Robert L. Blum

Discovery, confirmation, and incorporation of causal relationships from a large time-oriented clinical data base: The RX project

Abstract The objectives of the methods and computer implementation presented here are (1) to automate the process of hypothesis generation and exploratory analysis of data in large nonrandomized, time-oriented clinical data bases, (2) to provide knowledgeable assistance in performing studies on large data bases, and (3) to increase the validity of medical knowledge derived from nonprotocol data. The RX computer program consists of a knowledge base (KB), a discovery module, a study module, and a clinical data base. Utilizing techniques from the field of artificial intelligence, the KB contains medical and statistical knowledge hierarchically organized, and is used to assist in the discovery and study of new hypotheses. Confirmed results from the data base are automatically encoded into the KB. The discovery module uses lagged, nonparametric correlations to generate hypotheses. These are then studied in detail by the study module which automatically determines confounding variables and methods for controlling their influence. In determining the confounders of a new hypothesis the study module uses previously “learned” causal relationships. The study module selects a study design and statistical method based on knowledge of confounders and their distribution in the data base. Most studies have used a longitudinal design involving a multiple regression model applied to individual patient records. Data for system development were obtained from the American Rheumatism Association Medical Information System.

Inferring Knowledge From Clinical Data Banks Utilizing Techniques From Artificial Intelligence

This paper introduces the RX Project, a software design for facilitating the induction and storage of medical knowledge from clinical data banks. The RX Project is an attempt to bridge the gap between clinical data bank systems and knowledge-based medical consultation systems. By integrating these two approaches our expectation is that the benefits of both may be realized and that limitations may be overcome that are present in data bank systems which use only statistical techniques and in knowledge-based methodologies which use mainly artificial intelligence techniques. The RX Project is an outgrowth of two existing computer systems at Stanford Medical Center: ARAMIS Project (American Rheumatism Association Medical Information System) and the MYCIN Project, a knowledge-based system which advises physicians on antimicrobial therapy. We shall first discuss each of these projects individually as examples of 1) statistical and 2) symbolic approaches to medical consultation systems, respectively, in order to present their relative strengths and limitations. Then we shall present an overview of the RX Project architecture, which integrates these methodologies.

Acquisition of knowledge from data

The work described here addresses two problems: information overload of database users, and knowledge acquisition for use in Al systems. We have implemented programs that use artificial intelligence techniques to prepare high-level, intelligent summaries of databases, and that use empirical databases in turn, in combination with statistical and Al methods, to generate new domain knowledge base. Both programs are examples of the aquisition of knowledge from data: the Summarization Module fuses large amounts of data succinctly, the Discovery Module extracts new knowledge present implicitly in data. We describe the implementation of our programs and outline planned extensions which combine both approaches. This work is distinguished from current knowledge engineering approaches in that we prime the system with expert knowledge, and then use factual data to learn more about the domain.

Displaying clinical data from a time-oriented database

This paper details the principles and implementation of a time-oriented clinical database and a program for displaying data on individual patients as graphs, line drawings of episodes and scatter diagrams. The article presents efficient data structures for storage of graphs in which plots of multiple variables are overlayed. The display program permits plotting of patient data either by visit or by time, telescoping of the time axis, and changing of the display size. Line drawings of concurrent clinical episodes may be aligned beneath the graphs. No special graphics equipment is required. Issues of data representation, transformation, and display layout are discussed.

Studying Hypotheses on a Time-Oriented Clinical Database: An Overview of the RX Project

The RX computer program examines a time-oriented clinical database and attempts to derive a set of (possibly) causal relationships. First, a Discovery Module uses lagged, nonparametric correlations to generate an ordered list of tentative relationships. Second, a Study Module uses a small knowledge base (KB) of medicine and statistics to create a study design to control for known confounders. The study design is then executed by an on-line statistical package, and the results are automatically incorporated into the KB as a machine-readable record. In determining the confounders of a new hypothesis the Study Module uses previousbly “learned” causal relationships.

An integration of knowledge and data representation

A variety of types of linkages from knowledge bases to databases have been proposed, and a few have been implemented [MW84]. In this research note, we summarize a technique which was employed in a specific context: knowledge extraction from a copy of an existing clinical database. The knowledge base is also used to drive the extracting process. RX builds causal models in its domain to generate input for statistical hypothesis verification. We distinguish two information types: knowledge and data, and recognize four types of knowledge: categorical, definitional, causal (represented in frames), and operational, represented by rules. Based on our experience, we speculate about the generalization of the approach.

Statistical Analysis of Longitudinal Data

The objective of the Study Module is to formulate a statistical model of a hypothesis using medical and statistical knowledge derived from the knowledge base. The hypothesis may have been obtained either manually from a researcher or automatically from the Discovery Module. Once the statistical model has been formulated, a statistical package is automatically invoked, and the model is tested on the appropriate set of data from the database.

Minimycin: A Miniature Rule-Based System

Artificial intelligence (AI) systems, in which computers perform tasks that would show intelligence if performed by a human, are already in frequent clinical use in a number of hospitals [1–2, 7]. Articles on artificial intelligence have appeared in the New England Journal of Medicine [3] and the Journal of the American Medical Association [4]. (Good collections on AI and medicine include [5] and [6].)

The Discovery Module

The objective of the Discovery Module is to find candidate causal relationships. It is the most recent addition to the RX program; hence, its structure is apt to be expanded and improved over the next few years. In this chapter I will present its current structure, some preliminary experimental results, and designs for the future.

The LISP Programming Language and Environments

The LISP language provides a useful set of features for prototyping knowledge- intensive, clinical applications software that is not found in most other programming environments. Medical computer programs that need large medical knowledge bases—such as programs for diagnosis, therapeutic consultation, education, simulation, and peer review—are hard to design, evolve continually, and often require major revisions. They necessitate an efficient and flexible program development environment.