You must understand alarm fatigue before building a monitoring tool which uses AI.
What is alarm fatigue? Is it how you feel as you hit snooze on an alarm that was set too optimistically the night before? Well, maybe, but in the case of your alarm clock, there is only one specific condition under which it goes off - the time of day. If your alarm clock went off at other times, for no good reason, you’d quickly develop alarm fatigue. You’d also get rid of this alarm clock. It’s a wonder car alarms still exist.
Alarm fatigue is the detrimental effect of automated alerts on anyone who experiences them. They occur in a wide variety of settings, but have been experienced by most of us in a healthcare setting, whether as a patient or visitor. Alarm fatigue is also a factor in industrial settings, mining, utility control, aviation and automated driver assistance systems.
In healthcare, the negative effects of alarm fatigue on both patients and care staff is well documented. However, for a variety of reasons, clinicians remain powerless in adjusting these systems. No accommodations are made for user selectable thresholds, frequency and silencing periods. This is because this would modify the risk profile under which these systems have obtained their regulatory status. This is despite known adverse effects on quality of care and the recognized phenomenon of alarm pollution, which causes desensitisation by diluting important alerts amongst inconsequential ones.
A similar burden is posed by similar systems across industries and settings and even on a consumer facing level, from a smoke detector that’s a tad too sensitive for your kitchen to alerts and nudges in vehicles with automated driver assistance systems. Those who design these systems often seek to balance their exposure to liability with the optimal operating range for the users of the system.
Yet, those burdened by this effect will attest to the burden imposed on them. They often also evolve their attention using an unintended paradigm. That is, through experience, they develop an emergent recognition of the patterns and frequencies of the false alarms which then serves as the baseline from which to make decisions. This is a common scenario which strays very far from the designed application.
To be sure, we often get this wrong for the case of traditional, rules-based, monitoring systems.
What will come of this effect as we inject the non-deterministic nature of AI into such systems?