Applying interpretable deep learning models to identify chronic cough patients using EHR data.

Area of Science:

• Computational medicine
• Artificial intelligence in healthcare
• Electronic health record analysis

Background:

• Chronic cough affects 10% of adults and is linked to various conditions.
• Identifying chronic cough patients is difficult due to the lack of a specific ICD code.
• Electronic health records (EHRs) are a valuable resource for patient identification.

Purpose of the Study:

• To investigate data representations and deep learning algorithms for chronic cough prediction.
• To develop computational methods for identifying chronic cough cases from EHR data.
• To improve patient identification for clinical and research purposes.

Main Methods:

• Utilized real-world EHR data from October 2005 to September 2015.
• Evaluated deep learning and traditional machine learning models using structured (medication, diagnosis) and unstructured (clinical notes) data.
• Employed Natural Language Representation and a transformer-based deep learning algorithm (BERT with attention).

Main Results:

• A BERT with attention model achieved 0.856 sensitivity and 0.866 specificity using structured data.
• Combining structured data with symptoms from clinical notes improved sensitivity to 0.952 and specificity to 0.930.
• The attention mechanism identified key predictive features, and deep learning outperformed a rule-based algorithm.

Conclusions:

• Deep learning models reliably identify chronic cough patients using EHR data, enhancing research generalizability.
• Integrating structured and unstructured EHR data further boosts the sensitivity and specificity of chronic cough identification.
• These models show potential for predicting other diseases and identifying patient cases.