Interpretable Active Learning for Pedigree Data Deduplication in Cancer Genetics.

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Data Science

Background:

• Multicenter studies are crucial for rare genetic conditions but face challenges with duplicate records due to deidentified data.
• Duplicate pedigree data can introduce bias in family-based genetic studies, necessitating robust deduplication methods.
• The Li-Fraumeni and TP53: Understanding and Progress (LiFT UP) study involves families with TP53 mutations, highlighting the need for accurate data.

Purpose of the Study:

• To develop and evaluate an interpretable, active learning-based approach for efficient pedigree deduplication in multicenter genetic studies.
• To address the challenge of duplicate records arising from relatives enrolled at different sites in deidentified datasets.
• To improve data quality for genetic studies, specifically focusing on families with TP53 mutations.

Main Methods:

• Combined heuristic labeling with graph-based features and a machine learning model for iterative duplicate detection.
• Generated a partially labeled dataset using mutation variant diversity and family characteristics.
• Trained a random forest classifier and employed active learning for refinement, applied to LiFT UP study data.

Main Results:

• Achieved 99.95% automated processing in the pedigree deduplication workflow for the LiFT UP study data.
• Minimized manual effort by prioritizing likely duplicates for human review, ensuring high specificity.
• Demonstrated a scalable, automated solution that avoids reliance on traditional identifier-matching filters.

Conclusions:

• Interpretable active learning is an effective strategy for pedigree deduplication in multicenter genetic research.
• The developed method offers a scalable solution for enhancing data quality in genetic studies.
• Future research will focus on refining duplicate identification and assessing generalizability to other genetic datasets.