Real-time Alarm Model Adaptation Based on User Feedback

Abstract

Zie ook de paper hieronder in ditzelfde tijdschrif/boek. This paper presents real-time adaptation of alarm detection models in a remote health monitoring system based on user feedback. Real-time adaptation enables systems to fine-tune to the needs and preferences of the user and changes in the environment. This way, the system performance is improved in terms of technical accuracy and subjective user wishes. Two types of alarm detection models were used: (1) models in the form of rules created by domain expert and (2) models induced by machine learning. The problem of adaptation for the rule-based models is defined as Markov decision process. Machine-learning models are adapted by rebuilding the model every time new data is obtained. We tested the adaptation capabilities of the two types of alarm detection models based on their accuracy and time-to-alarm (needed length of possibly critical activity, such as lying on the ground, which causes the models to raise an alarm). Both types of models achieved 90% alarm detection accuracy. The rule-based models decreased time-to-alarm when user-triggered alarms were raised and increased it when the user indicated false alarm. We did not observe this process for the machine-learning models.