Combining deep learning with token selection for patient phenotyping from electronic health records

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Electronic health records (EHRs) contain vast clinical data, but free-form text notes pose analysis challenges.
• Variability in writing styles across EHRs complicates automated information extraction.
• Patient phenotyping from clinical notes requires advanced analytical methods.

Purpose of the Study:

• To investigate deep learning neural networks combined with natural language processing for analyzing clinical discharge summaries.
• To evaluate the impact of network architectures, sample sizes, and token information content on patient phenotyping accuracy.
• To identify optimal methods for classifying ten patient disorders, including challenging cases like Chronic Pain.

Main Methods:

• Utilized deep learning neural networks and natural language processing (NLP) to analyze text data from clinical discharge summaries.
• Investigated various network architectures, including convolutional neural networks (CNNs) at word-level and combined word- and sentence-level (ws-CNN).
• Analyzed the influence of sample size and token information content, employing a token selection mechanism based on Zipf's law.

Main Results:

• The combined word- and sentence-level input CNN (ws-CNN) demonstrated the largest performance gain for classifying Chronic Pain patients.
• Both data quality and quantity are crucial for leveraging complex network architectures beyond simple word-level CNNs.
• Larger sample sizes are necessary for complex models due to limited information per sample, often carried by few tokens.

Conclusions:

• Deep learning models, particularly ws-CNNs, show significant potential for accurate patient phenotyping from EHR discharge summaries.
• The findings highlight the importance of data quantity and quality in enhancing deep learning model performance for clinical text analysis.
• A token selection mechanism, informed by Zipf's law, aids in understanding token information content and contributes to explainable AI.