Introduction
In the development of medical diagnostic systems during the late 1980s, researchers faced a significant trade-off between efficiency and robustness. "Compiled" knowledge systems, which utilized simple if-then rules, were computationally fast but fragile when encountering complex or novel cases. In contrast, "causal models" provided deep, robust explanations by simulating physiological states but were computationally expensive, often requiring the examination of tens of thousands of states to reach a conclusion.
Phyllis Koton addressed this efficiency gap by introducing CASEY, a system that maintains the depth of a causal model while operating at the speed of a rule-based system. CASEY achieves this by leveraging Case-Based Reasoning (CBR), using memory to recall similar past experiences and focusing computational effort only on the differences between cases.
Methodology: The CASEY Architecture
CASEY operates as an intelligent layer atop the "Heart Failure" program, a large-scale causal model of cardiac disease. The system follows a five-step process to generate diagnostic explanations:
- Retrieval: The system searches its memory for a previous patient (a "precedent") with similar clinical features.
- Justification: This core innovation uses evidence principles to evaluate differences between the new patient and the precedent. It determines if a difference is significant or if it falls within the same "qualitative region" (e.g., comparing heart rates that are both categorized as "moderately high").
- Adaptation: If differences are deemed "repairable," CASEY applies repair strategies to modify the old causal graph—adding or removing nodes and links—rather than rebuilding the entire model from scratch.
- Storage: The new case and its adapted solution are stored to facilitate future reasoning.
- Feature Evaluation: The system identifies which features were critical to the final explanation to prioritize them in future searches.
Justification and Evidence Principles
The "Justifier" acts as a filter. By proving that a new problem can be explained by a modified version of an old solution, the system avoids redundant computation.
As shown in the comparison above, CASEY identifies differences in features like mean arterial pressure and pulse. If the Justifier fails to reconcile these differences using domain-independent rules (such as "Rule out" or "Same qualitative region"), the system falls back on the exhaustive causal model.
Experimental Evaluation and Results
The system was evaluated using the Heart Failure Model and a set of 45 test cases provided by medical professionals, covering approximately 15 different diseases.
Performance and Accuracy
The most striking result was the massive speedup in processing. In a head-to-head comparison for a patient named "Newman," CASEY produced an explanation by examining only 674 states. In contrast, the original Heart Failure program examined approximately 76,000 states to reach the same conclusion—a performance increase of over two orders of magnitude.
CASEY produced successful or satisfactory explanations for 86% of the test cases where a similar precedent existed in memory. Even in cases where a perfect match was impossible, CASEY often performed the bulk of the diagnostic work, leaving only a few unexplained features for the more intensive program to handle.
Conclusion
CASEY demonstrates that reasoning by analogy is significantly more efficient than reasoning from first principles in complex domains. By utilizing domain-independent evidence principles, the system provides a framework that could theoretically be applied to other fields governed by causal networks, such as mechanical troubleshooting or geology. While the current version does not fully integrate quantitative probabilities, its ability to learn through use—becoming more efficient as its library of cases grows—marks a significant advancement in diagnostic AI.
Summary updated