Identification of Semantically Similar Sentences in Clinical Notes: Iterative Intermediate Training Using Multi-Task...

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Electronic health records (EHRs) contain redundant information due to templated notes and copy-pasting.
• Effective utilization of EHR data is hindered by information redundancy.
• Measuring semantic similarity in clinical text is crucial for information condensation.

Purpose of the Study:

• To enhance semantic textual similarity performance in the clinical domain.
• To improve the robustness of clinical text similarity models.
• To leverage manually labeled data and contextualized embeddings for better results.

Main Methods:

• Utilized the ClinicalSTS dataset of 1642 annotated clinical text pairs.
• Developed an iterative intermediate training approach using multi-task learning (IIT-MTL).
• Applied IIT-MTL to ClinicalBERT and ensembled with BioBERT, MT-DNN, RoBERTa, and handcrafted features.

Main Results:

• Achieved state-of-the-art performance in the 2019 n2c2/OHNLP ClinicalSTS challenge.
• Ranked first among 87 submitted systems with a Pearson correlation coefficient of 0.9010.
• The winning system was an ensemble model incorporating IIT-MTL on ClinicalBERT, BioBERT, MT-DNN, and medication features.

Conclusions:

• IIT-MTL effectively utilizes annotated data from related tasks for improved performance on data-limited target tasks.
• This approach offers new possibilities for optimized dataset selection in the clinical domain.
• The study contributes to generating more robust and universal contextual text representations.