Related Experiment Videos
Named entity recognition in electronic health records using transfer learning bootstrapped Neural Networks
Luka Gligic1, Andrey Kormilitzin1, Paul Goldberg1
1University of Oxford, United Kingdom of Great Britain and Northern Ireland.
Summary
This study introduces a novel transfer learning approach for neural networks (NNs) in medical natural language processing. The method significantly improves named entity recognition and relationship extraction accuracy, even with limited electronic health record data.
Area of Science:
- Computational linguistics
- Machine learning in healthcare
- Bioinformatics
Background:
- Neural networks (NNs) excel in various machine learning tasks but require large labeled datasets, which are often scarce in specialized domains like electronic health records (EHRs).
- Limited data hinders the practical accuracy of NNs in extracting valuable information from EHRs.
- Existing methods struggle with the data scarcity inherent in medical natural language processing tasks.
Purpose of the Study:
- To develop a robust approach for named entity recognition (NER) and relationship extraction (RE) in electronic health records (EHRs) despite data limitations.
- To enhance the performance of neural network models through transfer learning and pretraining on unannotated data.
- To improve upon state-of-the-art results in medical information extraction challenges.
Main Methods:
- Developed a transfer learning strategy by pretraining word embeddings on a large corpus of unannotated EHRs.
- Utilized pretrained embeddings as a foundation for various neural network architectures.
- Applied attention-based sequence-to-sequence models for relationship extraction tasks.
Main Results:
- Achieved a 94.6 F1 score for named entity recognition in the I2B2 2009 Medical Extraction Challenge, surpassing the competition winner by 4.3 F1 points.
- Demonstrated strong performance in extracting relationships between medical terms, reaching an 82.4 F1 score using bootstrapped attention-based seq2seq models.
- Successfully addressed the challenge of data scarcity in EHR analysis for NNs.
Conclusions:
- The proposed transfer learning approach effectively overcomes data limitations in medical NLP tasks.
- Pretraining word embeddings on unannotated EHRs significantly boosts NN performance for NER and RE.
- This method offers a practical solution for leveraging NNs in healthcare settings with limited labeled data.