Model-Based Interpretation of Time-Ordered Data

Reference: Kahn, M. G. Model-Based Interpretation of Time-Ordered Data. Working Paper, March, 1987.

Abstract: A patient's medical record is often cited as the single most important source of information for making clinical decisions. This document chronicles the individual's history of disease and response to therapy. Many important clinical features, such as the progression of disease symptoms or the development of chronic toxicities, are expressed only as a sequence of observations over time. Since the clinical environment is often complex and changing due to interacting pathologic and therapeutic factors, recorded measurements and observations must be interpreted from the clinical context that existed when they were observed. Thus the task of reviewing and summarizing a patient's medical record is a necessary but difficult step for the physician faced with understanding a patient's current clinical state. Many fields are faced with large quantities of time-ordered data that must be interpreted to understand the important concepts that are implied by the observations. Abstraction, the task of selecting only those features that are most relevant to answering a question, and summarization, the task of combining multiple observations or features into a more general statement, require extensive knowledge about the entities and relationships that exisit between the observations. For example, serum creatinine and serum BUN are both measures of renal function; an elevation in one of these items should be associated with a rise in the other item. An increase in either measurement implies renal dysfunction. Temporal trends such as increasing, decreasing, or fluctuating values may have a significant impact on the expected value of other entities or on the conclusions drawn from the observations. An abstract representation of the presumed relationships between data elements is called a model. A model provides a set of organizing principles to guide the interpretation process by limiting the nuber of relationships that are assumed to be present in the observations. This thesis proposes that models can be used to construct useful summaries of complex data. In particular, models that capture the temporal nature of the observed system can be used to interpret time-ordered observations. The atemporal setting is viewed as a degenerate case where only a single temporal order is embodied in the model. A general methodology for summarizing time- ordered data is presented. The key element in the methodology is a sequence of models that are created as new observations are incorporated into an initial model of the observed system. These differences are used to detect important changes and unusual differences that are implied by the data. The methodology emphasizes four main elements in summarizing time-ordered data: 1. Starting with an initial model based only on genreal knowledge, a series of new models is generated by incorporating new observations. The collection of altered models captures state changes that evolve over time. 2. Differences between models detect changes in states or unusual features of the observed system. Model differences form the basis for abstraction and summarization. 3. A motivating question, provided by the user, guides the selection of model differences that are pertinent for inclusion in the summarization. The abstraction process uses this information to examine only those model differences that are potentially relevant to answering the user's question. 4. Additional contextual information modifies the summarization process so that the style and content of the generated summary text conforms to the user's expectations and requirements. The methodology will be demonstrated by implementing a computer program that summarizes the clinical course of individual patients receiving experimental cancer chemotherapy. In this context, a medical record contains "the observations" and a particular patient is "the observed system". The goal of the program is to produce text output that is similar to a treating oncologist summarizing the patient's clinical course for a senior consultant.

Notes: 18 pages.

Jump to... [KSL] [SMI] [Reports by Author] [Reports by KSL Number] [Reports by Year]
Send mail to: ksl-info@ksl.stanford.edu to send a message to the maintainer of the KSL Reports.