ROC: Difference between revisions

From 太極
Jump to navigation Jump to search
 
(24 intermediate revisions by the same user not shown)
Line 1: Line 1:
= Confusion matrix, Sensitivity/Specificity/Accuracy =
= Confusion matrix, Sensitivity/Specificity/Accuracy =
[https://en.wikipedia.org/wiki/Confusion_matrix Wikipedia]
[https://en.wikipedia.org/wiki/Confusion_matrix Wikipedia]. Note in wikipedia, it uses positive/negative to represent two classes.


{| border="1" style="border-collapse:collapse; text-align:center;"
{| border="1" style="border-collapse:collapse; text-align:center;"
Line 25: Line 25:
* Negative predictive value (NPV) = TN /  # negative calls = TN / (FN + TN)
* Negative predictive value (NPV) = TN /  # negative calls = TN / (FN + TN)
* Prevalence 盛行率 = (TP + FN) / N.
* Prevalence 盛行率 = (TP + FN) / N.
* Note that PPV & NPV can also be computed from sensitivity, specificity, and prevalence:
* Note that PPV & NPV can also be computed from sensitivity, specificity, and '''prevalence''':
** [https://en.wikipedia.org/wiki/Positive_and_negative_predictive_values#cite_note-AltmanBland1994-2 PPV is directly proportional to the prevalence of the disease or condition.].  
** [https://en.wikipedia.org/wiki/Positive_and_negative_predictive_values#cite_note-AltmanBland1994-2 PPV is directly proportional to the prevalence of the disease or condition.].  
** For example, in the extreme case if the prevalence =1, then PPV is always 1.
** For example, in the extreme case if the prevalence =1, then PPV is always 1.
Line 43: Line 43:


== PPA, NPA ==  
== PPA, NPA ==  
[https://www.fda.gov/regulatory-information/search-fda-guidance-documents/statistical-guidance-reporting-results-studies-evaluating-diagnostic-tests-guidance-industry-and-fda Positive percent agreement (PPA) and negative percent agreement (NPA)]
<ul>
<li>Cf '''PPV'''/positive predictive value & '''NPV'''/negative predictive value
<li>Positive percent agreement (PPA) and negative percent agreement (NPA) - [https://www.fda.gov/regulatory-information/search-fda-guidance-documents/statistical-guidance-reporting-results-studies-evaluating-diagnostic-tests-guidance-industry-and-fda Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests - Guidance for Industry and FDA Staff]
* The ''percentage of subjects (or samples) that test positive by one test (TA)'' that are found positive by a second test (TB); the two tests are indicated with the ordered subscripts, as <math>PPA_{TA:TB}</math>. The test listed first is treated as the '''“reference” test''' (denominator a+b in the formula given below).
<li>[https://analyse-it.com/blog/2020/4/diagnostic-accuracy-sensitivity-specificity-versus-agreement-ppa-npa-statistics Diagnostic accuracy (sensitivity/specificity) versus agreement (PPA/NPA) statistics]
* Positive agreement is the ''proportion of '''comparative/reference method''' positive results'' in which the '''test method''' result is positive.
* Negative agreement is the proportion of comparative/reference method negative results in which the test method result is negative.
<li>The focus on PPA or NPA depends on the context and the specific objectives of the test evaluation.
* <span style="color: red">PPA is particularly important when the consequence of a false negative is high. For example, in a disease screening context, a false negative (a sick person being incorrectly identified as healthy) could lead to delayed treatment and worse health outcomes. Therefore, a high PPA (few false negatives) is desirable.</span>
* <span style="color: red">NPA is crucial when the cost of a false positive is significant</span>. For instance, in a scenario where a positive test result could lead to an invasive follow-up procedure, a false positive (a healthy person being incorrectly identified as sick) could lead to unnecessary risk and anxiety for the patient. In this case, a high NPA (few false positives) is important.
<li>Paper examples.
* [https://www.frontiersin.org/articles/10.3389/fnagi.2022.1003792/full Primary progressive aphasia and motor neuron disease: A review]
<li>Other evaluation of diagnostic tests: Sensitivity/Specificity, Positive Predictive Value (PPV)/Negative Predictive Value (NPV), Overall Test Accuray, Diagnostic Odds Ratio (DOR), Youden Index (YI), Positive Likelihood Ratio (LR+).  See
* [https://www.degruyter.com/document/doi/10.1515/tjb-2020-0337/html Evaluation of binary diagnostic tests accuracy for medical researches] 2020,
* [https://journals.sagepub.com/doi/epdf/10.1177/201010581102000411 Measures of Diagnostic Accuracy: Sensitivity, Specificity, PPV and NPV] 2011 where Fig 1 shows the effect of '''disease prevalence''' on PPV and NPV.
* [https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0223832 Diagnostic test evaluation methodology: A systematic review of methods employed to evaluate diagnostic tests in the absence of gold standard – An update] 2019.
 
