Evaluating Binary Diagnostic Tests with EBDT

Authors: Miguel Angel Montero-Alonso & Juan de Dios Luna del Castillo.
Package: ebdt v1.0.1

Introduction

The ebdt package evaluates the quality of a binary diagnostic test under complete verification. It computes point estimates and confidence intervals for sensitivity, specificity, Youden index, positive and negative predictive values, positive and negative likelihood ratios, weighted kappa coefficient, and disease prevalence, for both, cross-sectional (prospective) and retrospective study designs.

Reference:

Quick Start

library(ebdt)
ebdt(s1 = 40, r1 = 5, s0 = 10, r0 = 45)

The 2x2 Table

A binary diagnostic test produces a 2x2 contingency table:

• TP (s1): True positives — diseased correctly classified as positive
• FP (r1): False positives — non-diseased incorrectly classified as positive
• FN (s0): False negatives — diseased incorrectly classified as negative
• TN (r0): True negatives — non-diseased correctly classified as negative

Study Types

Cross-sectional / Prospective (study = TRUE)

All quality parameters are estimable:

ebdt(40, 5, 10, 45, study = TRUE, conflev = 0.95)

Computed: Sensitivity, Specificity, Youden Index, Prevalence, PPV, NPV, PLR, NLR, Weighted Kappa.

Retrospective / Case-Control (study = FALSE)

Only Se, Sp, Youden, PLR, and NLR are estimable (prevalence, PPV, NPV not defined under retrospective sampling):

ebdt(40, 5, 10, 45, study = FALSE, conflev = 0.95)

Individual Metrics

Sensitivity

ebdt_se(40, 5, 10, 45)

Specificity

ebdt_sp(40, 5, 10, 45)

Youden Index

ebdt_you(40, 5, 10, 45)

Positive Predictive Value (PPV)

ebdt_ppv(40, 5, 10, 45)

Negative Predictive Value (NPV)

ebdt_npv(40, 5, 10, 45)

Positive Likelihood Ratio (LR+)

ebdt_plr(40, 5, 10, 45)

Two confidence intervals are provided: Simel (log-normal, delta method on log LR+) and Gart & Nam (exact, inverting the score statistic). The narrower interval is recommended.

Negative Likelihood Ratio (LR-)

ebdt_nlr(40, 5, 10, 45)

Prevalence (cross-sectional only)

ebdt_prev(40, 5, 10, 45)

Weighted Kappa Coefficient

ebdt_kap(40, 5, 10, 45)

Kappa is computed for c = 0.1, 0.2, ..., 0.9. Both Wald and Logit-transformed confidence intervals are provided.

Confidence Interval Methods

Zero-Cell Correction

When any cell of a related pair contains a zero, the Haldane-Anscombe continuity correction (+0.5) is applied to both cells of that pair:

• If s1 == 0 or s0 == 0: add 0.5 to both s1 and s0
• If r1 == 0 or r0 == 0: add 0.5 to both r1 and r0

A warning is issued when the correction is applied.

ebdt(0, 5, 10, 45, quiet = TRUE)

Return Values (Programmatic Use)

All functions return structured lists in addition to printing to the console:

res <- ebdt_se(40, 5, 10, 45)
res$est        # point estimate
res$ci_lower   # lower CI bound
res$ci_upper   # upper CI bound
res$ci_method  # "Wilson" or "Agresti-Coull"

res_plr <- ebdt_plr(40, 5, 10, 45)
res_plr$gartnam_ci_lower
res_plr$gartnam_ci_upper

res_kap <- ebdt_kap(40, 5, 10, 45)
head(res_kap, 3)

Function Return Elements

Shiny App

The package includes an interactive Shiny application:

shiny::runApp(system.file("app", package = "ebdt"))

and you can also find it at:

https://migmontal.shinyapps.io/ebdt/

Features: manual entry of TP/FP/FN/TN counts, Excel (.xlsx) upload with 2x2 table auto-population, study type selection, individual metric or all-metric mode, and text file export.

Reading from Excel

ebdt_from_excel("test_ebdt.xlsx")
ebdt_from_excel("test_ebdt.xlsx", study = FALSE, conflev = 0.99)

The Excel file must contain a 2x2 numeric matrix in the top-left corner:

# Row 1: TP  FP
# Row 2: FN  TN

GitHub

All content and source code on this site is available in our GitHub repository.

Package ebdt v1.0.1 — Universidad de Granada