Deep Learning Identification of Asthma Inhaler Techniques in Clinical Notes

Area of Science:

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

Background:

• Clinician documentation variability in asthma care impacts quality assessment.
• Manual chart review of electronic health records is labor-intensive and not scalable.
• National asthma guidelines are difficult to implement as real-time feedback tools.

Purpose of the Study:

• To develop and evaluate a deep learning model for identifying inhaler technique documentation in clinical narratives.
• To assess the efficacy of Bidirectional Encoder Representations from Transformers (BERT) with distant supervision for this task.
• To improve the feasibility of real-time feedback on guideline-concordant asthma care documentation.

Main Methods:

• Utilized a deep learning natural language model, Bidirectional Encoder Representations from Transformers (BERT).
• Employed distant supervision to train the BERT model on clinical narratives.
• Compared the performance of the BERT model against a rule-based approach.

Main Results:

• The BERT model with distant supervision demonstrated superior performance in identifying inhaler techniques compared to the rule-based method.
• A performance gain was observed for the BERT model with distant supervision over BERT without distant supervision.
• The deep learning approach showed promise in capturing contextual understanding of clinical narratives for documentation assessment.

Conclusions:

• Deep learning, specifically BERT with distant supervision, offers a more effective method for identifying specific clinical guideline elements like inhaler techniques from unstructured text.
• This automated approach can potentially enhance the assessment of guideline-concordant documentation in asthma care.
• The findings suggest a pathway for improving real-time feedback mechanisms to clinicians for better asthma management.