Improving heart failure information extraction by domain adaptation.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning (ML)
• Clinical Informatics

Background:

• Adapting information extraction applications to new domains typically requires extensive retraining.
• Previous training data may offer a way to improve adaptation efficiency.

Purpose of the Study:

• To investigate if machine learning models trained on a source domain can be reused to adapt an NLP application to a new target domain.
• To evaluate the effectiveness of domain adaptation techniques for clinical text analysis.

Main Methods:

• Developed an NLP application for extracting congestive heart failure treatment performance measures from echocardiogram reports (source domain).
• Adapted the application to a diverse set of clinical documents (target domain) using domain adaptation approaches.
• Experimented with reusing source model predictions and applying linear interpolation for adaptation.

Main Results:

• Achieved higher recall (92.4%) and precision (95.3%) when adapting the model compared to training only on the target domain.
• Demonstrated the effectiveness of reusing previously trained machine learning models for domain adaptation.

Conclusions:

• Domain adaptation using previously trained models can significantly outperform training from scratch on the target domain.
• Reusing source domain knowledge is a viable and effective strategy for adapting NLP applications in clinical settings.