<li>Assume rows are '''new test''' and columns are '''reference'''.
{| class="wikitable"
|-
|
| reference +
| reference -
|-
| test +
| TP (a)
| FP (c)
|-
| test -
| FN (b)
| TN (d)
|-
|
| x.1
| x.2
|}  PPA = TP/x.1 = TP/(TP+FN), NPA = TN/x.2 = TN/(FP+TN)
 
<li>APA ('''Average''' positive agreement) and ANA ('''Average''' negative agreement). See [https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200001.pdf#page=3 EVALUATION OF AUTOMATIC CLASS III DESIGNATION FOR Cell-Free DNA BCT] and [https://analyse-it.com/docs/user-guide/method-comparison/agreementmeasures Agreement measures for binary and semi-quantitative data]. '''Note, APA and ANA are not standard terms used in diagnostic test evaluation, and their definitions may vary depending on the context'''.
<ul>
<li>Suppose <math>PPA = \frac{a}{a+b}</math>, then <math>APA = \frac{a}{((a+b)+(a+c))/2} = \frac{2x_{11}}{x_{1.}+x_{.1}} =
\frac{\mbox{n. of concordant positives}}{\mbox{n. of concordant positives} + \frac{\mbox{n. of disconcordant calls}}{2}} </math>.
<li>Suppose <math>NPA = \frac{d}{c+d}</math>, then <math>ANA = \frac{d}{((d+b)+(d+c))/2} = \frac{2x_{22}}{x_{2.}+x_{.2}} =
\frac{\mbox{n. of concordant negatives}}{\mbox{n. of concordant negatives} + \frac{\mbox{n. of disconcordant calls}}{2}} </math>.
<li>The overall proportion of agreement is the sum of the diagonal entries divided by the total.
</ul>
<li>The above two APA and ANA statistics are in the same form as the Harrell's C-index. See Harrell's C-index formula at [https://statisticaloddsandends.wordpress.com/2019/10/26/what-is-harrells-c-index/ What is Harrell’s C-index?]
</ul>
 
== Random classifier ==
* Sensitivity and specificity should be close to 0.5 for a random classifier assuming a balanced class distribution in a binary classification problem.
* However, PPV or NPV is not close to 0.5 for a random classifier. PPV for a random classifier will depend on the class distribution in the dataset and the randomness of the classifier's guesses. It would vary randomly depending on the classifier's random guesses and the class distribution in the dataset.


= ROC curve =
= ROC curve =
Line 137: Line 187:
** [https://juliasilge.com/blog/sf-trees-random-tuning/ Tuning random forest hyperparameters with #TidyTuesday trees data]
** [https://juliasilge.com/blog/sf-trees-random-tuning/ Tuning random forest hyperparameters with #TidyTuesday trees data]


== Optimal threshold ==
== Optimal threshold, Youden's index/Youden's J statistic ==
* [https://homepage.stat.uiowa.edu/~rdecook/stat6220/Class_notes/ROC_introduction.pdf#page=25 Max of “sensitivity + specificity”]. See [https://cran.r-project.org/web/packages/Epi/index.html Epi::ROC()] function.
* [https://homepage.stat.uiowa.edu/~rdecook/stat6220/Class_notes/ROC_introduction.pdf#page=25 Max of “sensitivity + specificity”]. See [https://cran.r-project.org/web/packages/Epi/index.html Epi::ROC()] function.
* [https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.9432 On optimal biomarker cutoffs accounting for misclassification costs in diagnostic trilemmas with applications to pancreatic cancer] Bantis, 2022
* [https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.9432 On optimal biomarker cutoffs accounting for misclassification costs in diagnostic trilemmas with applications to pancreatic cancer] Bantis, 2022
* [https://en.wikipedia.org/wiki/Youden%27s_J_statistic Youden's index]
** Youden's index = sensitivity + specificity -1
** Youden's index ranges from 0 to 1 with higher values indicating better overall test performance
** The maximum value of the index may be used as a criterion for selecting the optimum cut-off point when a diagnostic test gives a numeric rather than a dichotomous result.
* [https://cran.r-project.org/web/packages/rocbc/index.html rocbc] package. Provides inferences and comparisons around the AUC, the Youden index, the sensitivity at a given specificity level (and vice versa), the optimal operating point of the ROC curve (in the Youden sense), and the Youden based cutoff.


== ROC Curve AUC for Hypothesis Testing ==
== ROC Curve AUC for Hypothesis Testing ==
Line 163: Line 218:
** Y-axis: Sensitivity = tp/(tp + fn) = Recall
** Y-axis: Sensitivity = tp/(tp + fn) = Recall
** X-axis: 1-Specificity = fp/(fp + tn)
** X-axis: 1-Specificity = fp/(fp + tn)
* ROC curves are appropriate when the observations are balanced between each class, whereas <span style="color: red">'''PR curves''' are appropriate for imbalanced datasets</span>. When dealing with highly skewed datasets, '''PR curves''' can give a more informative picture of an algorithm’s performance.
* [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0118432 The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets]
* [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0118432 The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets]


Line 200: Line 256:
confusionMatrix(t(xtab))$positive
confusionMatrix(t(xtab))$positive
# [1] "A"
# [1] "A"
</pre>
<li>[https://stackoverflow.com/a/49467039 Using F1 score metric in KNN through caret package]
<pre>
library(caret)
library(mlbench)
data(Sonar)
f1Summary <- function(data, lev = NULL, model = NULL) {
  # Calculate precision and recall
  precision <- posPredValue(data$pred, data$obs, positive = lev[1])
  recall <- sensitivity(data$pred, data$obs, positive = lev[1])
 
