Identifying Similar Cases in Document Networks Using Cross-Reference Structures.

Area of Science:

• Computational biology
• Data science
• Pharmacovigilance

Background:

• Identifying similar clinical cases from large datasets is crucial for pharmacovigilance.
• Existing methods may not effectively capture complex relationships within adverse event reports.

Purpose of the Study:

• To explore document network creation using varying information thresholds and clustering algorithms.
• To identify document clusters representing similar clinical cases from vaccine adverse event data.

Main Methods:

• Constructed networks from vaccine adverse event reports using structured data (Medical Dictionary for Regulatory Activities Preferred Terms) and text mining output.
• Linked reports based on shared terms (PTs) and their combinations (dyads to sextuplets), and applied information-theoretic weighting.
• Applied three clustering algorithms: Visualization of Similarities (VOS), Louvain, and k-means, evaluating performance on simulated and real-world datasets.

Main Results:

• The VOS algorithm generally outperformed Louvain and k-means in identifying known clusters.
• The optimal weighting scheme depended on the complexity of the clinical case cluster.
• Singleton weighting was effective for clusters driven by a single term, such as intussusception.
• Clustering based on text mining output showed marginal differences compared to structured data.

Conclusions:

• The developed approach effectively supports the identification of similar nodes (documents) within a document network.
• Network construction and clustering offer a viable strategy for discovering similar clinical cases in pharmacovigilance.
• Algorithm and weighting scheme selection should consider the nature of the clinical cases being investigated.