Exploring generalized association rule mining for disease co-occurrences.

Area of Science:

• Medical Informatics
• Data Mining
• Public Health

Background:

• Association rule mining is key for automated knowledge discovery in healthcare.
• Constraining the search space in rule mining is crucial to avoid computational complexity and information loss.
• Identifying disease co-occurrences can reveal complex health patterns and inform clinical practice.

Purpose of the Study:

• To apply generalized association rule mining to identify disease co-occurrences using ICD-9-CM codes.
• To compare generalized associations with those derived from raw data to understand information loss and gain.
• To investigate the impact of generalization hierarchies on discovered disease associations.

Main Methods:

• Utilized generalized association rule mining techniques on a statewide hospital discharge dataset.
• Employed the Clinical Classifications Software (CCS) and ICD-9-CM numerical hierarchy for code generalization.
• Maintained data links between raw and generalized codes to track association changes.

Main Results:

• Identified disease co-occurrences by generalizing ICD-9-CM codes.
• Quantified associations lost, overlapping, and newly discovered through generalization.
• Observed that the choice of concept hierarchy influences the resulting disease associations.

Conclusions:

• Generalized association rule mining effectively identifies disease co-occurrences while managing computational complexity.
• The generalization process can lead to both information loss and the discovery of novel associations.
• Concept hierarchy selection is a critical factor in the outcomes of disease association studies.