  # Calculate F1 score
  f1 <- (2 * precision * recall) / (precision + recall)
 
  # Return F1 score
  c(F1 = f1)
}
# Create a trainControl object that uses the custom summary function
trainControl <- trainControl(summaryFunction = f1Summary)
# Use the trainControl object when calling the train() function
model <- train(Class ~., data = Sonar, method = "rf",
              trControl = trainControl, metric = "F1")
> model
Random Forest
208 samples
60 predictor
  2 classes: 'M', 'R'
No pre-processing
Resampling: Bootstrapped (25 reps)
Summary of sample sizes: 208, 208, 208, 208, 208, 208, ...
Resampling results across tuning parameters:
  mtry  F1
  2    0.8416808
  31    0.7986326
  60    0.7820982
F1 was used to select the optimal model using the largest value.
The final value used for the model was mtry = 2.
</pre>
</pre>
<li>Examples in papers
<li>Examples in papers
* [https://www.biorxiv.org/content/10.1101/2023.04.19.537463v1.full.pdf A Comprehensive Benchmark Study on Biomedical Text Generation and Mining with ChatGPT]
* [https://www.biorxiv.org/content/10.1101/2023.04.19.537463v1.full.pdf A Comprehensive Benchmark Study on Biomedical Text Generation and Mining with ChatGPT]
</ul>
</ul>
== PPV/Precision ==
The relationship between PPV and sensitivity depends on the '''prevalence''' of the condition being tested for in the population. If the condition is rare, a test with high sensitivity may still have a low PPV because there will be many false positive results. On the other hand, if the condition is common, a test with high sensitivity is more likely to also have a high PPV.
Here we have a diagnostic test for a rare disease that affects 1 in 1000 people and the test has a sensitivity of 100% and a false positive rate of 5%:
<pre>
                Test| positive  negative
-----------------------------------------
Disease present    |  1          0
Disease absent      |  50        949
</pre>
The prevalence of the disease is 1/1000.


= Incidence, Prevalence =
= Incidence, Prevalence =
Line 239: Line 352:
* [https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-019-1466-7 Calibration: the Achilles heel of predictive analytics] 2019
* [https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-019-1466-7 Calibration: the Achilles heel of predictive analytics] 2019


= Performance evaluation =
= Comparison of two AUCs =
* [https://onlinelibrary.wiley.com/doi/epdf/10.1002/sim.5727 Testing for improvement in prediction model performance] by Pepe et al 2013.
* [https://onlinelibrary.wiley.com/doi/epdf/10.1002/sim.5727 Testing for improvement in prediction model performance] by Pepe et al 2013.
* [https://statcompute.wordpress.com/2018/12/25/statistical-assessments-of-auc/ Statistical Assessments of AUC]. This is using the '''pROC::roc.test''' function.
* [https://cran.r-project.org/web/packages/prioritylasso/vignettes/prioritylasso_vignette.html prioritylasso]. It is using roc(), auc(), roc.test(), plot.roc() from the '''pROC''' package. The calculation based on the training data is biased so we need to report the one based on test data.


== Youden's index/Youden's J statistic ==
== DeLong test for comparing two ROC curves ==
* [https://en.wikipedia.org/wiki/Youden%27s_J_statistic Youden's index]
* [https://glassboxmedicine.com/2020/02/04/comparing-aucs-of-machine-learning-models-with-delongs-test/ Comparing AUCs of Machine Learning Models with DeLong’s Test]
* [https://cran.r-project.org/web/packages/rocbc/index.html rocbc] package. Provides inferences and comparisons around the AUC, the Youden index, the sensitivity at a given specificity level (and vice versa), the optimal operating point of the ROC curve (in the Youden sense), and the Youden based cutoff.
* [https://pubmed.ncbi.nlm.nih.gov/22415937/ Misuse of DeLong test to compare AUCs for nested models]
* [https://statisticaloddsandends.wordpress.com/2020/06/07/what-is-the-delong-test-for-comparing-aucs/ What is the DeLong test for comparing AUCs?]
* [https://www.jianshu.com/p/9b0556134eec R语言,ROC曲线,deLong test]. [https://cran.r-project.org/web/packages/pROC/ pROC]::roc.test() was used.
* [https://www.rdocumentation.org/packages/Daim/versions/1.1.0/topics/deLong.test Daim::deLong.test()]


