Introduction
Extracting family history (FH) from unstructured clinical notes is a critical task for personalized medicine, yet this information is often buried in the free-text sections of Electronic Health Records (EHRs). Traditional Natural Language Processing (NLP) approaches typically utilize a pipeline method, first identifying entities (e.g., family members and diseases) and then determining the relations between them. However, this sequential process is prone to error propagation, where inaccuracies in entity recognition negatively impact subsequent relation extraction. To overcome this, researchers have developed a deep joint learning framework that simultaneously performs both tasks within a single neural network architecture.
Methodology
The core of the proposed system is a unified deep learning model that integrates entity recognition and relation extraction. This approach allows the two tasks to share parameters and inform each other during the training process.
Model Architecture
The architecture consists of several distinct layers:
- Input Layer: Words are converted into dense vectors using GloVe word embeddings, supplemented by Part-of-Speech (POS) tags to capture grammatical context.
- Entity Recognition: A Bidirectional LSTM (Bi-LSTM) processes the input sequence to understand context from both directions. A softmax layer then predicts tags to identify family members (FM), medical observations, and living status (LS).
- Relation Extraction: This layer takes the output from the entity recognition step—including specific "label embeddings"—and uses another Bi-LSTM to analyze the relationships between identified entities.
Joint Optimization
The model is optimized using a shared loss function $L = \alpha L_e + (1-\alpha)L_r$, where $L_e$ and $L_r$ represent the losses for entity recognition and relation extraction, respectively. The combination coefficient $\alpha$ is used to balance the importance of each task.
Experimental Results
The model was evaluated using the OHNLP2018-FH dataset, which contains 149 manually annotated patient records. The joint learning approach demonstrated superior performance compared to standard pipeline baselines, achieving state-of-the-art results in the BioCreative/OHNLP 2018 challenge.
- Entity Identification (Subtask 1): The model achieved an F1-score of 0.8901.
- Family History Extraction (Subtask 2): The model achieved an F1-score of 0.6359.
Notably, the joint method improved relation extraction performance by 6.31% over the pipeline method. The researchers also found that the optimal value for $\alpha$ varied by task: $\alpha = 0.6$ was best for entity recognition, while $\alpha = 0.4$ improved complex relation extraction.
Discussion and Future Work
While the deep joint learning model significantly reduces error propagation, it still faces challenges with indirect relatives (e.g., identifying a grandmother's siblings correctly). Future research will focus on incorporating Bi-LSTM-CRF architectures to improve tagging consistency and implementing attention mechanisms to better capture long-range dependencies in complex clinical sentences.
Summary updated