Classifying patient portal messages using Convolutional Neural Networks.

Area of Science:

• Health Informatics
• Natural Language Processing
• Machine Learning

Background:

• Patient portals are increasingly used for healthcare communication.
• Growing message volumes can overwhelm healthcare providers.
• Automated classification of messages can improve triage and response.

Purpose of the Study:

• To evaluate if semantic features and word context enhance patient portal message classification.
• To compare different feature representations and machine learning models for message categorization.

Main Methods:

• Classified messages into informational, medical, social, and logistical categories.
• Utilized bag of words, graph representations, and Word2Vec embeddings as features.
• Trained random forest, logistic regression, and convolutional neural network (CNN) classifiers.
• Evaluated performance using Area Under the Curve (AUC).

Main Results:

• CNN with Word2Vec embeddings achieved the highest AUCs (0.908-0.943) across all categories.
• Graph representations outperformed bag of words.
• Semantic features significantly improved classification accuracy compared to non-semantic features.

Conclusions:

• Word2Vec embeddings and graph representations enhance patient portal message classification accuracy.
• CNN models combined with semantic features offer superior performance for message categorization.