Discovery and Representation of Causal Relationships from a Large Time-Oriented Clinical Database: The RX Project

As a hospital physician it is impossible to escape the notion that the difficult medical problems one encounters are also being confronted by other' physicians throughout the world. It is equally apparent that without special effort one's own patient observations will not be shared with others. Without the medical literature there would be almost no meaningful shadng of experience. Medical textbooks and journals contain reports of the latest tests and treatments. from university hospitals and research centers. There are, however, definite limitations to the medical literature. First. the literature records only a miniscule and highly select portion of medical experience. Second, because of this selectivity, it may be difficult to apply the findings and recommendations in the literature to one's own patients. One serious consequence of these characteristics of the medical literature is that patients are largely overtreated. Tests and treatments are over-prescribed, and adverse effe

Contenu

1. The RX Project: An Overview.- 1.1. Introduction.- 1.1.1. Medical Databases.- 1.1.2. Time-Oriented Clinical Databases.- 1.2. Evolution of Empirical Knowledge.- 1.3. Inference from Non-Randomized Databases: The Problems.- 1.3.1. The Database Research Team.- 1.4. Causal Models: The RX Knowledge Base.- 1.4.1. Causal Models: An Overview.- 1.4.2. Causal Models: Path Analysis.- 1.4.2.1. Example: A Clinical Causal Model of Coronary Heart Disease.- 1.5. The Discovery Module.- 1.5.1. An Operational Definition of Causality.- 1.5.2. Methodology of the Discovery Module.- 1.6. The RX Knowledge Base: Its Role.- 1.6.1. Discerning Time Precedence and Association.- 1.7. The Study Module.- 1.7.1. Selection of Causal Dominators.- 1.7.2. Determination of Methods for Controlling Confounding Variables.- 1.7.3. Database Access Functions.- 1.7.4. Selection of Method of Statistical Analysis.- 1.7.5. Determination Eligibility Criteria.- 1.7.6. Statistical Analysis.- 1.7.7. Weighted Multiple Regressions.- 1.7.8. Interpretation of the Results.- 1.7.9. Incorporation of the New Causal Relationship into the KB.- 1.8. Conclusions.- 2. The Time-Oriented Database.- 2.1. Introduction.- 2.1.1. The ARAMIS Database of Rheumatology.- 2.1.2. The RX Database: A Subset of ARAMIS.- 2.2. Computer Facilities.- 2.3. The RX Database: Overview of the Logical Structure.- 2.3.1. Headers.- 2.3.2. Point-Events.- 2.3.3. Interval-Events.- 2.3.4. Internal Representation of A Patient Record.- 2.3.5. Displaying Time-Oriented Clinical Data.- 2.3.6. Attribute Schemas.- 2.3.7. Representation of Missing Values of Attributes.- 2.3.8. Attributes and Derived Variables.- 2.4. Database Implementation Issues.- 2.4.1. Indexing.- 2.4.2. Hashing.- 2.4.3. Database Access Functions: Primitives.- 2.4.4. Time-Dependent Functions.- 2.4.5. Conversion of Patient Data to Array Format.- 2.5. Summary.- 3. The RX Knowledge Base: An Overview.- 3.1. Introduction.- 3.2. Categories of Schema Properties.- 3.2.1. Database Schema Properties.- 3.2.2. Hierarchical Relationship Properties.- 3.2.3. Properties Pertaining to the Definition and Intrinsic Characteristics of 62 an Object.- 3.2.4. Properties Specifying Causal Relationships to Other Objects.- 3.2.5. Summary.- 3.3. Contents of the RX Knowledge Base.- 3.3.1. Medical Schemata.- 3.3.1.1. States.- 3.3.1.2. Actions.- 3.3.2. Statistical Schemata.- 3.3.3. Schemata for Schemata.- 3.4. Inheritance Mechanisms.- 3.5. The RX Knowledge Base: Interactive Use.- 4. The Properties and Representation of Causal Relationships.- 4.1. An Operational Definition of Causality.- 4.1.1. Time Precedence.- 4.1.2. Covariation.- 4.1.3. Nonspuriousness.