Transfer learning based clinical concept extraction on data from multiple sources.

Area of Science:

• Natural Language Processing
• Machine Learning
• Biomedical Informatics

Background:

• Machine learning models typically assume identical training and test data distributions.
• This assumption is often violated in clinical concept extraction tasks, hindering model generalizability.
• Adapting models to new datasets from different institutions with varying data distributions is a significant challenge.

Purpose of the Study:

• To develop a robust clinical concept extraction model transferable to data from different institutions.
• To address the challenge of distribution shift in clinical text data using transfer learning.
• To evaluate a novel transfer learning approach that requires minimal target domain data.

Main Methods:

• An instance-based transfer learning method, TrAdaBoost, was employed.
• TrAdaBoost was integrated with Bagging to mitigate negative transfer and enable softer weight updates.
• The method was validated using BETH, PARTNERS, and BETHBIO datasets from the 2010 i2b2/VA challenge and a custom dataset.

Main Results:

• The proposed method demonstrated superior performance compared to the baseline model in cross-institutional experiments.
• Performance improvements of 2.3% and 4.4% were observed in BETH vs. PARTNERS and BETHBIO vs. PARTNERS comparisons, respectively.
• Statistical significance was confirmed through confidence intervals, and the need for only a small amount of target domain data was highlighted.

Conclusions:

• The integrated TrAdaBoost and Bagging approach effectively addresses distribution shifts in clinical concept extraction.
• This method enables the development of high-performing concept extraction models with limited labeled data from the target domain.
• The findings suggest broad applicability for improving cross-institutional model adaptation in biomedical natural language processing.