Architecture for a Risk Processor. C. D. Carrington, US FDA, 200 C Street SW, Washington, DC 20204
Translation of a data-driven procedure for answering questions of risk into a computer program may be expected to result in the following benefits: a) The process may be more complex and therefore more adaptable to the available data; b) the results will be reproducible, and differences between assessments will be attributable to differences in the available data rather than the process; c) expertise may be directed to specific components of the process; d) all presumptions would be explicit and (with open code) subject to open evaluation; e) continuous refinement of the process to make it more accurate and more responsive to the needs of the decisionmaker would be facilitiated. The following specifications for risk processor are proposed. The risk processor should respond to the decision maker's need for information and any and all available data. Specifically, the processor should describe, rather than manage, the relationship between a potentially regulated activity and x) the magnitude of the effect that may occur in an individual, y) the variability in the magnitude of the effect that may be anticipated in a population (heterogeneity), and z) the magnitude of the uncertainty associated with the estimate (error). The risk processor should produce numerical and graphical output that describe the three dimensions of risk, and should identify areas where additional information would be useful for reducing uncertainty. To accomplish these goals, the following architecture is proposed: 1) An Input procedure in which a list of questions which can currently be answered by the risk processor are drawn up for the decision maker to select. The critical part of this procedure is Model Design, in which a scheme for calculating the risk is assimilated. The list of available questions will be contingent on the available data and analytical procedures at the disposal of the risk processor. 2) An Analytical procedure in which two- (for parameters) or three dimensional (for cause-effect relationships) models are optimized to describe the available data. Analytical routines will generate representations of either single or multiple data sets which include separate descriptions of population variability and uncertainty. 3) A Synthetic procedure in which the components are assembled into a predictive model. A two-dimensional Monte-Carlo procedure may be used to separately integrate the distributions in each component describing population heterogeneity and uncertainty. 4) An Output procedure in which the three dimensional output is presented to the decision maker either in the form of numerical output or as a visual display. The output procedure should be able to adapt to the technical competence of the decision maker either by reducing the amount of information presented or by educating the decision maker.