gms | German Medical Science

Introduction: In recent years, a variety of methods for analyzing diagnostic trials have been developed. Hereby different measures of diagnostic accuracy such as sensitivity, specificity or the area under the ROC-Curve (AUC) are most commonly used to assessing diagnostic performance. Hence a multitude of approaches exist to analyze these most common assessments of diagnostic accuracy even in complex study designs [1], [2], [3], [4]. More recently the partial area under the curve (pAUC) has come to the fore and is becoming increasingly common. However, there still is limited research investigating analysis methods for the pAUC especially when considering more complex designs. To overcome these shortcomings, we present a new perspective on the pAUC and thus make it possible to apply the vast array of existing analysis methods to pAUC studies.

Methods: Since most diagnostic tests used in clinical practice require a minimum sensitivity and a minimum specificity some parts of the ROC-curve are less relevant than others. This consideration leads to the concept of the partial area under the curve (pAUC). Here only those parts of the ROC-curve are taken into account where specificity and sensitivity exceed certain thresholds – the other parts of the ROC-curve are not relevant. This leads to the fact that only certain parts of the original distribution functions are relevant for the determination of the pAUC while the other parts of these distribution functions can be completely neglected. We show that by editing the irrelevant parts of the distribution functions, the pAUC can be viewed as a special form of the usual AUC and thus, under certain assumptions, the pAUC can be analyzed using AUC analysis methods.

Results: By showing that the pAUC can be interpreted as a special form of the AUC, we open the toolbox of (nonparametric) AUC analysis methods to the pAUC, giving access to a large number of analysis methods for the pAUC, such as methods for factorial designs [1], clustered data [2], [3], and adjustment for covariates [4]. Existing software packages can be used.

Discussion: The major advantage of the presented analysis approach is its broad applicability even in complex study designs. Thus, we fill an important gap since our methodology provides analysis tools for the pAUC for some study designs for the first time (to the best of our knowledge, e.g., for covariates [4] or clustered data [2], [3]). However, for certain study designs, there are evaluation procedures that have been developed specifically for the pAUC. In these cases, careful consideration should be given beforehand as to which method is most appropriate.

Conclusion: Since the pAUC is a specific AUC, it can be analyzed using the same methodology and software as the AUC under certain conditions.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.