= Some R packages =
= Some R packages =
Line 261: Line 379:
<ul>
<ul>
<li>https://cran.r-project.org/web/packages/pROC/index.html
<li>https://cran.r-project.org/web/packages/pROC/index.html
* response: a factor, numeric or character vector of responses (true class), typically encoded with 0 (controls) and 1 (cases).
* Pay attention to the message printed on screen after calling roc(). It will show which one is control and which one is case.
<li>[https://web.expasy.org/pROC/ pROC: display and analyze ROC curves in R and S+]
<li>[https://web.expasy.org/pROC/ pROC: display and analyze ROC curves in R and S+]
<li>[https://medium.com/swlh/roc-curve-and-auc-detailed-understanding-and-r-proc-package-86d1430a3191 ROC Curve and AUC in Machine learning and R pROC Package]. Pay attention to the '''legacy.axes''' parameter.
<li>[https://medium.com/swlh/roc-curve-and-auc-detailed-understanding-and-r-proc-package-86d1430a3191 ROC Curve and AUC in Machine learning and R pROC Package]. Pay attention to the '''legacy.axes''' parameter.
Line 298: Line 418:
# Area under the curve: 0.7314
# Area under the curve: 0.7314
</pre>
</pre>
[[File:Roc asah.png|350px]]
</li>
<li>Plot multiple ROC curves in one plot.
<pre>
roc1 <- roc()
roc2 <- roc()
roc3 <- roc()
plot(roc1)
lines(roc2, col="red")
lines(roc3, col="blue")
</pre>
Another solution is to use [https://rdrr.io/cran/pROC/man/ggroc.html pROC::ggroc()]. The annotation was taken care automatically.
<pre>
rocobj <- roc(aSAH$outcome, aSAH$s100b)
rocobj2 <- roc(aSAH$outcome, aSAH$wfns)
g2 <- ggroc(list(s100b=rocobj,
                wfns=rocobj2,
                ndka=roc(aSAH$outcome, aSAH$ndka))) # 3 curves
</pre>
<li>Plot a panel of ROC plots by using facet_wrap().
* [https://stackoverflow.com/a/69394230 How to add AUC to a multiple ROC graph with pROC's ggroc]. ggroc() is part of pROC package.
* See [https://stackoverflow.com/a/11889798 Annotating text on individual facet in ggplot2]
* See [[Ggplot2#facet_wrap_and_facet_grid_to_create_a_panel_of_plots|facet_wrap]] on how to change the labels in each plots.
</li>
</li>
</ul>
</ul>
Line 355: Line 500:
= mean ROC curve =
= mean ROC curve =
[https://stackoverflow.com/a/66001395 ROC with cross-validation for linear regression in R]
[https://stackoverflow.com/a/66001395 ROC with cross-validation for linear regression in R]
= Comparison of two AUCs =
* [https://statcompute.wordpress.com/2018/12/25/statistical-assessments-of-auc/ Statistical Assessments of AUC]. This is using the '''pROC::roc.test''' function.
* [https://cran.r-project.org/web/packages/prioritylasso/vignettes/prioritylasso_vignette.html prioritylasso]. It is using roc(), auc(), roc.test(), plot.roc() from the '''pROC''' package. The calculation based on the training data is biased so we need to report the one based on test data.


= Assess risk of bias =
= Assess risk of bias =
Line 365: Line 506:
= Confidence interval of AUC =
= Confidence interval of AUC =
[http://theautomatic.net/2019/08/20/how-to-get-an-auc-confidence-interval/ How to get an AUC confidence interval]. [https://cran.r-project.org/web/packages/pROC/index.html pROC] package was used.
[http://theautomatic.net/2019/08/20/how-to-get-an-auc-confidence-interval/ How to get an AUC confidence interval]. [https://cran.r-project.org/web/packages/pROC/index.html pROC] package was used.
= DeLong test for comparing two ROC curves =
* [https://glassboxmedicine.com/2020/02/04/comparing-aucs-of-machine-learning-models-with-delongs-test/ Comparing AUCs of Machine Learning Models with DeLong’s Test]
* [https://pubmed.ncbi.nlm.nih.gov/22415937/ Misuse of DeLong test to compare AUCs for nested models]
* [https://statisticaloddsandends.wordpress.com/2020/06/07/what-is-the-delong-test-for-comparing-aucs/ What is the DeLong test for comparing AUCs?]
* [https://www.jianshu.com/p/9b0556134eec R语言,ROC曲线,deLong test]. [https://cran.r-project.org/web/packages/pROC/ pROC]::roc.test() was used.
* [https://www.rdocumentation.org/packages/Daim/versions/1.1.0/topics/deLong.test Daim::deLong.test()]


= AUC can be a misleading measure of performance =
= AUC can be a misleading measure of performance =
Line 388: Line 522:
* [https://onlinelibrary.wiley.com/doi/full/10.1002/sim.9412 Sample size methods for evaluation of predictive biomarkers] 2022
* [https://onlinelibrary.wiley.com/doi/full/10.1002/sim.9412 Sample size methods for evaluation of predictive biomarkers] 2022


= Picking a threshold based on model performance/utility =
= Threshold =
[https://win-vector.com/2020/10/05/squeezing-the-most-utility-from-your-models/ Squeezing the Most Utility from Your Models]
== Myths about risk thresholds for prediction models ==
[https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-019-1425-3 Three myths about risk thresholds for prediction models] 2019
 
== Picking a threshold based on model performance/utility ==
[https://win-vector.com/2020/10/05/squeezing-the-most-utility-from-your-models/ Squeezing the Most Utility from Your Models] 2020


= Why does my ROC curve look like a V =
= Why does my ROC curve look like a V =
Line 434: Line 572:
* [https://bioconductor.org/packages/release/bioc/html/compcodeR.html compcodeR]: RNAseq data simulation, differential expression analysis and performance comparison of differential expression methods
* [https://bioconductor.org/packages/release/bioc/html/compcodeR.html compcodeR]: RNAseq data simulation, differential expression analysis and performance comparison of differential expression methods
* [https://academic.oup.com/bioinformatics/article/31/17/2778/183245 Polyester]: simulating RNA-seq datasets with differential transcript expression, [https://github.com/leekgroup/polyester_code github], [https://htmlpreview.github.io/?https://github.com/leekgroup/polyester_code/blob/master/polyester_manuscript.html HTML]
* [https://academic.oup.com/bioinformatics/article/31/17/2778/183245 Polyester]: simulating RNA-seq datasets with differential transcript expression, [https://github.com/leekgroup/polyester_code github], [https://htmlpreview.github.io/?https://github.com/leekgroup/polyester_code/blob/master/polyester_manuscript.html HTML]
* [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0232271 Benchmarking RNA-seq differential expression analysis methods using spike-in and simulation data] Baik 2020


= Reporting =
= Reporting =

Latest revision as of 17:06, 29 October 2023

Confusion matrix, Sensitivity/Specificity/Accuracy

Wikipedia. Note in wikipedia, it uses positive/negative to represent two classes.

