A1572
Title: General adapted-threshold monitoring in discrete environments and rules for imbalanced classes
Authors: Ansgar Steland - RWTH Aachen University (Germany) [presenting]
Abstract: A framework for monitoring is studied where observed variables $X_t$ are thresholded to detect relevant changes. In the novel approach, the threshold depends on external information $Z_t$, which is used to control the detector's sensitivity. As an example, consider individualized intense-care monitoring, where clinically relevant alarm thresholds need to depend on a patients' health status. Discrete-valued $Z_t$ is studied, which splits the sample into classes, and threshold functions of $Z_t$ are derived, controlling the false alarm rate $\alpha$. A proportional threshold is proposed, which favors classes with small class probabilities, thus mitigating the imbalanced classes problem. Asymptotic theory is developed for i.i.d. and dependent learning samples. Two-stage designs are suggested, allowing the distribution of the budget in a controlled manner over an a priori partition of the sample space of $Z_t$.
