Identifying the Perceived Severity of Patient-Generated Telemedical Queries Regarding COVID: Developing and...

Area of Science:

• Natural Language Processing (NLP)
• Medical Informatics
• Machine Learning

Background:

• Telemedical query triage is critical for resource allocation in remote healthcare.
• Prioritizing severe patient queries ensures timely care for time-sensitive needs.

Purpose of the Study:

• Evaluate transfer learning for telemedical triage effectiveness.
• Analyze errors in NLP systems for challenging telemedical queries.
• Release a public dataset for respiratory telemedical triage severity classification.

Main Methods:

• Annotated 573 medical queries from online health platforms.
• Evaluated lexical models (TF-IDF) as a baseline.
• Assessed GloVe embeddings with SVM, bi-LSTM, and HAN models.
• Tested transformer architectures: BERT, Bio+Clinical-BERT, and SBERT.

Main Results:

• Transformer models (BERT, Bio+Clinical-BERT, SBERT) achieved a mean F1 score of 0.914 (SD 0.034) to 0.917 (SD 0.037).
• SBERT demonstrated statistically significant improvement over lexical and GloVe-based models.
• Error analysis identified challenges in queries with negations, temporal relations, and complex intents.

Conclusions:

• State-of-the-art transfer learning significantly enhances telemedical triage performance.
• A public dataset for telemedical triage severity classification has been released.
• Future work should focus on modeling complex query characteristics for improved NLP performance.