Predict
1 0
True 1 TP FN Sens=TP/(TP+FN)=Recall=TPR
FNR=FN/(TP+FN)
0 FP TN Spec=TN/(FP+TN), 1-Spec=FPR
PPV=TP/(TP+FP)=Precision
FDR=FP/(TP+FP) =1-PPV
NPV=TN/(FN+TN) N = TP + FP + FN + TN
  • Sensitivity 敏感度 = TP / (TP + FN) = Recall
  • Specificity 特異度 = TN / (TN + FP)
  • Accuracy = (TP + TN) / N
  • False discovery rate FDR = FP / (TP + FP)
  • False negative rate FNR = FN / (TP + FN)
  • False positive rate FPR = FP / (FP + TN) = 1 - Spec
  • True positive rate = TP / (TP + FN) = Sensitivity
  • Positive predictive value (PPV) = TP / # positive calls = TP / (TP + FP) = 1 - FDR = Precision
  • Negative predictive value (NPV) = TN / # negative calls = TN / (FN + TN)
  • Prevalence 盛行率 = (TP + FN) / N.
  • Note that PPV & NPV can also be computed from sensitivity, specificity, and prevalence:
[math]\displaystyle{ \text{PPV} = \frac{\text{sensitivity} \times \text{prevalence}}{\text{sensitivity} \times \text{prevalence}+(1-\text{specificity}) \times (1-\text{prevalence})} }[/math]
[math]\displaystyle{ \text{NPV} = \frac{\text{specificity} \times (1-\text{prevalence})}{(1-\text{sensitivity}) \times \text{prevalence}+\text{specificity} \times (1-\text{prevalence})} }[/math]

caret::confusionMatrix

False positive rates vs false positive rates

FPR (false positive rate) vs FDR (false discovery rate)

PPA, NPA

  • Cf PPV/positive predictive value & NPV/negative predictive value
  • Positive percent agreement (PPA) and negative percent agreement (NPA) - Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests - Guidance for Industry and FDA Staff
    • The percentage of subjects (or samples) that test positive by one test (TA) that are found positive by a second test (TB); the two tests are indicated with the ordered subscripts, as [math]\displaystyle{ PPA_{TA:TB} }[/math]. The test listed first is treated as the “reference” test (denominator a+b in the formula given below).
  • Diagnostic accuracy (sensitivity/specificity) versus agreement (PPA/NPA) statistics
    • Positive agreement is the proportion of comparative/reference method positive results in which the test method result is positive.
    • Negative agreement is the proportion of comparative/reference method negative results in which the test method result is negative.
  • The focus on PPA or NPA depends on the context and the specific objectives of the test evaluation.
    • PPA is particularly important when the consequence of a false negative is high. For example, in a disease screening context, a false negative (a sick person being incorrectly identified as healthy) could lead to delayed treatment and worse health outcomes. Therefore, a high PPA (few false negatives) is desirable.
    • NPA is crucial when the cost of a false positive is significant. For instance, in a scenario where a positive test result could lead to an invasive follow-up procedure, a false positive (a healthy person being incorrectly identified as sick) could lead to unnecessary risk and anxiety for the patient. In this case, a high NPA (few false positives) is important.
  • Paper examples.
  • Other evaluation of diagnostic tests: Sensitivity/Specificity, Positive Predictive Value (PPV)/Negative Predictive Value (NPV), Overall Test Accuray, Diagnostic Odds Ratio (DOR), Youden Index (YI), Positive Likelihood Ratio (LR+). See
  • Assume rows are new test and columns are reference.
    reference + reference -
    test + TP (a) FP (c)
    test - FN (b) TN (d)
    x.1 x.2
    PPA = TP/x.1 = TP/(TP+FN), NPA = TN/x.2 = TN/(FP+TN)
  • APA (Average positive agreement) and ANA (Average negative agreement). See EVALUATION OF AUTOMATIC CLASS III DESIGNATION FOR Cell-Free DNA BCT and Agreement measures for binary and semi-quantitative data. Note, APA and ANA are not standard terms used in diagnostic test evaluation, and their definitions may vary depending on the context.
    • Suppose [math]\displaystyle{ PPA = \frac{a}{a+b} }[/math], then [math]\displaystyle{ APA = \frac{a}{((a+b)+(a+c))/2} = \frac{2x_{11}}{x_{1.}+x_{.1}} = \frac{\mbox{n. of concordant positives}}{\mbox{n. of concordant positives} + \frac{\mbox{n. of disconcordant calls}}{2}} }[/math].
    • Suppose [math]\displaystyle{ NPA = \frac{d}{c+d} }[/math], then [math]\displaystyle{ ANA = \frac{d}{((d+b)+(d+c))/2} = \frac{2x_{22}}{x_{2.}+x_{.2}} = \frac{\mbox{n. of concordant negatives}}{\mbox{n. of concordant negatives} + \frac{\mbox{n. of disconcordant calls}}{2}} }[/math].
    • The overall proportion of agreement is the sum of the diagonal entries divided by the total.
  • The above two APA and ANA statistics are in the same form as the Harrell's C-index. See Harrell's C-index formula at What is Harrell’s C-index?

