Post on 01-Apr-2015
transcript
Diagnostic research
Lecture Contents
I. Diagnostics in practice
- Explained with a case
II. Scientific diagnostic research– Design– Data-analysis– Reporting
III. Exercises
IV. Summary
Diagnostics in practice
Diagnostics always start with a patient with a complaint/symptom
Case: neck stiffness • Child, 2 years-old, comes to ER with parents • Child turns out to have a very stiff neck
What is the physician’s aim?
Diagnostics in practice
Aim of the physician• Quickly and efficiently determine the correct
diagnosis
Why diagnose?• Basis medical handling • Determines treatment choice• Gives information about prognosis
What are possible diagnoses for neck stiffness?
Diagnostics in practice
Differential diagnosis (DD)• Bacterial meningitis • Viral meningitis• Pneumonia• ENT infection• Other (e.g. myalgia)
What is the most important diagnosis? Which one does the physician not want to miss?
Diagnostics in practice
Most important diagnosis• Bacterial meningitis (BM) • If missed: often fatal
Diagnostics in practice
Suppose: 20% of all children on the ER with neck stiffness has BM – 20% with disease in that population =
prevalence– Prior-probability
What is your decision for the child in this case?
Diagnostics in practice
Decision for child in case • Prior-probability too low to treat• Prior-probability too high to send home
Decision: reduce uncertainty diagnostics
What is the best test?
Diagnostics in practice
Best test
Lumbal punction (liquor culture)
Diagnostics in practice
Gold standard• True disease status; ‘truth’
– Never 24 karat• Reference standard/test• Decisive test with doubt
Perform reference test for everybody (=every child on ER with neck stiffness)?
Reference test for everybody?• Unethical too invasive/risky• Inefficient too expensive• Do not perform unnecessarily
How should we then determine the probability of disease presence and what would be ideal?
Diagnostics in practice
How then?• Simpler diagnostics:
– Usually anamnesis, physical exam, simple lab tests, imaging, etc.
– Ideal: diagnosis without reference test
• Diagnostic process in practice: – Stepwise process: less more invasive– Not one diagnosis based on 1 test – Each item: separate test
Diagnostics in practice
Diagnostics in practice
Suppose: after anamnesis & PE 10% probability of BM
• Probability of disease given test results = posterior-probability
• The bigger the difference between prior and posterior probability, the better the diagnostic value of the tests
Our decision for child in case: probability is too high to send home --> next step?
Diagnostics in practice
Next step– Additional research, e.g.
blood tests (leucocytes,
CRP, sedimentation, etc.)
Diagnostics in practice
Suppose: 1% posterior-probability after anamnesis, PE+ simple lab tests posterior probability low enough to send home
• Ideal diagnostic process: simple tests reduce posterior probability to 0 or 100% (without reference)• Most often physician continues testing until sufficiently sure (approximation of 0 or 100%)• Choose when sufficiently sure: depends on prognosis of disease if untreated + risks/costs treatment
Diagnostics in practice
Summarizing • What does diagnosing involve in practice?
– Estimation of probability of disease presence based on test results of the patient
When is the probability of disease best estimated? Why is this usually not done?
Diagnostics in practice
Why not all possible tests?– Invasive (for patient and budget)– Unnecessary: different test results give same info– However: In practice often more tested than
necessary!
What diagnostics truly necessary scientific diagnostic research
BREAK
Study design
Scientific diagnostic research– What tests truly contribute to probability estimation?– Has to serve practice follow practice
Study design
• Research question• Domain• Study population • Determinant(s): test(s) to study• Endpoint: presence/absence disease (outcome)• Study design: design• Data analysis, interpretation + reporting
Research question
• With as few as possible simple, safe, and cheap tests estimate the probability of the presence/absence of disease.
• Determinant-outcome relation:– probability of disease as a function of test results– outcome = probability of disease = % = prevalence– test results = determinants
Research question
Case• What tests contribute to probability estimation of
presence or absence of BM in children with neck stiffness at the ER?
• Or: Determinants of presence/absence disease (BM)?
