Diagnostic test accuracy in longitudinal study settings: Theoretical approaches with use cases from clinical practice

Objective In this study we evaluate how to estimate diagnostic test accuracy (DTA) correctly in the presence of longitudinal patient data (i.e., repeated test applications per patient). Study design and Setting We used a nonparametric approach to estimate sensitivity and specificity of three tests for different target conditions with varying characteristics (i.e., episode length and disease-free intervals between episodes): 1) systemic inflammatory response syndrome (n = 36), 2) depression (n = 33), and 3) epilepsy (n = 30). DTA was estimated on the levels ‘time’, ‘block’, and ‘patient-time’ for each diagnosis, representing different research questions. The estimation was conducted for the time units per minute, per hour, and per day. Results A comparison of DTA per and across use cases showed variations in the estimates, which resulted from the used level, the time unit, the resulting number of observations per patient, and the diagnosis-specific characteristics. Intra- and inter-use-case comparisons showed that the time-level had the highest DTA, particularly the larger the time unit, and that patient-time-level approximated 50% sensitivity and specificity. Conclusion Researchers need to predefine their choices (i.e., estimation levels and time units) based on their individual research aims, estimands, and diagnosis-specific characteristics of the target outcomes to make sure that unbiased and clinically relevant measures are communicated. In case of uncertainty, researchers could report the DTA of the test using more than one estimation level and/or time unit.