</strong><strong>in</strong>g strength <strong>of</strong> tests


Positive Likelihood Ratio<br />

LR+ = sensitivity/(1-specificity)<br />

Sensitivity = 90%<br />

Specificity = 95%<br />

LR+ = .90/.05 = 18 =<br />

<strong>The</strong> probability <strong>of</strong> a positive result <strong>in</strong> a patient with disease<br />

<strong>The</strong> probability <strong>of</strong> a positive result <strong>in</strong> a patient without disease


Positive Likelihood Ratio<br />

• 10 or<br />

strong<br />

• 5-10<br />

medium<br />

• 2-5<br />

weak<br />

• 2 or<br />

too weak


Test Sensitivity Specificity<br />

A 80% 75%<br />

B %65 %85<br />

C %70 %80<br />

D %78 %78


Test Sensitivity Specificity LR+<br />

A 80% 75%<br />

3.2<br />

B %65 %85<br />

4.3<br />

C %70 %80<br />

3.5<br />

D %78 %78<br />

3.54


Test Sensitivity Specificity LR+ LR-<br />

A 80% 75%<br />

3.2 0.26<br />

B %65 %85<br />

4.3 0.41<br />

C %70 %80<br />

3.5 0.37<br />

D %78 %78<br />

3.54 0.28


Negative Likelihood Ratio<br />

LR- = (1- sensitivity) / specificity<br />

Sensitivity = %90<br />

Specificity = %95<br />

LR- = 0.10 / 0.95 = 0.105


Negative Likelihood Ratio<br />

• 0.1 ve<br />

strong<br />

• 0.1 - 0.2<br />

medium<br />

• 0.2 - 0.5<br />

weak<br />

• 0.5 ve<br />

too weak


Likelihood Ratios<br />

Strong<br />

medium<br />

medium<br />

Strong


LR helps to predict the post-test<br />

probability


What about Predictive Values?<br />

Positive predictive value<br />

• Probability <strong>of</strong><br />

hav<strong>in</strong>g disease<br />

when the test is<br />

positive<br />

Negative predictive value<br />

• Probability <strong>of</strong> not<br />

hav<strong>in</strong>g disease<br />

when the test is<br />

negative


Disease<br />

Disease<br />

Total<br />

+<br />

-<br />

Test +<br />

a<br />

b<br />

a+b<br />

TP<br />

FP<br />

TP+FP<br />

Test -<br />

c<br />

d<br />

c+d<br />

FN<br />

TN<br />

FN+TN<br />

Total<br />

a+c<br />

b+d<br />

a+b+c+d<br />

TP+FN<br />

TN+FP<br />

Positive Predictive Value: a/a+b<br />

Negative Predictive Value: d/c+d<br />

Sensitivity:<br />

Specificity:<br />

a/a+c<br />

d/b+d


Test Characteristics<br />

(sensitivity, specificity)<br />

Predictive values<br />

• May be<br />

generalized to<br />

different<br />

populations<br />

• Can not be<br />

generalized to<br />

other groups<br />

Unique for a<br />

specific study<br />

group


D+ D - Total<br />

T + 63 18 81<br />

T - 37 82 119<br />

Sensitivity: 63%<br />

Specificity: 82%<br />

PPV: 77%<br />

Toplam 100 100 200<br />

D + D - Total<br />

T + 63 162 225<br />

T - 37 738 775<br />

Sensitivity: 63%<br />

Specificity: 82%<br />

PPV: 28%<br />

Total 100 900 1000


How do we calculate the<br />

probability after the test?


Probability revision<br />

<strong>The</strong> probability after a positive result<br />

Pre test probability<br />

<strong>The</strong> probability after a negative result


Post-test probability can be<br />

calculated only if there is a pre-test<br />

probability


Post-test probability calculation<br />

methods<br />

• Bayes formulation<br />

• Decision tree<br />

• Cont<strong>in</strong>gency table revision<br />

• Nomogram<br />

• Conditional probability graphs<br />

Post-test Odds = Pre-test Odds x LR


Post-test Odds = Pre-test Odds x LR<br />

Post-test probability<br />

Pre-test probability<br />

1- post-test probability = 1- pre-test probability<br />

x LR


Nomogram<br />

Pre-t Prob<br />

LR<br />

Post-t Prob.


Post-test probability<br />

Probability graph:<br />

1<br />

.9<br />

Test +<br />

.8<br />

.7<br />

.6<br />

.5<br />

.4<br />

.3<br />

.2<br />

.1<br />

Test -<br />

0 .1 .2 .3 .4<br />

Pre-test probability<br />

.5 .6 .7 .8 .9 1


Question 1<br />

LR+=4.5<br />

Pre-test probability= ???<br />

A CT report <strong>in</strong>dicates a lymphadenopathy <strong>in</strong> mediast<strong>in</strong>um.<br />

Sensitivity 90%<br />

Specificity 80%<br />

What is really the probability <strong>of</strong> the lymphadenopathy?<br />

A) 2-3%<br />

B) 20-25%<br />

C) 50-60%<br />

D) 80-90%<br />

E) Can not be calculated


Question 2<br />

LR+=19<br />

Pre-test probability=1/1000<br />

• Mammography screen<strong>in</strong>g program<br />

• Prevalence <strong>of</strong> the Breast Ca is 0.1%<br />

• A positive mammogram was reported.<br />

sensitivity 95%<br />

specificity 95%<br />

What is really the probability <strong>of</strong> the patient hav<strong>in</strong>g a cancer?<br />

A) 2%<br />

B) 10-20%<br />

C) 50-60%<br />

D) 90-95%<br />

E) Can not be calculated<br />

%1.93


Summary - I<br />

• Radiologists frequently deal with probabilities<br />

while mak<strong>in</strong>g <strong>decision</strong><br />

• To be familiar with the <strong>math</strong> <strong>of</strong> the <strong>decision</strong><br />

mak<strong>in</strong>g process is helpful


Summary - II<br />

• Likelihood ratios are used to def<strong>in</strong>e;<br />

– strength <strong>of</strong> test<br />

– post test probability


Summary - III<br />

• Likelihood ratios are calculated by sensitivity<br />

and specificity<br />

LR + = sensitivity/(1-specificity)<br />

LR - = (1-sensitivity)/specificity


Summary - IV<br />

• Post test probability can not be calculated without<br />

pre-test probability.<br />

• So, cl<strong>in</strong>ical <strong>in</strong>formation is crucial !!


Conclusion-IV<br />

Post test probability can be calculated based on<br />

Bayes pr<strong>in</strong>ciples:<br />

Post-test Odds = Pre-test Odds X LR

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!