A fuzzy relation-based extension of Reggia's relational model for diagnosis handling uncertain and incomplete information

Didier Dubois, Henri PradeUAI 1993

TL;DR

This paper introduces a novel fuzzy relational model for medical diagnosis that extends existing approaches by explicitly handling uncertainty in both the presence of symptoms and the causal relationships between disorders and symptoms. It offers a more nuanced representation of diagnostic information, moving beyond binary states to graded possibilities and certainties, thereby improving diagnostic reasoning under incomplete data.

What Problem Does This Paper Address?

Diagnosing diseases often involves dealing with incomplete or uncertain information. Traditional diagnostic models, like Reggia's relational model, describe associations between disorders and their resulting manifestations (symptoms). However, these models typically assume a binary world: a symptom is either present or absent, and a disorder either causes a symptom or it doesn't. This is a significant simplification, as in reality, symptoms can be observed with varying degrees of certainty, and the causal links themselves can be uncertain. For instance, a patient might only partially exhibit a symptom, or a disorder might possibly lead to a symptom but not always. This paper aims to bridge this gap by developing a diagnostic model that can gracefully handle these nuances of uncertainty and incompleteness, moving beyond a purely binary or probabilistic framework.

Methods

The authors propose a fuzzy relation-based extension to Reggia's relational model. The core innovation lies in using twofold fuzzy sets to represent uncertainty. Instead of just stating a symptom is present or absent, they introduce:

  • Fuzzy sets for manifestations: Distinguishing between manifestations that are certainly present (M+) and those that are certainly absent (M-), and crucially, allowing for manifestations that are more or less certainly present or more or less certainly absent.
  • Fuzzy causal relations: Modeling the relationship between disorders (D) and manifestations (M) not as a simple binary link, but as fuzzy sets. This includes:
  • Manifestations certainly caused by a disorder (M(d)+).
  • Manifestations certainly not caused by a disorder (M(d)-).
  • And importantly, manifestations that are possibly caused by a disorder, allowing for graded uncertainty.

This approach leverages possibility theory, which provides an ordinal way to represent uncertainty based on comparisons of certainty levels, rather than strict probabilities. The model extends Reggia's work by incorporating these graded uncertainties into the diagnostic inference process, allowing for a more refined ranking of potential disorders.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and model development, rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Beyond Binary: The Power of Graded Uncertainty: The most significant insight is that real-world diagnostic information is rarely black and white. Symptoms and causal links exist on a spectrum of certainty and possibility, and a model that captures this nuance can lead to more robust diagnoses.
  • Twofold Fuzzy Sets as a Natural Fit: The concept of twofold fuzzy sets, which represent elements that are "more or less certainly in" and "more or less certainly out" of a set, is a powerful tool for modeling this dual nature of uncertainty (certainty vs. possibility).
  • Possibility Theory for Ordinal Reasoning: By using possibility theory, the model offers a qualitative, ordinal approach to uncertainty. This is advantageous when precise probabilistic information is unavailable or difficult to obtain, focusing instead on relative degrees of certainty.
  • Improved Discrimination: While handling incomplete information generally increases the number of potential diagnoses, the graded nature of this fuzzy model allows for a ranking of these possibilities, effectively improving the discrimination power compared to a purely binary incomplete model.

Key Results

The paper presents a formal extension of Reggia's model to handle graded uncertainty. It provides equations (e.g., (5), (6), (11), (13)) that allow for the computation of potential disorders (D) and ranked sets of disorders (D) based on the fuzzy sets of manifestations and causal relationships. The authors demonstrate that their approach can recover results from both the completely informed and partially informed cases, while offering a more refined ranking of candidate disorders.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Direct Disorder-Manifestation Association: The current model directly links disorders to manifestations. Future work could incorporate intermediary sets, creating a more complex causal chain (D → S → M).
  • General Incomplete Information: The model doesn't capture all forms of incomplete information, such as "manifestations mi or mj are present (but perhaps one of them is absent)" or temporal aspects of symptom onset.
  • Temporal Dynamics: The current model assumes all effects of a disorder occur simultaneously. Future research could incorporate time-dependent relationships between disorders and manifestations.
  • Logical Formalism: Expressing the proposed approach in a logical formalism would allow for integration with other reasoning systems, such as assumption-based truth-maintenance systems.
  • A Priori Information: Incorporating a priori knowledge about disorders, beyond what's captured by the fuzzy relations, is a key area for future development, potentially using possibilistic abduction.
  • Quantitative vs. Ordinal: While the current model is ordinal, future work could explore quantitative extensions or a unified framework that encompasses both probabilistic and ordinal models.

Summary updated

BibTeX
@inproceedings{dubois1993fuzzy,
  author = {Didier Dubois and Henri Prade},
  title = {{A fuzzy relation-based extension of Reggia's relational model for diagnosis handling uncertain and incomplete information}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}

Loading related content...

Figure