Construction of Genealogical Knowledge Graphs From Obituaries: Multitask Neural Network Extraction System

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Genealogical information is crucial for biomedical research, including disease heritability and risk prediction.
• Electronic health records (EHRs) and claims data are used to infer family relationships, but completeness can be limited.
• Online obituaries offer a novel, rich data source for constructing comprehensive family trees.

Purpose of the Study:

• To develop an end-to-end information extraction system for constructing Genealogical Knowledge Graphs (GKGs) from online obituaries.
• To enrich EHR data with detailed genealogical information for advanced biomedical research.
• To create a system that can assemble individual GKGs into larger, more comprehensive family structures.

Main Methods:

• A corpus of 1700 online obituaries was curated, with data augmentation used to address data scarcity.
• A multitask artificial neural network was developed to simultaneously detect names, extract relationships, and assign attributes (dates, residence, gender, age).
• A predefined family relationship map with 4 entity types and 71 relationship types was utilized.
• Related GKGs were assembled into larger graphs by identifying individuals present in multiple obituaries.

Main Results:

• The system achieved high performance with precision (94.79%), recall (91.45%), and F1-score (93.09%) on 10-fold cross-validation.
• A total of 12,407 GKGs were constructed, including one spanning 4 generations and 30 individuals.
• The system demonstrated the potential for enriching EHR data with genealogical insights.

Conclusions:

• Genealogical Knowledge Graphs derived from online obituaries can significantly enhance biomedical research capabilities.
• The developed multitask deep neural system provides an effective method for constructing and assembling GKGs.
• The study highlights the value of novel data sources for building comprehensive family relationship information, with shared code available for the scientific community.