Model-based interpretation of time-ordered medical data

Time-ordered data often contain implications that physicians infer only by recognizing subtle temporal patterns. This dissertation is concerned with representing and using medical knowledge about temporal relationships so that decision-support programs can deduce abstract clinical concepts that appear as temporal patterns in time-ordered clinical observations. Computational models of time that are based on a single representational method are limited by the expressive power of the chosen method. I propose that combining multiple temporal formalisms facilitates the creation of robust temporal models. The use of multiple modeling techniques permits each method to represent those temporal relationships that are best expressed within a particular formalism. This dissertation proposes a methodology for interpreting time-ordered clinical data using a combination of numeric and symbolic techniques. The methodology proposes three transformations: (1) model-based transformation of time-ordered observations into underlying concepts, (2) interval-based transformation of dynamic model predictions into domain-specific temporal abstractions, and (3) state-based transformation of temporal abstractions into clinically meaningful discourse fragments. The temporal model used to convert observations into underlying system (patient) concepts encodes physiological process knowledge. The temporal model used to convert model-based predictions into temporal abstractions encodes domain-specific event knowledge. These two temporal models differ in the temporal structure they assume and in the temporal concepts they encode. Without multiple temporal models, this diversity of temporal knowledge could not be represented. Effective medical decision making demands that physicians recognize and exploit the complex temporal changes that permeate the medical record. In this dissertation, I examine the temporal features of time-ordered clinical data in the context of summarizing a patient's medical record. I have demonstrated the three-transformation methodology in a computer program, call TOPAZ, that summarizes the clinical course of individual patients who are receiving experimental cancer chemotherapy.