Abstraction in Belief Networks: The Role of Intermediate States in Diagnostic Reasoning

Gregory M. ProvanUAI 1995

TL;DR

This paper investigates whether including intermediate "pathophysiological states" in Bayesian belief networks (BNs) for medical diagnosis improves diagnostic accuracy compared to simpler two-level networks (disease-finding only). Experiments on a liver disease BN show that these intermediate states have a statistically insignificant impact on diagnostic performance, suggesting that simpler abstractions can be sufficient.

What Problem Does This Paper Address?

Bayesian belief networks (BNs) are powerful tools for diagnostic reasoning, but their complexity can be a bottleneck. Building and querying large BNs requires significant knowledge engineering effort and computational resources. A key question is how much detail is necessary in a BN's representation. Specifically, this paper explores the trade-off between the accuracy of a detailed, multi-level BN (which includes intermediate states like pathophysiological conditions) and a more abstract, two-level BN (which only models direct relationships between diseases and findings). The goal is to provide empirical and theoretical guidance on when such abstractions are appropriate without sacrificing diagnostic performance.

Methods

The authors use a three-level Bayesian network, CPCS-BN, which models diseases, intermediate pathophysiological states (IPSs), and findings. They then create a simplified, two-level version of this network by abstracting away the IPS nodes. This abstraction involves a specific mapping of probabilities:

  • Link Activation Probability Mapping: For a path from a disease node to an IPS node and then to a finding node, the activation probability in the two-level network is the product of the probabilities on the original three-level path (e.g., $p \times q$).
  • Leak Activation Probability Mapping: This involves a more complex process of updating leak probabilities to account for the removed intermediate nodes, using techniques from prior work.

The core of their experimental approach is to compare the diagnostic performance of the original three-level network against its abstracted two-level counterpart using a set of real-world medical diagnosis subnetworks and synthetically generated test cases.

Datasets

The study uses subnetworks derived from the CPCS-BN, a large BN for medical diagnosis of liver and bile diseases. Three specific subnetworks were used, varying in the number of diseases:

  • BN 2: 2 diseases, 42 nodes total
  • BN 3: 3 diseases, 146 nodes total
  • BN 4: 4 diseases, 245 nodes total

Test cases were generated synthetically from these networks by simulating findings according to the network's probabilities. To assess performance under varying evidence, findings were revealed in five "phases," mimicking a medical consultation.

Key Insights and Ideas

  • Abstraction is Often Sufficient: The most striking insight is that, for the studied domain (liver disease diagnosis), removing intermediate pathophysiological states from the BN did not significantly degrade diagnostic accuracy. This suggests that simpler, more abstract models can be as effective as complex ones.
  • Trade-off Between Complexity and Performance: The paper highlights the inherent trade-off between the richness of a BN's representation (and thus its potential accuracy) and the computational and knowledge engineering costs associated with it.
  • Topology Matters for Overestimation: While the overall difference was often insignificant, the study notes that the two-level network slightly overestimates posterior probabilities for diseases, especially in larger networks. This overestimation is linked to the fan-in and fan-out of intermediate nodes in the original network.
  • Cost of Knowledge Engineering: The paper implicitly underscores the significant effort required to model intermediate states, noting that building the CPCS-BN involved substantial additional knowledge engineering compared to a two-level representation.

Key Results

  • Statistically Insignificant Differences: For most test cases and subnetworks, paired t-tests showed statistically insignificant differences (at the 95% confidence level) in the expected posterior probabilities assigned to true diagnoses between the two-level and three-level networks.
  • Slight Overestimation in Two-Level Networks: The two-level network tended to slightly overestimate true-positive and false-positive probabilities, particularly for larger networks. This effect was more pronounced when intermediate nodes had higher fan-in and fan-out.
  • Disease-Specific Differences: In one instance (hepatitis-acute-viral), a statistically significant difference was observed, attributed to the specific topology of IPS nodes in the path to its findings.

Limitations and Future Work

  • Domain Specificity: The authors acknowledge that these results are specific to the studied domain (hepatobiliary diseases) and may not generalize to all diagnostic domains. In other domains, intermediate nodes might be crucial for accuracy or simpler to model.
  • Focus on Accuracy: The study primarily focused on diagnostic accuracy and did not analyze the impact of abstraction on inference speed.
  • Generalization of Abstraction Methods: While the paper proposes an approximate mapping, it notes that more accurate (but computationally expensive) methods exist.
  • Further Analysis: The authors plan to investigate similar abstractions in other domains, analyze the impact on inference speed, and explore more general situations with both positive and negative evidence.

Summary updated

BibTeX
@inproceedings{provan1995abstraction,
  author = {Gregory M. Provan},
  title = {{Abstraction in Belief Networks: The Role of Intermediate States in Diagnostic Reasoning}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1995},
}
Figure