Abstract Explanations of Strategy in a Diagnostic Consultation System

Diane Warner HaslingAAAI 1983

Abstract

NEOMYCIN is a medical expert system designed to provide abstract, principled explanations of its diagnostic strategies. By decoupling strategic metaknowledge from domain-specific facts, it addresses the transparency limitations of earlier systems like MYCIN. This paper details the task-metarule architecture that enables NEOMYCIN to explain not just its conclusions, but the underlying diagnostic plan, serving as a significant advancement for medical education and system debugging.

The Transparency Problem in Expert Systems

Early expert systems, most notably MYCIN, were revolutionary in their ability to perform medical diagnosis. However, they suffered from a significant "transparency" problem. While they could provide a final diagnosis, their internal reasoning process—the strategy governing why certain questions were asked in a specific order—was often implicit. This strategic knowledge was frequently buried in the arbitrary ordering of rules or hard-coded into the inference engine.

For medical students, simply seeing a list of triggered rules is insufficient for learning. They need to understand the high-level diagnostic "plan," such as the intent to differentiate between two infectious processes. NEOMYCIN was developed to make this strategic reasoning explicit, general, and explainable.

Architecture: Separating Strategy from Domain Knowledge

The fundamental innovation in NEOMYCIN is the explicit separation of strategic knowledge from domain-specific medical facts. This is achieved through a hierarchical structure of tasks and metarules.

Hierarchical task decomposition in NEOMYCIN's diagnostic process.

As shown in the task hierarchy, NEOMYCIN decomposes high-level goals like Make-Diagnosis into sub-tasks such as Identify-Problem, Collect-Information, and Explore-and-Refine. These tasks are achieved via "metarules"—methods that dictate how to proceed without referencing specific medical terms. This domain independence makes the strategy abstract and potentially portable to other fields.

The Mechanics of Explanation: Why and How

NEOMYCIN maintains an audit trail of task calls and successful metarules. This allows it to answer user queries based on a principled mapping:

  • WHY: When a user asks "WHY" a question is being asked, the system cites the premise of the metarule currently being evaluated (the reason the task was triggered).
  • HOW: When a user asks "HOW" a goal is achieved, the system describes the actions taken by the metarules associated with that task.

By providing both concrete explanations (referencing specific patient data) and abstract explanations (referencing general diagnostic principles), the system helps users recognize patterns that apply across different cases, facilitating learning by analogy.

Comparative Performance and Results

The primary success of NEOMYCIN lies in its ability to generate multi-level explanations that mirror human diagnostic behavior. Unlike MYCIN, which was limited to backward-chaining through domain rules, NEOMYCIN provides a rich "forward-reasoning" context.

Comparison of reasoning and explanation mechanisms between MYCIN and NEOMYCIN.

As summarized in the comparison above, NEOMYCIN shifts the focus from simple goal-rule-subgoal chains to a more sophisticated task-metarule-subtask framework. This allows the system to explain why a task is done by citing the invoking metarule and how it is accomplished by citing the achieving metarules.

Limitations and Future Work

Despite its advancements, NEOMYCIN faces challenges regarding "information overload." Providing exhaustive detail for every task can overwhelm users, suggesting a need for discourse rules to determine the appropriate "grain size" for explanations. Additionally, while the abstract representation enables analogy, the current system does not automatically generate comparative case summaries. Future work involves refining these discourse strategies and ensuring that all control logic is strictly maintained within the task/metarule structure to preserve explanation integrity.

Summary updated

BibTeX
@inproceedings{hasling1983abstract,
  author = {Diane Warner Hasling},
  title = {{Abstract Explanations of Strategy in a Diagnostic Consultation System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure