|
Abstract:
|
Patient data from clinical trials exhibit dynamics both (i) from top-down rules of the intended protocol and (ii) natural variability in patients, time, and clinical settings. Rules-based algorithmic monitoring of clinical trial data provides a foundation for detecting anomalous deviations from the trial protocol, but suffers because sensitivity and specificity of detection is heavily dependent on a priori definitions of the monitoring rules. Statistical monitoring allows greater flexibility by replacing dichotomized thresholds with continuities over distributions. Furthermore, probabilistic inference provides a principled manner in which to reason when conditioning data distributions, ex. with respect to a specific patient, time, or covariate. We illustrate how patient-specific data can enrich rules-based systems by ensuring the rules are built with greater clinical personalized time-dependent context. Next we demonstrate how building patient-specific models can identify poor quality data that would be missed by population-based statistics or rule-based methods. These personalized models can be visualized and interpreted by nontechnical experts monitoring the trial.
|
