Laboratory Work Preparation: Lab Work 9: Binary Classification

LABORATORY WORK PREPARATION
Lab Work 9: Binary classification

Biomedical Informatics
Assoc. Prof. Tomaž Vrtovec, Ph.D.
University of Ljubljana, Faculty of Electrical Engineering Electrical Engineering, level 2

Laboratory of Imaging Technologies International course
Lab Work 9: Binary classification 2 / 18
BINARY CLASSIFICATION
What is binary classification?
Binary or binomial classification is the task of assigning samples into two classes
according to a decision. In medicine, such binary decisions are most frequent, as
we commonly want to know whether the observed subject is diseased or healthy.
Actual condition: Test results:
- positive samples – P - positive results – RP
Number of diseased subjects. Number of positive subjects.
- negative samples – N - negative results – RN
Number of healthy subjects. Number of negative subjects.
University of Ljubljana, Faculty of Electrical Engineering BIOMEDICAL INFORMATICS Electrical Engineering, level 2
Laboratory of Imaging Technologies Assoc. Prof. Tomaž Vrtovec, Ph.D. International course
Classification of the results
The samples can be, according to the actual condition and the test results,
classified into four groups:
- true positives – TP - true negatives – TN
The number of diseased subjects The number of healthy subjects
(P) that are correctly classified (N) that are correctly classified
as diseased (RP). as healthy (RN).
Classification of the results (2)
- false positives – FP - false negatives – FN

The number of healthy subjects The number of diseased subjects
(N) that are wrongly classified as (P) that are wrongly classified as
diseased (RP). healthy (RN).
Contingency table
The contingency table represents a condensed description of sample

classification according to the actual condition and the test results.
Positive test Negative test

(RP) (RN)
Diseased
TP FN P = TP + FN
subjects (P)
Healthy
FP TN N = FP + TN
subjects (N)
P+N=
RP = TP + FP RN = FN + TN
= RP + RN
Ideal classification
In the case of an ideal classification, there are no false

positives and false negatives (FP = FN = 0), as all
positive test results actually correspond to positive
samples (RP = P = TP) and all negative test results
actually correspond to negative samples (RN = N = TN).
threshold t threshold t
RP RN RP RN
diseased (P)
Subjects (P, N)
TP TP TN
result
result
healthy (N)
TN
Real classification
In the case of a real classification, false positives and

false negatives exist (FP ≠ 0, FN ≠ 0) mostly due to:
- the influence of the biological variability of samples
- the influence of the variability of reference data
- the influence of the variability of the test
threshold t threshold t
RP RN RP RN
diseased (P)
Subjects (P, N)
TP TP TN
FN FP FN
result
result
FP
healthy (N)
TN
Performance measures
The measures of consistency are:

- true positive rate (TPR) - true negative rate (TNR)
TP TP TN TN
TPR = = TNR = =
P TP + FN N TN + FP
Sensitivity is the probability that Specificity is the probability that
when testing a diseased subject (P) when testing a healthy subject (N) the
the result will be positive (RP). result will be negative (RN).
Performance measures (2)
The measures of inconsistency are:

- false positive rate (FPR) - false negative rate (FNR)
FP FP FN FN
FPR = = FNR = =
N TN + FP P TP + FN
Non-specificity is the probability that Non-specificity is the probability that
when testing a healthy subject (N) the when testing a diseased subject (P)
result will be positive (RP). the result will be negative (RN).
The predictive values are:

- positive predictive value (PPV) - negative predictive value (NPV)
TP TP TN TN
PPV = = NPV = =
RP TP + FP RN TN + FN
The probability that a subject that The probability that a subject that
tested positive (RP) is actually tested negative (RN) is actually
diseased (P). healthy (N).
Two interesting measures are also

