Introduction
Genetic research and personalized medicine rely heavily on accurate biological pedigrees to understand the heritability of complex diseases. However, Electronic Health Records (EHRs) often lack systematic kinship data, frequently capturing administrative relationships (e.g., insurance dependents) rather than comprehensive biological lineages. Online obituaries represent a significant, publicly available resource for multi-generational kinship information. This paper presents a joint neural network model designed to automatically extract these relationships, transforming unstructured death notices into structured data for genetic research.
Methodology: The Joint Bi-LSTM Model
Traditional approaches to information extraction typically utilize a "pipeline" method, where Named Entity Recognition (NER) and Relation Extraction (RE) are performed sequentially. Such systems are susceptible to error propagation. To mitigate this, we proposed a joint bidirectional Long Short-Term Memory (Bi-LSTM) model that identifies names and their specific kinship roles simultaneously.
The model processes text through an embedding layer followed by Bi-LSTM layers, which capture context from both directions. A softmax output layer generates initial tags, which are then refined by a rule-based result improver to ensure logical consistency in the extracted kinship roles.
Deceased-Centric Tagging Scheme
A key innovation of this work is the "deceased-centric" tagging scheme. Since the subject of an obituary is always the deceased, the task is simplified to identifying the "Object" (the relative) and their relationship to that single subject. We employed a modified BIESO (Begin, Inside, End, Single, Other) scheme combined with kinship types.
For instance, in the phrase "survived by his sister, Joyce M. Tottingham," the name is tagged with sister_B, sister_I, and sister_E to denote the beginning, inside, and end of the entity associated with the "sister" relationship. This approach allows the model to leverage kinship keywords as strong contextual clues for entity recognition.
Dataset and Linguistic Features
The model was trained and validated on a corpus of 1,809 obituaries from Rochester, Minnesota, containing 30,035 sentences and nearly 30,000 annotated names. The dataset covers 71 distinct kinship types across multiple generations.
The system was specifically tuned to handle linguistic patterns common in obituaries, such as:
- Last Name Distribution: Recognizing that a shared last name applies to multiple preceding first names (e.g., "parents John and Mary Smith").
- Parenthetical Names: Identifying maiden names or nicknames often enclosed in parentheses.
- Semantic vs. Syntactic Distinctions: Treating terms like "spouse" and "married to" as separate categories during training to account for differing sentence structures, despite their semantic equivalence.
Experimental Results
The joint model demonstrated high performance across various metrics. For the 57 kinship types with sufficient training data (at least 10 examples), the model achieved a micro-averaged F1-measure of 96.98%. At the macro level, which weighs rare and common categories equally, the model reached an F1-score of 74.93%.
Crucially, the joint model outperformed traditional pipeline approaches by 6.5% to 9.7% in macro-F1 scores. Common relationships, such as "child" or "grandchild," were extracted with near-perfect precision (exceeding 96%).
Conclusion and Future Work
This research demonstrates the feasibility of using joint neural models to extract high-fidelity kinship data from public obituaries. While the current model is limited to sentence-level extraction, future work will focus on cross-sentence context and entity resolution—linking individuals across different obituaries to construct large-scale, multi-generational family trees. Integrating these public pedigrees with clinical EHR data holds the potential to significantly advance heritability studies and genomic medicine.
Summary updated