A Neural Expert System with Automated Extraction of Fuzzy If-Then Rules and Its Application to Medical Diagnosis

NeurIPS 1990

TL;DR

This paper introduces a novel fuzzy neural expert system (FNES) capable of automatically extracting fuzzy If-Then rules with linguistic relative importance from trained neural networks. This approach aims to bridge the gap between the black-box nature of neural networks and the need for interpretable knowledge bases, demonstrated through its successful application in diagnosing hepatobiliary disorders.

What Problem Does This Paper Address?

Traditional expert systems, while interpretable, often struggle with knowledge acquisition and handling uncertainty. Neural networks excel at learning from data and managing complex patterns but are notoriously opaque ("black boxes"), making it difficult to understand their decision-making process. This paper addresses the challenge of creating expert systems that combine the learning power of neural networks with the interpretability of rule-based systems. Specifically, it tackles the problem of automatically extracting understandable rules from a trained neural network, going beyond simple rule extraction to include the relative importance of conditions within those rules, expressed linguistically.

Methods

The core of the proposed system is a fuzzy neural expert system (FNES) built upon a distributed neural network architecture. This network includes input, intermediate ("hidden"), and output cells, with connections having integer weights. A key innovation is the automated extraction of fuzzy If-Then rules from this trained network. The process involves:

  1. Embodiment and Generalization: Information from training data is learned and embedded within the fuzzy neural network's structure and weights.
  2. Fuzzy If-Then Rule Extraction: A multi-step algorithm is employed to extract rules. This includes:
  3. Framework Extraction: Identifying propositions (input items) that form the antecedent (If-part) of a rule by analyzing cell activations and connection weights. This involves both forward and backward search mechanisms to determine which input patterns lead to specific output states.
  4. Linguistic Truth Value Assignment: Assigning a degree of certainty (e.g., "Very True," "Possibly True") to each extracted rule based on weighted sums of output cells.
  5. Linguistic Relative Importance Assignment: Quantifying the influence of each proposition within the antecedent on the rule's conclusion (e.g., "Very Important," "Moderately Important"). This is a novel aspect that enhances rule expressiveness.

The input layer of the fuzzy neural network is designed to handle both fuzzy and crisp data, utilizing fuzzy cell groups and crisp cell groups, respectively.

Datasets

The paper utilizes a real medical database for diagnosing hepatobiliary disorders. This database contains information on sex and the results of nine biochemical tests for 536 patients. The patients were clinically and pathologically diagnosed by physicians. The dataset was split into 373 training data points and 163 test (external) data points.

Key Insights and Ideas

  • Bridging the Interpretability Gap: The FNES offers a way to make neural network knowledge more transparent by extracting human-readable fuzzy If-Then rules.
  • Linguistic Nuance in Rules: Incorporating "linguistic relative importance" for propositions within rule antecedents allows for more flexible and expressive rules, capturing subtle relationships in the data.
  • Automated Knowledge Discovery: The system automates the complex and often manual process of knowledge acquisition from data, reducing the need for extensive human expert intervention.
  • Handling Uncertainty: The use of fuzzy logic inherently allows the system to manage and reason with uncertain or imprecise information, which is common in medical diagnosis.

Key Results

The proposed fuzzy neural expert system was applied to diagnose hepatobiliary disorders.

  • After training on 345 patients, the FNES achieved a 75.5% diagnostic accuracy on the test data (163 patients) and 100% accuracy on the training data.
  • In comparison, a linear discriminant analysis achieved only 65.0% accuracy on the test data and 68.4% on the training data.
  • The FNES demonstrated significantly higher diagnostic accuracy than the linear discriminant analysis, both on training and test datasets.
  • 48 general fuzzy If-Then rules were extracted for diagnosing hepatobiliary disorders (12 for confirming diseases, 36 for excluding them).
  • An evaluation of twelve extracted rules (four confirming, eight excluding) using a specific reasoning method yielded a total diagnostic accuracy of 87.7%.

Limitations and Future Work

  • Computational Complexity: While not explicitly detailed as a limitation, the multi-step extraction algorithm and the use of a distributed neural network suggest potential computational demands.
  • Generalizability of Extraction Algorithm: The paper mentions that exact algorithms for Steps 2 and 3 of rule extraction can be derived from previous work, implying that the current presentation is a high-level overview.
  • Future Work: The authors suggest that this approach is a promising and useful technique for generating practical knowledge bases from various databases. They highlight that enhancing the interpretation capability of real data and embodying implicit/subjective knowledge can lead to a significant reduction in manpower for knowledge acquisition in expert system development. The paper implies further research into refining the reasoning methods for extracted fuzzy If-Then rules.

Summary updated

BibTeX
@inproceedings{1990neural,
  title = {{A Neural Expert System with Automated Extraction of Fuzzy If-Then Rules and Its Application to Medical Diagnosis}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1990},
}
Figure