Precursor-induced conditional random fields: connecting separate entities by induction for improved clinical named...

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Biomedical Informatics

Background:

• Clinical named entity recognition (NER) is crucial for processing health records.
• Traditional first-order Conditional Random Fields (CRFs) struggle with long-distance dependencies between clinical entities.
• Clinical narratives often contain relationships (causal, posterior) between separated entities that standard CRFs cannot leverage.

Purpose of the Study:

• To develop an improved CRF model for clinical NER that captures high-order label transition factors.
• To leverage non-entity tokens as information carriers for better entity recognition.
• To enhance the performance and efficiency of clinical NER systems.

Main Methods:

• Introduced a precursor-induced CRF model that uses non-entity tokens to transmit information between separated entities.
• The model's non-entity states memorize precursor entity information, allowing it to propagate forward.
• Compared the proposed model against first- and second-order CRFs on two clinical NER datasets.

Main Results:

• The precursor-induced CRF model achieved superior F1-scores compared to both first- and second-order CRFs.
• The proposed model demonstrated greater computational efficiency than higher-order CRF models.
• Performance improvements were observed on the i2b2 2012 challenge and SNUH EHR datasets.

Conclusions:

• The precursor-induced CRF effectively exploits long-distance transition factors for improved clinical NER.
• This approach enhances F1 scores without exponentially increasing computational time.
• The proposed model offers a significant advancement over existing CRF-based clinical NER methods.