Identification of asthma control factor in clinical notes using a hybrid deep learning model

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Artificial Intelligence in Healthcare

Background:

• Variability in guideline-concordant asthma care documentation is a challenge.
• Assessing documentation accuracy requires labor-intensive manual review of electronic health records (EHRs).
• Capturing specific guideline elements like inhaler technique review from EHR free text demands contextual understanding.

Purpose of the Study:

• To develop and evaluate a context-aware language model (BERT) for identifying inhaler techniques in EHR free text.
• To assess the impact of distant supervision and a hybrid approach on model performance.
• To reduce the burden of manual chart review for training deep learning models.

Main Methods:

• Utilized two datasets: manually reviewed notes (1039) and weakly labeled notes (27,363).
• Applied Bidirectional Encoder Representations from Transformers (BERT) and clinical BioBERT (cBERT) models.
• Incorporated distant supervision with rule-based weak labels and explored a hybrid approach with post-hoc rules.

Main Results:

• BERT models without distant supervision performed similarly to rule-based models (F1-score ~0.84).
• BERT models with distant supervision showed improved performance (F1-score up to 0.88).
• Hybrid models achieved the best performance (F1-score up to 0.904), outperforming other methods.

Conclusions:

• BERT models with distant supervision effectively identify inhaler techniques in EHR free text.
• Distant supervision alleviates the need for costly manual chart review in deep learning model development.
• Hybrid models further enhance performance by correcting BERT model errors.