Random classifier

  • Sensitivity and specificity should be close to 0.5 for a random classifier assuming a balanced class distribution in a binary classification problem.
  • However, PPV or NPV is not close to 0.5 for a random classifier. PPV for a random classifier will depend on the class distribution in the dataset and the randomness of the classifier's guesses. It would vary randomly depending on the classifier's random guesses and the class distribution in the dataset.

ROC curve

  • Binary case:
    • Y = true positive rate = sensitivity,
    • X = false positive rate = 1-specificity = 假陽性率
  • Area under the curve AUC from the wikipedia: the probability that a classifier will rank a randomly chosen positive instance higher than a randomly chosen negative one (assuming 'positive' ranks higher than 'negative').
    [math]\displaystyle{ A = \int_{\infty}^{-\infty} \mbox{TPR}(T) \mbox{FPR}'(T) \, dT = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} I(T'\gt T)f_1(T') f_0(T) \, dT' \, dT = P(X_1 \gt X_0) }[/math]
    where [math]\displaystyle{ X_1 }[/math] is the score for a positive instance and [math]\displaystyle{ X_0 }[/math] is the score for a negative instance, and [math]\displaystyle{ f_0 }[/math] and [math]\displaystyle{ f_1 }[/math] are probability densities as defined in previous section.
  • Interpretation of the AUC. A small toy example (n=12=4+8) was used to calculate the exact probability [math]\displaystyle{ P(X_1 \gt X_0) }[/math] (4*8=32 all combinations).
    • It is a discrimination measure which tells us how well we can classify patients in two groups: those with and those without the outcome of interest.
    • Since the measure is based on ranks, it is not sensitive to systematic errors in the calibration of the quantitative tests.
    • The AUC can be defined as The probability that a randomly selected case will have a higher test result than a randomly selected control.
    • Plot of sensitivity/specificity (y-axis) vs cutoff points of the biomarker
    • The Mann-Whitney U test statistic (or Wilcoxon or Kruskall-Wallis test statistic) is equivalent to the AUC (Mason, 2002). Verify in R.
library(pROC)
set.seed(123)
group1 <- rnorm(50)
group2 <- rnorm(50, mean=1)

roc_obj <- roc(rep(c(0,1), each=50), c(group1, group2))
# Setting levels: control = 0, case = 1
# Setting direction: controls < cases
auc(roc_obj)
# Area under the curve: 0.8036

wilcox.test(group2, group1)$statistic/(50*50) # notice the group1,group2 order
#      W 
# 0.8036
    • The p-value of the Mann-Whitney U test can thus safely be used to test whether the AUC differs significantly from 0.5 (AUC of an uninformative test).
  • Calculate AUC by hand. AUC is equal to the probability that a true positive is scored greater than a true negative.
  • See the uROC() function in <functions.R> from the supplementary of the paper (need access right) Bivariate Marker Measurements and ROC Analysis Wang 2012. Let [math]\displaystyle{ n_1 }[/math] be the number of obs from X1 and [math]\displaystyle{ n_0 }[/math] be the number of obs from X0. X1 and X0 are the predict values for data from group 1 and 0. [math]\displaystyle{ TP_i=Prob(X_1\gt X_{0i})=\sum_j (X_{1j} \gt X_{0i})/n_1, ~ FP_i=Prob(X_0\gt X_{0i}) = \sum_j (X_{0j} \gt X_{0i}) / n_0 }[/math]. We can draw a scatter plot or smooth.spline() of TP(y-axis) vs FP(x-axis) for the ROC curve.
uROC <- function(marker, status)   ### ROC function for univariate marker ###
{
    x <- marker
    bad <-  is.na(status) | is.na(x) 
    status <- status[!bad]
    x <- x[!bad]
    if (sum(bad) > 0) 
        cat(paste("\n", sum(bad), "records with missing values dropped. \n"))
	no_case <- sum(status==1)
	no_control <- sum(status==0)
	TP <- rep(0, no_control)
	FP <- rep(0, no_control)
	for (i in 1: no_control){	
	  TP[i] <- sum(x[status==1]>x[status==0][i])/no_case	
	  FP[i] <- sum(x[status==0]>x[status==0][i])/no_control	
    }
    list(TP = TP, FP = FP)
}

partial AUC

summary ROC

Weighted ROC

Adjusted AUC

Difficult to compute for some models

Optimal threshold, Youden's index/Youden's J statistic

ROC Curve AUC for Hypothesis Testing

Challenges, issues

Methodological conduct of prognostic prediction models developed using machine learning in oncology: a systematic review 2022. class imbalance, data pre-processing, and hyperparameter tuning. twitter.

