Monitoring diseases with empirical and model-generated histories

摘要

Diagnostic monitoring systems track disease hypotheses over time, symbolically interpreting the time-varying patient data produced by medical instrumentation. The need to track multiple interacting diseases recommends a hypothesize, test and refine reasoning architecture which incorporates a robust knowledge representation. Rule-based systems are inadequate, and deep or model-based representations capable of first principles reasoning are currently favoured. However, the model-based approach may be too low level for many monitoring tasks. While disease interactions may present novel patterns to a monitor, usually the diseases themselves will be familiar. It is proposed that disease histories generated from pathophysiological models are at an appropriate level of abstraction for many monitoring tasks. Histories lie between disease models and rules in depth. Using the QSIM representation, results are presented for model-generated histories that define some limits of their utility in reasoning systems. In particular, if the underlying system is non-linear then restrictions exist on the predictions possible with histories alone. These results are extended to poorly modelled domains which may be tractable to reasoning with empirically derived histories. As a consequence, we can also specify when a monitoring system must switch from history to model-based representations.