D-PDDM provably monitors model deterioration requiring no training data during deployment, and performs well in real-worlds datasets.