Reference: Shahar, Y. A Knowledge-Based Method for Temporal Abstraction of Clinical Data. Ph.D. Thesis, Knowledge Systems Laboratory, Medical Computer Science, 1994.
Abstract: This dissertation describes a reasoning framework for knowledge-based systems, that is specific to the task of abstracting higher-level concepts from time-stamped data, but that is independent of any particular domain. I specify the theory underlying the framework by a logical model of time, parameters, events, and contexts: a knowledge-based temporal-abstraction theory. The domain-specific knowledge requirements and the semantics of the inference structure that I propose are well defined and can be instantiated for particular domains. I have applied my framework to the domain of clinical medicine. My goal is to create, from primary time-stamped patient data, interval-based temporal abstractions, such as "severe anemia for 3 weeks in the context of administering the drug AZT," and more complex patterns, involving several such intervals. These intervals can be used for planning interventions for diagnostic or therapeutic reasons, for monitoring plans during execution, and for creating high-level summaries of electronic medical records. Temporal abstractions are also helpful for explanation purposes. Finally, temporal abstractions can be a useful representation for comparing a therapy planner's recommendation with that of the human user, when the goals in both plans can be described in terms of creation, maintenance, or avoidance of certain temporal patterns. I define a knowledge-based temporal-abstraction method that decomposes the task of abstracting higher-level, interval-based abstractions from input data into five subtasks. These subtasks are then solved by five separate, domain-independent, temporal-abstraction mechanisms. The temporal-abstraction mechanisms depend on four domain-specific knowledge types. The semantics of the four knowledge types and the role they play in each mechanism are defined formally. The knowledge needed to instantiate the temporal-abstraction mechanisms in any particular domain can be parameterized and can be acquired from domain experts manually or with automated tools. I present a computer program implementing the knowledge-based temporal-abstraction method: RESUME. The architecture of the RESUME system demonstrates several computational and organizational claims with respect to the desired use and representation of temporal-reasoning knowledge. The RESUME system accepts input and returns output at all levels of abstraction; generates context-sensitive and controlled output; accepts and uses data out of temporal order, modifying a view of the past or of the present, as necessary; maintains several possible concurrent interpretations of the data; represents uncertainty in time and value; and facilitates its application to additional domains by editing only the domain-specific temporal-abstraction knowledge. The temporal-abstraction knowledge is organized in the RESUME system as three ontologies (domain-specific theories of relations and properties) of parameters, events, and interpretation contexts, respectively, in each domain. I have evaluated the RESUME system in the domains of protocol-based care, monitoring of children's growth, and therapy of insulin-dependent diabetic patients. I have demonstrated that the knowledge required for instantiating the temporal-abstraction mechanisms can be acquired in a reasonable amount of time from domain experts, can be easily maintained, and can be used for creating application systems that solve the temporal-abstraction task in these domains. Understanding the knowledge required for abstracting clinical data over time is a useful undertaking. A clear specification of that knowledge, and its representation in an ontology specific to the task of abstracting concepts over time, as was done in the architecture of the RESUME system, supports designing new medical and other knowledge-based systems that perform temporal-reasoning tasks. The formal specification of the temporal-abstraction knowledge also supports acquisition of that knowledge from domain experts, maintenance of that knowledge once acquired, reusing the problem-solving knowledge for temporal abstraction in other domains, and sharing the domain-specific knowledge with other problem solvers that might need access to the domain's temporal-reasoning knowledge.