The Discovery of the Causes of Leprosy: A Computational Analysis

AAAI 1996

TL;DR

This paper explores how computational models can reconstruct historical scientific discovery, specifically focusing on the discovery of leprosy's causes. By developing an "indulgent induction" algorithm, the authors show how to model scientific reasoning that allows for exceptions and implicit knowledge, offering a more nuanced understanding of how scientific hypotheses are formed and evolve.

What Problem Does This Paper Address?

The paper tackles the challenge of computationally modeling scientific discovery, particularly in medicine, where inductive reasoning often operates within complex, evolving conceptual frameworks. Traditional data-driven induction struggles to account for historical scientific reasoning, which frequently involved implicit assumptions, "satisficing" theories (simple but not perfectly accurate), and the gradual incorporation of exceptions. The authors aim to bridge this gap by developing an inductive algorithm that can handle these complexities, using the historical discovery of leprosy's causes as a case study. They question whether purely data-driven approaches can truly reconstruct the path to scientific understanding, especially when faced with incomplete data and evolving conceptualizations of disease.

Methods

The core methodological contribution is the development of a new inductive inference approach called "indulgent induction." This approach departs from classical generalization-based induction by explicitly modeling exceptions. The authors then introduce PASTEUR, a new inductive algorithm designed to implement this principle. PASTEUR explores a hypothesis space structured as decision graphs, where rules are connected by a specialization relation. It prioritizes refining exceptions, allowing for the discovery of "satisficing" theories that are useful for structuring reasoning even if not perfectly consistent with all data. This is contrasted with traditional methods like Version Space, Decision Trees, and DNF algorithms, highlighting PASTEUR's unique ability to handle exceptions and implicit knowledge.

Datasets

The primary dataset used is a compilation of 61 cases of leprosy from an Indian leprosy asylum in the 1880s, reported by Phineas (1889). This dataset was collected with a pro-heredity bias, making it an interesting challenge for an algorithm to potentially uncover alternative hypotheses like contagion. The paper also references previous work using data from the discovery of scurvy.

Key Insights and Ideas

  • Indulgent Induction is Crucial for Historical Reconstruction: Purely data-driven induction is insufficient for understanding historical scientific reasoning. Implicit background knowledge and the ability to handle exceptions are vital.
  • "Satisficing" Theories Drive Progress: Simple, imperfect hypotheses (like early theories of heredity or contagion for leprosy) are not necessarily failures but crucial stepping stones that allow researchers to structure their thinking and move towards more accurate theories.
  • The Concept of Disease Matters: The way a disease is conceptualized (e.g., as a penetrating entity vs. a humoral imbalance) significantly influences inductive reasoning and the hypotheses generated.
  • Absence of Disease Can Be Informative: The model's detection of the importance of the absence of disease (implying immunity) in predicting children's health was a surprising and forward-looking insight, aligning with modern research.
  • Computational Models Can Reveal Historical Biases: By simulating historical reasoning with different conceptual frameworks, the authors can demonstrate how biases (like the pro-heredity bias in the leprosy dataset) shaped the scientific discourse.

Key Results

The paper presents two main experimental results:

  1. Descriptive Point of View: When given the 1880s leprosy data with a pro-heredity bias, PASTEUR generated a model that largely reflected the theory of heredity, identifying the father's health as a key factor. However, it also highlighted the importance of the absence of disease, hinting at a more advanced understanding.
  2. Normative Point of View: By introducing an "archaic" concept of disease (where disease is an entity that penetrates the body), PASTEUR was able to induce a model that closely mirrored the Nineteenth-century reasoning on leprosy, emphasizing disease type and heredity. This demonstrates the algorithm's ability to reconstruct historical reasoning by incorporating specific a priori knowledge.

Limitations and Future Work

  • Limitations: The authors acknowledge that the dataset was collected with a pro-heredity bias, which influenced the initial findings. The "indulgent" nature of the induction means hypotheses are not guaranteed to be correct, but rather are exploratory. The paper also notes that the specific "mix" in Hippocratic medicine remained ill-defined.
  • Future Work: The authors suggest that their approach can be used to dynamically account for Nineteenth-century medical reasoning by incorporating knowledge about the concept of disease. They also imply that further exploration of how different conceptual frameworks influence inductive discovery is warranted. The PASTEUR algorithm itself is presented as a flexible tool for exploring satisficing theories and modeling exceptions.

Summary updated

BibTeX
@inproceedings{1996discovery,
  title = {{The Discovery of the Causes of Leprosy: A Computational Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}

Loading related content...

Figure