• %BM = ƒ(age, gender, fever, blood leucocytes, blood CRP, etc)
Research population
Case: • All children with neck stiffness
in 2002 at ER Utrecht
Domain
• For whom domain, generalisation = type of patient with certain symptom/complaint +
setting• Research population = 1 sample from domain
Case: All children (e.g. in Western world) suspected of disease (BM) based on neck stiffness (characteristic) in secondary care (setting)
Determinants
= Tests to study
• Diagnostic determinants• All possible important tests (in domain)
CaseItems anamnesis, PE and lab (blood and urine)
tests
Endpoint
‘True’ presence/absence disease = Diagnostic outcome = Results reference test
NB: reference = not infallible but always best available test in practice at that moment
Case• Positive liquor culture
PICO EBM
• Population/ problem• Intervention• Comparison/ control• Outcome
• Domain• Determinant• Reference test• Outcome
Measure determinants/endpoint
• Determinants– Without knowledge (blinded) of the outcome – Same method in study and practice
never measure more precisely than in practice (overestimation information yield)
• Endpoint– Assessment blind for determinants– With the best possible test known in practice
Study design
• Observational and descriptive – Observational = no manipulation of determinants– Descriptive = not causal
– if the determinant only predicts– no hypothesis functional mechanism determinant-
outcome
• >1 determinant
Study design
• Cross-sectional= Simultaneously measure determinants and outcome
Data-analysis
After data collection, per patient– Value determinants (test results)– Diagnostic outcome (reference test)
Data-analysis• Data analysis: 3 steps
1) Estimation a priori probability (without test results)
2) Compare each test result separately with reference = univariate3) Compare combination of test results with reference = multivariate (via model)
- Following order in practice - Determine added value test result to already collected (previous) test results
Data-analysis
CaseData scientific research available:200 patients with neck stiffness at ER
Liquor culture positive (BM+) n=40Liquor culture negative (BM-) n=160
Step 1: A priori probability (prevalence) of BM?
= % BM+ = 40/ 200 patients = 20%
Data-analysis reading 2 by 2 table
Disease
Presence Absence
Test Positive True positiveA
False positiveB
Negative CFalse negative
DTrue negative
• Step 2: Analysis per determinant (univariate) • Use 2 by 2 table
Data-analysis reading 2 by 2 table
Horizontally• Positive predictive value (PV+)
= probability Disease + if test + PV+ = A / A + B
• Negative predictive value (PV-)= probability disease - if test -
PV- = D / C + DVertically• Sensitivity (SE) = probability test + if disease +
SE = A / A + C• Specificity (SP) = probability test - if disease -
SP = D / B + D
What numbers do you think are most useful in practice (PV+ and PV- or SE and SP)?
TP A
FN C
B FP
Gold standardDisease + Disease –
Test +
Test –D TN
Data-analysis
Perfect diagnostic testFalse Positive = 0 False Negative = 0
e.g. Fever > 380C as predictor for BM
20
40 160
70 90
200
20 90 110
BM+BM- tot.
Yes (+)Fever > 380C
No (-)
Data-analysis reading 2 by 2 table
Horizontally• probability BM+ if fever+ = 20/110 = 18%
PV+ = A / A + B• probability BM - if fever- = 70/90 = 78%
PV- = D / C + DVertically• probability fever+ if BM+ = 20/40 = 50%
SE = A / A + C• probability fever- if BM- = 70/160 = 44%
SP = D / B + D
What numbers do you think are most useful in practice (PV+ and PV- or SE and SP)?
20
TP A
FN C
20
90
B FP
Gold standardBM+ BM–
Fever +
Fever –D TN
70
BREAK
Exercise 1
Mercury thermometer or timpanic membrane infrared meter
Exercise 1
Ad question 1
Research question: Can fever be determined with the TIM?
