DxGoals: A Software Tool for Determining and Analyzing Clinically Meaningful Classification Accuracy Goals for Diagnostic Tests

Abstract

Abstract Background To evaluate diagnostic tests for low prevalence conditions, classification accuracy metrics such as sensitivity, specificity, and positive likelihood ratio (PLR) and negative likelihood ratio (NLR) are advantageous because they are prevalence-independent and thus estimable in studies enriched for the condition. However, classification accuracy goals are often chosen without a clear understanding of whether they are clinically meaningful. Pennello (2021) proposed a risk stratification framework for determining classification accuracy goals. A software application is needed to determine the goals and provide data analysis. Methods We introduce DxGoals, a freely available, R-Shiny software application for determining, visualizing, and analyzing classification accuracy goals for diagnostic tests. Given prevalence p for the target condition and specification that a test's positive and negative predictive values PPVand NPV=1−cNPV should satisfy PPV>PPV* and cNPV<cNPV*, DxGoals uses Bayes Theorem to determine equivalent goals for PLR and NLR and implied goals for sensitivity and specificity. When study data are provided, DxGoals analyzes whether the determined goals are met with statistical significance. When comparing 2 tests, DxGoals translates a superiority or noninferiority goals for the differences PPV−p and p−cNPV to equivalent goals for PLR and NLR and analyzes the goals when data are provided. Results We illustrate DxGoals on tests for penicillin allergy, ovarian cancer, and cervical cancer. The inputs cNPV*,p, and PPV* were informed by clinical management guidelines. Conclusions DxGoals facilitates determination, visualization, and analysis of clinically meaningful standalone and comparative classification accuracy goals. It is a potentially useful tool for diagnostic test evaluation.