Survival data

'Survival Model Predictive Accuracy and ROC Curves' by Heagerty & Zheng 2005

  • Recall Sensitivity= [math]\displaystyle{ P(\hat{p_i} \gt c | Y_i=1) }[/math], Specificity= [math]\displaystyle{ P(\hat{p}_i \le c | Y_i=0 }[/math]), [math]\displaystyle{ Y_i }[/math] is binary outcomes, [math]\displaystyle{ \hat{p}_i }[/math] is a prediction, [math]\displaystyle{ c }[/math] is a criterion for classifying the prediction as positive ([math]\displaystyle{ \hat{p}_i \gt c }[/math]) or negative ([math]\displaystyle{ \hat{p}_i \le c }[/math]).
  • For survival data, we need to use a fixed time/horizon (t) to classify the data as either a case or a control. Following Heagerty and Zheng's definition in Survival Model Predictive Accuracy and ROC Curves (Incident/dynamic) 2005, Sensitivity(c, t)= [math]\displaystyle{ P(M_i \gt c | T_i = t) }[/math], Specificity= [math]\displaystyle{ P(M_i \le c | T_i \gt t }[/math]) where M is a marker value or [math]\displaystyle{ Z^T \beta }[/math]. Here sensitivity measures the expected fraction of subjects with a marker greater than c among the subpopulation of individuals who die at time t, while specificity measures the fraction of subjects with a marker less than or equal to c among those who survive beyond time t.
  • The AUC measures the probability that the marker value for a randomly selected case exceeds the marker value for a randomly selected control
  • ROC curves are useful for comparing the discriminatory capacity of different potential biomarkers.


Precision recall curve

F1-score

  • F-score/F1-score.
    • F1 = 2 * Precision * Recall / (Precision + Recall) = 2 / (1/Precision + 1/Recall) Harmonic mean of precision and recall.
    • This means that comparison of the F-score across different problems with differing class ratios is problematic.
    • Recall = Sensitivity = true positive rate. Precision = PPV. They are from the two directions of the "1" class in the confusion matrix.
    • The F1 score
    • How to interpret F1 score (simply explained). It is a popular metric to use for classification models as it provides accurate results for both balanced and imbalanced datasets, and takes into account both the precision and recall ability of the model. F1 score tells you the model’s balanced ability to both capture positive cases (recall) and be accurate with the cases it does capture (precision).
    • What is Considered a “Good” F1 Score? It gives an example of F1-score where all the predictions are on one class (baseline model) for comparison purpose.
    • An example. confusionMatrix() assumes the rows are predicted and columns are true; that's why I put a transpose on my table object. The 'positive' class seems to be always the 1st row of the 2x2 table. If the levels are not specified, it will assign 'A' to the first row and 'B' to the 2nd row. For this case, we can see accuracy (0.615) is between sensitivity (0.429) and specificity (0.684) and F1-score (0.375) is between precision (0.333) and recall (0.429) . Also F1-score is smaller than sensitivity and specificity.
                  predicted
             sensitive   resistant
    -------------------------------
    true
    sensitive     3          4
    resistant     6         13
    
    xtab <- matrix(c(3,6,4,13),nr=2)
    confusionMatrix(t(xtab))$byClass
    #         Sensitivity          Specificity       Pos Pred Value       Neg Pred Value 
    #           0.4285714            0.6842105            0.3333333            0.7647059 
    #           Precision               Recall                   F1           Prevalence 
    #           0.3333333            0.4285714            0.3750000            0.2692308 
    #      Detection Rate Detection Prevalence    Balanced Accuracy 
    #           0.1153846            0.3461538            0.5563910
    
    confusionMatrix(t(xtab))$overall
    #       Accuracy          Kappa  AccuracyLower  AccuracyUpper   AccuracyNull 
    #      0.6153846      0.1034483      0.4057075      0.7977398      0.7307692 
    # AccuracyPValue  McnemarPValue 
    #      0.9348215      0.7518296
    
    confusionMatrix(t(xtab))$positive
    # [1] "A"
    
  • Using F1 score metric in KNN through caret package
    library(caret)
    library(mlbench)
    data(Sonar)
    
    f1Summary <- function(data, lev = NULL, model = NULL) {
      # Calculate precision and recall
      precision <- posPredValue(data$pred, data$obs, positive = lev[1])
      recall <- sensitivity(data$pred, data$obs, positive = lev[1])
      
      # Calculate F1 score
      f1 <- (2 * precision * recall) / (precision + recall)
      
      # Return F1 score
      c(F1 = f1)
    }
    
    # Create a trainControl object that uses the custom summary function
    trainControl <- trainControl(summaryFunction = f1Summary)
    
    # Use the trainControl object when calling the train() function
    model <- train(Class ~., data = Sonar, method = "rf", 
                   trControl = trainControl, metric = "F1")
    > model
    Random Forest
    
    208 samples
     60 predictor
      2 classes: 'M', 'R'
    
    No pre-processing
    Resampling: Bootstrapped (25 reps)
    Summary of sample sizes: 208, 208, 208, 208, 208, 208, ...
    Resampling results across tuning parameters:
    
      mtry  F1
       2    0.8416808
      31    0.7986326
      60    0.7820982
    
    F1 was used to select the optimal model using the largest value.
    The final value used for the model was mtry = 2.
    
  • Examples in papers

PPV/Precision

The relationship between PPV and sensitivity depends on the prevalence of the condition being tested for in the population. If the condition is rare, a test with high sensitivity may still have a low PPV because there will be many false positive results. On the other hand, if the condition is common, a test with high sensitivity is more likely to also have a high PPV.

Here we have a diagnostic test for a rare disease that affects 1 in 1000 people and the test has a sensitivity of 100% and a false positive rate of 5%:

                Test| positive  negative
-----------------------------------------
Disease present     |  1          0

Disease absent      |  50        949

The prevalence of the disease is 1/1000.

