Intelligent alarms : allocating attention among concurrent processes

I have developed and evaluated a computable, normative framework for intelligent alarms: automated agents that allocate scarce attention resources to concurrent processes in a globally optimal manner. My approach is decision-theoretic, and relies on Markov decision processes to model time-varying, stochastic systems that respond to externally applied actions. Given a collection of continuing processes and a specified time horizon, my framework computes, for each process: (1) an attention allocation, which reflects how much attention the process is awarded, and (2) an activation price, which reflects the process''s priority in receiving the allocated attention amount. I have developed a prototype, Simon, that computes these alarm signals for a simulated ICU. My validity experiments investigate whether sensible input results in sensible output. The results show that Simon produces alarm signals that are consistent with sound clinical judgment. To assess computability, I used Simon to generate alarm signals for an ICU that contained 144 simulated patients; the entire computation took about 2 seconds on a machine with only moderate processing capabilities. I thus conclude that my alarm framework is valid and computable, and therefore is potentially useful in a real-world ICU setting.