Mining Patterns and Linkages in Medical Data

1
Mining Patterns and Linkages in Medical Data
Department of Computer Science
EngineeringCollege of Engineering
Samah Fodeh and Pang-Ning Tan

• Health care data contains a wealth of information
  that can be utilized to help improve the quality
  of patient care and to enhance the capability for
  public health surveillance.
• Challenges
• Large database size ( 100 GB).
• Need efficient mining algorithms.
• 2. Trend in data.
• Few records prior to 2001.
• Evaluation validation of data mining results.
• Large number of patterns, some of which may be
  spurious.
• Need a principled approach for evaluating and
  interpreting patterns
• 4. Noise in data.
• Need a robust algorithm.

Overall Framework

• Querying Domain Knowledge
• Domain knowledge is obtained from the Pubmed
  database provided by National Library of
  Medicine.
• Challenge
• Not all articles may be relevant to queried term.
• Cost for querying

Experimental Results

• Methodology
• Preprocess the data.
• Apply an efficient frequent pattern mining
  algorithm called PGMiner.
• Extract association rules from the frequent
  patterns.
• For each rule
• Query the PubMed database to find articles about
  the concepts.
• Use the query results to evaluate and interpret
  the rule.

helicobacter
lesions
gastric
bleeding
carcinona
dysplasia
therapy
peptic
congenital

• Rule Evaluation
• To evaluate the rule, compare the confidence
  measure computed from medical database (C1) to
  the estimated confidence obtained from medical
  literature (C2).
• A rule is considered interesting if
• Purpose is to uncover previously unknown
  relationships

ulcer
healing
gastrointestinal
cancer
infection
cancer
cells
pylori

• Summary
• Combine data mining with domain knowledge
  automatically extracted from medical literature
  to detect unexpected relationships and to aid the
  interpretation of discovered rules.