Determinant: test under study = timpanic membrane infrared meter
Outcome: fever determined with rectal mercury thermometer (RMT)
Domain: Children in secondary/tertiary care (ER hospital)
Exercise 1
Ad question 2
77
TP A
FN C
19
9
B FP TIM >38°
TIM 38°D TN
108
GS RMTFever+ Fever–
Se = probability TIM+ if RMT+ = 77/96 = 80 %
SP = probability TIM- if RMT- = 108/117 = 92%
Exercise 1
Ad question 3
77
TP A
FN C
19
9
B FP TIM >38°
TIM 38°D TN
108
GS RMTFever+ Fever–
PV+ = probability RMT+ if TIM+ = 77/86 = 90 %
PV- = probability RMT- if TIM- = 108/127 = 85%
Exercise 1
Ad question 4– The prior probability of fever in the general practice is
lower, e.g. 20% (X/213=0,2 X=43)– For similar Se and SP:
(A/43=0,8 A=34)
(D/170=0,92 D=156)– PV+ becomes lower (34/48 = 70%)– PV – becomes higher (156/164 = 95%) 9
43 170
156 164
213
34 14 48 TIM+
TIM-
GS RMTFever+ Fever–
Exercise 1
Ad question 5
– In the general practice an unjustly referred or treated child is less of a problem than an unjust reassurance of the parents
– Especially the negative predictive value must therefore be sufficiently high
Data-analysis: combination of determinants
• In practice not one single diagnosis based on 1 test
– Tests together distinguish ill/non-ill– Method: statistical model
• Moreover: diagnostic process is hierarchical– (simple --> invasive/expensive) --> always start with
anamnesis model --> see case
Data-analysisCase: model with all anamnestic tests gender + age + fever + pain
%BM = ƒ(gender, age, fever, pain)
• Statistical model can be seen as 1 (composed) test
• Quantify diagnostic value model with area under ROC curve (Receiver Operating Characteristic =Area Under Curve (AUC))
Data-analysis
ROC Curve
1 - Specificity
1,00,75,50,250,00
Se
nsi
tivity
1,00
,75
,50
,25
0,00
Data-analysis
Case: AUC anamnesis model = 0,71
Informal interpretation AUC = % correctly diagnosed
The larger the ROC area the better the model AUC range: 0,5 1,0
AUC = 0,5 bad (Se = 1- Sp diagonal [coin])AUC > 0,7 reasonableAUC > 0,8 goodAUC > 0,9 excellent AUC = 1,0 perfect (Se=100% & 1-Sp=0%)
Data-analysis
Quantify added value additional tests to previous tests
• Extend previous model (follow order practice)• Quantify change in AUC
CaseModel 1 anamnesis model + physical exam (5 extra tests) -->
AUC = 0,72 interpretation?Model 2 anamnesis model + 3 blood tests ---> AUC = 0,90
interpretation?
Data-analysis
ROC Curve
1 - Specificity
1,00,75,50,250,00
Se
nsi
tivity
1,00
,75
,50
,25
0,00
Coin flip
Patient hisotry
Pat hist + test
Data-analysis
• The AUC does not directly say anything about individual patients and is therefore not directly applicable
Reporting
Research question
Study set-up• Research population, setting, determinants, outcome,
design
Results• Predictive values (new) test and/or ROC curve• ROC curve combination of tests• Added value new test --> ROC curve
Ad question 1- Cross-sectional study in patients suspected of a
stomach or duodenum ulcer
- For all patients anamnestic data were collected
- For all patients a gastroscopy was done
- Independent diagnostic value of anamnestic factors (determinant) for the diagnosis of ulcer (outcome: gastroscopy) were calculated
Exercise 2
Exercise 2
Ad question 2
Adults with stomach complaints referred to a gastro-enterology policlinic in a peripheral hospital
Exercise 2
Ad question 3
Score is 5, risk is 57%
Exercise 2
Ad question 4- Everybody above the cut-off point has the same risk
(and the same below the cut-off point)
- Of course this is not true and the score loses precision
- Preferably predictive values for score-categories and predictive values for more cut-off points
Exercise 2
Ad question 5
20
TP A
FN C
5
11
B FP Test +
Test -D TN
64
Peptic ulcus+ –
PV+ = 20/31 = 65%
PV- = 64/69 = 93%
Exercise 2
Ad question 6- Predictive values more favourable and therefore
preferred
- But it is not about the isolated predictive value but about the added diagnostic value given the results of the anamnestic score
Exercise 2
Ad question 7• Perform the anamnestic score and the breath test for a
population from the domain. Subsequently perform the reference test (endoscopy) for everybody
• Compare the next determinant-outcome relations: • P(ulcus) = ƒ (age, gender, anamnesis, ...)• P(ulcus) = ƒ (age, gender, anamnesis, ..., breath test)• Then compare the Receiver Operating Characteristic
(ROC)-curve of the models
Exercise 2
Ad question 8
- Breath test partially contains the same information as the score
- Suppose that the breath test is more often positive with age, then the breath test also measures age and therefore the added value is less than when the breath test would be completely independent of the score
Exercise 2
Ad question 9
- Preferably not, but if the assessor is informed of the data in the score in practice, than it should be the same in the study
Diagnostics Summary (1)
Diagnostics in practice– Uncertainty reduction– Determines prognosis & determines policy
Diagnostic research
Design– Observational– Descriptive
– Cross-sectional• Simultaneous measurement determinant
and outcome (reference standard) – Always study >1 determinant
Design– Assess determinants as in practice
– Assess disease status & determinant status with double blinding
Diagnostics Summary (2)
Analysis– Univariate (per determinant)– Multivariate: combination of test results in relation to
outcome • Endpoint = ƒ(combination of determinants)• Determine added value; first analyse least
invasive tests (as in practice)
Reporting– Mainly added value of test
Diagnostics Summary (3)