Imprecise meanings as a cause of uncertainty in medical knowledge-based systems

Steven J. HenkindUAI 1986

TL;DR

This paper argues that a significant source of uncertainty in medical knowledge-based systems (KBS) stems not from probabilistic inference, but from the inherent imprecision of the language used to represent medical knowledge. It highlights how subtle differences in word definitions between experts can lead to conflicting conclusions, even when dealing with seemingly objective medical terms. The authors propose that addressing this "lexical imprecision" is crucial for building more robust and reliable medical AI.

What Problem Does This Paper Address?

Building effective knowledge-based systems, especially in complex domains like medicine, relies on accurately representing knowledge. While much research has focused on handling uncertainty in the strength of inferences (e.g., "there's a 70% chance of this diagnosis"), this paper points to a more fundamental problem: the ambiguity and imprecision of the very words used to define medical concepts. Medical sublanguages, like any specialized vocabulary, are prone to this. A single term, like "oliguria" (low urine output), can have slightly different interpretations among physicians. This isn't about inherently subjective terms like "lethargic," but about ostensibly objective terms where subtle definitional variations can lead to drastically different conclusions. This imprecision can cripple a KBS, leading to user interface issues and flawed inferencing, ultimately impacting diagnostic accuracy.

Methods

The paper doesn't propose a specific new algorithm or technique for handling lexical imprecision. Instead, it focuses on identifying and characterizing the problem. The authors use examples from medical terminology, such as "pulsus paradoxus" and "petit mal seizures," to illustrate how different interpretations of these terms can arise even among experts. They then discuss potential strategies for mitigating the impact of this imprecision. These include:

  • Direct Input Acquisition: Minimizing the need for human interpretation by having the system gather data directly (e.g., monitoring a chemical plant).
  • Quantification: Replacing vague terms with precise numerical measurements whenever possible (e.g., asking for urine output volume instead of just "oliguria").
  • Incorporation of Definitions: Building explicit definitions of semantic primitives (words) into the KBS, so that all users and experts share a common understanding.
  • Fuzzy Set Theory: Using fuzzy logic to represent concepts that are based on cutoffs, allowing for a more nuanced representation than a simple binary "present" or "absent."

Datasets

Not discussed in the paper. The paper focuses on conceptual analysis and uses examples from existing medical literature and common medical understanding.

Key Insights and Ideas

  • Lexical Imprecision vs. Lexical Ambiguity: A crucial distinction is made. Lexical ambiguity refers to a single word having multiple, distinct meanings (e.g., "beat" as a verb or noun). Lexical imprecision, the paper's focus, is when a word has several similar but not identical meanings, leading to subtle but significant interpretation differences.
  • The "Pulsus Paradoxus" Problem: This example powerfully illustrates how different cutoffs for a medical sign (e.g., 10mm Hg vs. 13mm Hg decline in systolic pressure) can lead to opposite diagnoses, even when the underlying phenomenon is the same.
  • Expert Inconsistency: Even when a single expert's knowledge might be internally consistent, combining knowledge from multiple experts with slightly different definitions for the same terms can lead to logical inconsistencies and incorrect deductions in a KBS.
  • Ubiquity of the Problem: The authors argue that lexical imprecision is not a niche issue but a pervasive phenomenon in specialized fields, and will likely persist as knowledge domains evolve.

Key Results

The paper does not present quantitative results from experiments. Its primary contribution is a conceptual framework and a call to action. The "results" are the identification and detailed explanation of lexical imprecision as a significant challenge for medical KBS, supported by illustrative examples. The paper highlights that efforts to standardize medical terminology have had mixed success, underscoring the persistent nature of the problem.

Limitations and Future Work

The authors acknowledge that completely eliminating lexical imprecision is likely impossible due to the evolving nature of scientific knowledge. They also note that while incorporating definitions into a system is a promising approach, it can be resource-intensive and may not be feasible for all primitives in a large system. Furthermore, the recursive nature of definitions means complete consistency is hard to enforce.

Future work, as suggested by the paper, involves actively developing and incorporating techniques to handle lexical imprecision into medical KBS. This includes further exploring the application of direct input acquisition, quantification, definition inclusion, and fuzzy set methods. The authors indicate they are currently examining ways to integrate these techniques.

Summary updated

BibTeX
@inproceedings{henkind1986imprecise,
  author = {Steven J. Henkind},
  title = {{Imprecise meanings as a cause of uncertainty in medical knowledge-based systems}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1986},
}
Figure