Incidence, Prevalence

https://www.health.ny.gov/diseases/chronic/basicstat.htm

Jaccard index

Calculate area under curve by hand (using trapezoid), relation to concordance measure and the Wilcoxon–Mann–Whitney test

genefilter package and rowpAUCs function

  • rowpAUCs function in genefilter package. The aim is to find potential biomarkers whose expression level is able to distinguish between two groups.
# source("http://www.bioconductor.org/biocLite.R")
# biocLite("genefilter")
library(Biobase) # sample.ExpressionSet data
data(sample.ExpressionSet)

library(genefilter)
r2 = rowpAUCs(sample.ExpressionSet, "sex", p=0.1)
plot(r2[1]) # first gene, asking specificity = .9

r2 = rowpAUCs(sample.ExpressionSet, "sex", p=1.0)
plot(r2[1]) # it won't show pAUC

r2 = rowpAUCs(sample.ExpressionSet, "sex", p=.999)
plot(r2[1]) # pAUC is very close to AUC now

Use and Misuse of the Receiver Operating Characteristic Curve in Risk Prediction

Comparison of two AUCs

DeLong test for comparing two ROC curves

Some R packages

pROC

  • https://cran.r-project.org/web/packages/pROC/index.html
    • response: a factor, numeric or character vector of responses (true class), typically encoded with 0 (controls) and 1 (cases).
    • Pay attention to the message printed on screen after calling roc(). It will show which one is control and which one is case.
  • pROC: display and analyze ROC curves in R and S+
  • ROC Curve and AUC in Machine learning and R pROC Package. Pay attention to the legacy.axes parameter.
    library(pROC)
    data(aSAH)
    str(aSAH[,c('outcome', "s100b")])
    # 'data.frame':	113 obs. of  2 variables:
    #  $ outcome: Factor w/ 2 levels "Good","Poor": 1 1 1 1 2 2 1 2 1 1 ...
    #  $ s100b  : num  0.13 0.14 0.1 0.04 0.13 0.1 0.47 0.16 0.18 0.1 ...
    roc.s100b <- roc(aSAH$outcome, aSAH$s100b)
    
    roc(aSAH$outcome, aSAH$s100b, 
      plot=TRUE,
      auc=TRUE,  # already the default
      col="green",
      lwd =4,
      legacy.axes=TRUE,
      main="ROC Curves")
    # Data: aSAH$s100b in 72 controls (aSAH$outcome Good) < 41 cases (aSAH$outcome Poor).
    # Area under the curve: 0.7314
    
    auc(roc.s100b)
    # Area under the curve: 0.7314
    
    auc(aSAH$outcome, aSAH$s100b)
    # Area under the curve: 0.7314
    

    Note: in pROC::roc() or auc(), the response is on the 1st argument while in caret::confusionMatrix(), truth/reference is on the 2nd argument.

    If we flipped the outcomes, it won't affect AUC.

    aSAH$outcome2 <- factor(ifelse(aSAH$outcome == "Good", "Poor", "Good"), 
                            levels=c("Good", "Poor"))
    roc(aSAH$outcome2, aSAH$s100b)
    # Data: aSAH$s100b in 41 controls (aSAH$outcome2 Good) > 72 cases (aSAH$outcome2 Poor).
    # Area under the curve: 0.7314
    

    Roc asah.png

  • Plot multiple ROC curves in one plot.
    roc1 <- roc()
    roc2 <- roc()
    roc3 <- roc()
    plot(roc1)
    lines(roc2, col="red")
    lines(roc3, col="blue")
    

    Another solution is to use pROC::ggroc(). The annotation was taken care automatically.

    rocobj <- roc(aSAH$outcome, aSAH$s100b)
    rocobj2 <- roc(aSAH$outcome, aSAH$wfns)
    g2 <- ggroc(list(s100b=rocobj, 
                     wfns=rocobj2, 
                     ndka=roc(aSAH$outcome, aSAH$ndka))) # 3 curves
    
  • Plot a panel of ROC plots by using facet_wrap().

ROCR

Cross-validation ROC

mean ROC curve

ROC with cross-validation for linear regression in R

Assess risk of bias

PROBAST: A Tool to Assess Risk of Bias and Applicability of Prediction Model Studies: Explanation and Elaboration 2019. http://www.probast.org/

Confidence interval of AUC

How to get an AUC confidence interval. pROC package was used.

AUC can be a misleading measure of performance

AUC is high but precision is low (i.e. FDR is high). https://twitter.com/michaelhoffman/status/1398380674206285830?s=09.

Caveats and pitfalls of ROC analysis in clinical microarray research

Caveats and pitfalls of ROC analysis in clinical microarray research (and how to avoid them) Berrar 2011

Limitation in clinical data

Sample size

Threshold

Myths about risk thresholds for prediction models

Three myths about risk thresholds for prediction models 2019

Picking a threshold based on model performance/utility

Squeezing the Most Utility from Your Models 2020

Why does my ROC curve look like a V

https://stackoverflow.com/a/42917384

Unbalanced classes

Weights

Metric

Class comparison problem

Reporting

Applications

Lessons

  • Unbalanced data: kNN or nearest centroid is better than the traditional methods
  • Small sample size and large number of predictors: t-test can select predictors while lasso cannot