- 4.1.4. Mechanism and Intervening Variables.- 4.1.5. Summary.- 4.2. Features of Individual Cause/Effect Relationships.- 4.2.1. Frequency of Occurrence.- 4.2.2. Intensity of a Causal Relationship.- 4.2.3. Direction of Relationship.- 4.2.4. Setting.- 4.2.5. Functional Form.- 4.2.6. Certainty.- 4.2.7. Summary.- 4.3. Representation of Causal Relationships.- 4.3.1. Introduction.- 4.4. Representation of Causal Links in RX.- 4.4.1. Intensity.- 4.4.2. Frequency.- 4.4.3. Direction.- 4.4.4. Interactive Display of Causal Relationships and Paths.- 4.4.5. Setting.- 4.4.6. Functional Form.- 4.4.7. Validity.- 4.4.7.1. Uses of the Validity Feature.- 4.4.8. Evidence.- 4.4.9. Machine Representation.- 4.5. AI Research on Causal Models.- 4.6. Conclusion.- 5. Derived Variables, Proxy Variables, and Time-Dependent Access Functions.- 5.1. Introduction.- 5.1.1. The Uses of Derived Variables.- 5.1.1.1. Disease Episodes and other Interval-Events.- 5.1.1.2. Proxies for Latent Causal Variables.- 5.2. The Derivation of Interval-Events.- 5.2.1. Deriving Values for Interval-Events.- 5.3. Time-Dependent Database Access Functions.- 5.3.1. Access Functions Used in the Prednisone/Cholesterol Study.- 5.3.1.1. Function:Delayed-Action.- 5.3.1.2. Function:Delayed-Effect.- 5.3.1.3. Function:Delayed-Interval.- 5.3.2. Other Access Functions.- 5.4. Latent Variables and Proxies.- 6. The Discovery Module.- 6.1. Introduction.- 6.2. The Algorithm.- 6.2.1. Correlation within Patient Records.- 6.2.2. Time Delays in Correlations.- 6.2.3. Combining Correlations Across Patients.- 6.2.4. Using the Scores to Infer Causation.- 6.3. Automated Inference: A Comparison with Other Work.- 6.3.1. Statistical Work.- 6.3.2. AI Work.- 6.3.3. RX: A Hybrid Between Statistics and AI.- 7. The Study Module.- 7.1. Overview.- 7.2. Determination of Feasibility of Study.- 7.2.1. Parsing the Hypothesis.- 7.3. Confounding Variables and Causal Dominators.- 7.3.1. Causal Dominators.- 7.3.2. Controlling Other Variables.- 7.3.2.1. Variables Related to the Cause.- 7.3.2.2. Other Influences on the Effect.- 7.4. Determination of Methods for Controlling Confounding Variables.- 7.4.1. Production Rules.- 7.4.2. Controlling Confounders.- 7.4.3. Proxies for Confounders.- 7.5. Choice of Study Design and Statistical Method.- 7.5.1. Selection of Statistical Method.- 7.6. Formatting of Database Access Functions.- 7.7. Determination of Eligibility Criteria.- 7.8. Statistical Analysis: Fitting the Model.- 7.8.1. Analysis within IDL.- 7.9. Interpretation of Results.- 7.10. Incorporation of the New Causal Relationship into the KB.- 8. Statistical Analysis of Longitudinal Data.- 8.1. The Longitudinal Model.- 8.2. Regression Analysis.- 8.2.1. Combining Data Across Patients.- 8.2.2. Summary: Combining Regression Coefficients Across Patients.- 8.2.2.1. Testing the Weighted Average.- 8.3. Adequacy of the Model.- 8.3.1. Adequacy of the Model Within Individual Patient Records.- 8.3.2. Adequacy of the Model Across Patients.- 9. Medical Results.- 9.1. Introduction.- 9.2. Effects of Prednisone.- 9.3. Effect of Prednisone on Cholesterol.- 9.4. Refinements,.- 9.4.1. Pharmacokinetic Models.- 9.4.2. Automated Examination of Subsets.- 10. Summary, Applications, Future Development.- 10.1. Introduction.- 10.2. Project Summary.- 10.3. Applicability of the RX Prbject.- 10.4. Accession of Data 189 10.4.1. Post-Marketing Surveillance of Drugs.- 10.5. The RX Project: Limitations and Future Development.- 10.5.1. Study Module.- 10.5.1.1. User Interface.- 10.5.1.2. Statistical Procedures.- 10.5.1.3. Analysis of Residuals.- 10.5.2. Knowledge Base.- 10.5.2.1. Definitions of Proxy Variables and Other Derived Variables.- 10.5.2.2. Syntax of Causal Relationships.- 10.5.3. Database.- 10.5.4. Discovery Module.