- false discovery rate (FDR) - accuracy (ACC)
FP FP TP + TN
FDR = = ACC =
RP TP + FP P+N
The probability that a subject that The ration of all correctly
tested positive (RP) is actually classified subjects.
healthy (N).
ROC curve (2)
The receiver operating characteristic (ROC) curve displays the course of

sensitivity (TPR = 1 – FNR) against non-specificity (FPR = 1 – TNR) at
different values of threshold t.
FNR
FPR TNR
(sensitivity)
TPR
TPR
0
0 FPR 1
(non-specificity)
ROC curve (2)
RP RN
1
subjects
TP TN
1 1
result
2
2
3
TPR
TP TN
FP FN
result
3
0
FP FN 0 FPR 1
result
ROC curve (3)
The area under the [ROC] curve (AUC) is equal to the probability that
a (diagnostic) test will classify a randomly chosen positive sample better
than a randomly chosen negative sample.
The test (or its success) can be therefore 1
labeled as:
- excellent
1.0 ≥ AUC > 0.9
- good
TPR
0.9 ≥ AUC > 0.8 AUC
- fair
0.8 ≥ AUC > 0.7
- poor
0
0.7 ≥ AUC > 0.6
0 FPR 1
- fail
0.6 ≥ AUC > 0.5
Example
Positive test Negative test

(RP) (RN)
Diseased
TP = 8 FN = 12 P = 20
subjects (P)
Healthy
FP = 4 TN = 12 N = 16
subjects (N)
RP = 12 RN = 24 P + N = 36
TP TP 8 TP TP 8
TPR = = = = 40.0% PPV = = = = 66.7%
P TP + FN 8 + 12 RP TP + FP 8 + 4
TN TN 12 TN TN 12
TNR = = = = 75.0% NPV = = = = 50.0%
N TN + FP 12 + 4 RN TN + FN 12 + 12
FP FP 4 FP FP 4
FPR = = = = 25.0% FDR = = = = 33.3%
N TN + FP 12 + 4 RP TP + FP 8 + 4
FN FN 12 TP + TN 8 + 12
FNR = = = = 60.0% ACC = = = 55.6%
P TP + FN 8 + 12 P+N 36
LABORATORY WORK
Lab work 9: Binary classification
Your task will be to perform binary classification of results that represent

blood pressure measurements for a larger group of diseased and healthy
subjects according to four different methods (four diagnostic tests).
Classification will be based on comparison to reference data that defines

two classes of subjects, namely:
- a class labeled with “0” for healthy subjects
- a class labeled with “1” for diseased subjects
You will select a classification threshold and assign the subjects into groups
(TP, TN, FP and FN), and then compute the performance measures (TPR,
TNR, FPR, FNR, PPV, NPV, FDR and ACC). According to the obtained
results at different classification thresholds you will display ROC curves
and finally compute the AUC for the ROC curve.
LABORATORY WORK
Lab work 9: Binary classification
Test 1 Test 2 Test 3 Test 4

Reference results results results results
Subject data (mmHg) (mmHg) (mmHg) (mmHg)
⁞ ⁞ ⁞ ⁞ ⁞ ⁞
71 0 85.5 98.5 114.8 115.0
72 1 144.2 70.6 113.8 115.0
73 0 93.4 145.6 117.9 115.0
⁞ ⁞ ⁞ ⁞ ⁞ ⁞
consecutive
subject
number actual condition
results of four
(0 = healthy,
different tests
1 = diseased)
CONCLUSION
Discussion, comments, questions…
- What is binary classification and what is it useful for?

- Which groups can be the samples classified into?
- Describe the contingency table.
- What are the properties of ideal and what of real classification?
- Describe the consistency measures for classification performance.
- Describe the inconsistency measures for classification performance.
- Describe the predictive values for classification performance.
- What other measures for classification performance exist?
- What is the relationship between TPR and FNR, and what between
TNR and FPR?
- Describe the ROC curve.
- What does the area under the ROC curve represent, and how can you
label a test according to the resulting area?

Laboratory Work Preparation: Lab Work 9: Binary Classification

Uploaded by

Copyright:

Available Formats

Laboratory Work Preparation: Lab Work 9: Binary Classification

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Laboratory Work Preparation: Lab Work 9: Binary Classification

Uploaded by

Copyright:

Available Formats

LABORATORY WORK PREPARATION

Lab Work 9: Binary classification

University of Ljubljana, Faculty of Electrical Engineering Electrical Engineering, level 2

- false positives – FP - false negatives – FN

The contingency table represents a condensed description of sample

Positive test Negative test

In the case of an ideal classification, there are no false

In the case of a real classification, false positives and

The measures of consistency are:

The measures of inconsistency are:

The predictive values are:

Two interesting measures are also

The receiver operating characteristic (ROC) curve displays the course of

Positive test Negative test

Your task will be to perform binary classification of results that represent

Classification will be based on comparison to reference data that defines

Test 1 Test 2 Test 3 Test 4

- What is binary classification and what is it useful for?

You might also like