Induction and uncertainty management techniques applied to veterinary medical diagnosis

Matthew Cecile, Mary McLeish, P. Pascoe, W. TaylorUAI 1988

TL;DR

This paper explores applying machine learning techniques, specifically induction and uncertainty management, to build a diagnostic system for veterinary medicine. By leveraging existing patient data, the researchers aim to improve upon traditional diagnostic charts and create a more robust system for identifying surgical versus medical colic in horses.

What Problem Does This Paper Address?

Veterinary medicine, particularly in diagnosing conditions like colic in horses, faces challenges in accurately identifying cases requiring urgent surgical intervention. This distinction is critical, as misdiagnosis can lead to fatal outcomes for the animal and significant costs for owners who must transport horses for further testing. Current diagnostic aids, like charts, can be simplistic, and experienced clinicians' knowledge is difficult to codify. The paper addresses the need for a computerized diagnostic tool that can assist both in-hospital and remote veterinarians by extracting valuable patterns from extensive medical databases.

Methods

The paper investigates several machine learning approaches:

  • Classical Statistical Methods: Discriminant analysis and logistic regression were used to identify significant variables for diagnosis. These methods aim to find linear relationships between input parameters and outcomes.
  • Bayesian Classification (Autoclass I): This unsupervised learning method aims to discover natural groupings (classes) within the data without prior assumptions about the number of classes or similarity measures. It can handle both categorical and real-valued data and provides insights into attribute influence within each class.
  • Other Induction Techniques:
    • Probabilistic Learning System (PLS1): This system uses an information-theoretic criterion to discriminate classes and can handle dynamic and uncertain learning by representing concepts as prototypes and hyperrectangles. It aims to incorporate probability and its error.
  • Quinlan's Algorithm: This algorithm is used for rule extraction, generating decision trees from data.
  • Evidence Combination (Weight of Evidence): A novel method is proposed that combines evidence using Bayesian updating with an infinite-valued logic (fuzzy set theory). This approach aims to handle both probabilistic and "fuzzy" uncertainty, using the "weight of evidence" to quantify the support for a hypothesis. It also introduces the concept of discovering "important symptom sets" by considering higher-order dependencies.

Datasets

The primary dataset used in this study consists of 253 equine colic cases collected from the teaching hospital at Guelph. Each case includes approximately 20 clinical variables, along with diagnostic codes, presence of a surgical lesion, and lesion type. For evaluation, a subset of 89 cases from this training set was used. The paper notes that the animal hospital's VMIMS system stores a considerable amount of medical information, with the database holding about 48,000 records and 30,000 unique cases.

Key Insights and Ideas

  • Limitations of Classical Methods: While discriminant analysis and logistic regression can identify significant variables, they struggle with non-linear relationships and missing data. The paper highlights that these methods can miss important effects and that estimating missing data can be unreliable.
  • Bayesian Classification for Pattern Discovery: Autoclass I proved valuable in identifying distinct patient groups (classes) and revealing influential attributes within those classes, offering a more nuanced understanding than simple variable selection.
  • Handling Fuzzy Uncertainty: The proposed "weight of evidence" method, combined with fuzzy set theory, offers a unified approach to reasoning in domains with both probabilistic and subjective uncertainty, which is common in medical diagnosis.
  • Importance of Symptom Groups: The research suggests that considering symptom groups (higher-order dependencies) is crucial, as individual symptoms might be less informative or violate Bayesian independence assumptions when considered in isolation. This approach aligns with how clinicians often reason.
  • Clinician Agreement: A significant finding is that the important symptom sets identified by the "weight of evidence" method were also deemed significant by veterinary clinicians, validating the approach.

Key Results

  • The Weight of Evidence method achieved a Negative Predictive Value of 96.7% and a Positive Predictive Value of 86.6% on an 89-case test set, outperforming the clinician's Negative Predictive Value (87.6%) in differentiating surgical lesions.
  • For a specific case where clinicians decided against surgery, the Weight of Evidence method strongly supported a surgical lesion hypothesis with a probability of 0.804.
  • Classical regression analysis for the same case predicted a probability of 0.5818, which the authors found less "convulsive" due to fewer variables considered.
  • Combining Weight of Evidence with Bayesian classification provided a more comprehensive picture, suggesting a poor outcome prognosis for a particular case, which ultimately led to the decision for euthanasia, a decision not solely supported by regression or weight of evidence alone.

Limitations and Future Work

  • Missing Data: The paper acknowledges that missing data was a significant problem for several methods, leading to case elimination or unreliable estimation.
  • Completeness of Induction Methods: Quinlan's algorithm for decision trees faced implementational difficulties, resulting in an incomplete and unreliable tree.
  • Data Retrieval: Retrieving all necessary data for a new dataset proved challenging, requiring manual extraction from hard copy records.
  • Future Work: The authors plan to:
  • Gather a new dataset of 168 cases to further test and refine methodologies.
  • Explore more standard Bayesian models to understand dependencies and conditioning.
  • Implement a prototype system using the evidence combination methodology, considering symptom groups up to size three.
  • Develop a more advanced algorithm for testing independence between groups and estimating errors within a blackboard architecture.
  • Make results available on incoming cases to provide doctors with information from multiple methodologies, especially for difficult cases.

Summary updated

BibTeX
@inproceedings{cecile1988induction,
  author = {Matthew Cecile and Mary McLeish and P. Pascoe and W. Taylor},
  title = {{Induction and uncertainty management techniques applied to veterinary medical diagnosis}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1988},
}
Figure