Flexible outlier detection in multicenter clinical trials

Multicenter clinical trials have become increasingly larger and more complex yet assuring high-quality data remains an essential task for data coordinating centers.

A flexible algorithm is presented to exhaustively search for potential data anomalies by participant and site. The algorithm proceeds as a three-phase collaboration between the data coordinating center and clinical sites. First, participant-level data are examined in a univariate approach for all relevant variables. Values at the extreme tails of the distribution that also lie outside range checks and are previously unverified by clinical sites are deemed potential outliers. Second, participant-level data are examined in a multivariate machine learning approach among meaningful groups of related variables, e.g., weight and body mass index. Third, site-level differences are characterized using statistical tests and standardized differences, both adjusted and unadjusted for site demographics. Findings are discussed with sites and, if needed, alterations can be made to procedures for data collection. For illustration, the algorithm is applied to data collected in the Molecular Transducers of Physical Activity Consortium (MoTrPAC) study.

Application of the algorithm to MoTrPAC yielded an evaluation of over 1.9 million observations in n = 1029 study participants. Numerous individual univariate, multivariate, and site-level outliers were identified that were previously unidentified by existing data evaluation processes.

It is recommended to apply this algorithm to a subset of participants early in a study, with repeated explorations over subsequent intervals throughout the study. The goal is to maximize data quality, particularly critical to the increasing occurrence of open-source, data resources.