A causal bayesian model for the diagnosis of appendicitis

Stanley M. Schwartz, Jonathan Baron, John R. ClarkeUAI 1986

TL;DR

This paper introduces a "causal Bayesian model" for medical diagnosis, specifically appendicitis, that addresses limitations of traditional Bayesian approaches by incorporating causal relationships between symptoms and diseases. This novel approach significantly improves diagnostic accuracy by better handling symptom interdependence and providing more reliable probability estimates for patient management.

What Problem Does This Paper Address?

Medical diagnosis is a complex task involving inferring underlying diseases from observed symptoms. Traditional Bayesian approaches, while theoretically sound, struggle with the sheer number of conditional probabilities required when symptoms are not independent. This "combinatorial explosion" makes building comprehensive diagnostic systems difficult. Furthermore, standard Bayesian models often assume symptom independence, which is unrealistic and can lead to inaccurate probability estimates, especially when symptoms are correlated due to shared underlying causes. This paper tackles these issues by proposing a causal Bayesian framework that models these interdependencies more effectively.

Methods

The core innovation is the "causal Bayesian model." Instead of treating symptoms as directly independent or dependent on a disease, this model introduces intermediate "pathstates" that represent causal mechanisms linking diseases to symptoms. For example, appendicitis might cause a gastrointestinal disturbance (pathstate), which in turn causes anorexia and nausea (symptoms). This hierarchical, causal structure allows for more realistic modeling of symptom interdependence.

Key improvements over standard Bayesian methods include:

  • Modeling Symptom Interdependence: Symptoms are assumed to be independent with respect to their proximal causes (pathstates), rather than directly with respect to the disease. This significantly reduces the number of required probabilities.
  • Handling Different Levels of Severity: The model can incorporate pathstates that represent increasing levels of severity of a stressor, allowing for a more nuanced representation of how symptoms manifest.
  • Accounting for External Causes: The model allows for symptoms to be caused by factors outside the specific diseases being considered, addressing the issue that not all symptoms are directly caused by the diseases in the hypothesis pool.
  • Incorporating Time: Likelihood links are made conditional on the time since disease onset, reflecting the dynamic nature of symptom presentation.

The implementation uses a hierarchical graph structure (tree) where nodes represent diseases, pathstates, and symptoms. Links represent probabilities, and child nodes of the same parent are treated as probabilistically independent.

Datasets

The paper describes the development of a model for appendicitis diagnosis. The knowledge base was built using an expert physician. For evaluation, the authors used a dataset of 100 recorded cases of appendicitis and 100 cases of nonspecific abdominal pain. A comparison model was also tested on this dataset.

Key Insights and Ideas

  • Causality Matters for Diagnosis: Explicitly modeling the causal pathways from disease to symptoms, rather than just statistical correlations, leads to a more robust and accurate diagnostic system.
  • Pathstates as Causal Bridges: Intermediate causal states (pathstates) are crucial for bridging the gap between diseases and symptoms, effectively explaining why certain symptoms co-occur.
  • Real Probabilities are Key: Unlike some alternative AI approaches (e.g., certainty factors), this model outputs real probabilities, which are essential for informed patient management decisions based on treatment utilities.
  • Expert Knowledge is Quantifiable: The research demonstrates that expert medical knowledge, even subjective estimates, can be effectively structured and quantified within a causal Bayesian framework.

Key Results

The causal Bayesian model was tested against a standard Bayesian model that assumed symptom independence. The causal model demonstrated superior calibration (a measure of how well predicted probabilities match actual frequencies) with a statistically significant improvement. This suggests that the causal approach provides a more reliable estimation of disease probabilities, which is critical for making optimal treatment decisions.

Limitations and Future Work

  • Subjective Estimates: The model relies on subjective probability estimates from an expert, which can be prone to biases. The authors attempted to mitigate this through coherency checks.
  • Temporal Dynamics: While the model incorporates time, further work is planned to fully leverage symptom information from multiple time points.
  • External Causes Complexity: The handling of external causes for symptoms is a complex area, and the paper outlines ongoing work to refine this aspect.
  • Generalizability: The paper focuses on appendicitis. Future work would involve testing the model's applicability to other medical domains and comparing it with other "hot" AI approaches for uncertainty.

Summary updated

BibTeX
@inproceedings{schwartz1986causal,
  author = {Stanley M. Schwartz and Jonathan Baron and John R. Clarke},
  title = {{A causal bayesian model for the diagnosis of appendicitis}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1986},
}
Figure