Probing Patient Messages Enhanced by Natural Language Processing: A Top-Down Message Corpus Analysis

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Communication

Background:

• Patients increasingly use asynchronous communication platforms to interact with healthcare teams.
• Automating the classification and triage of patient messages via NLP can significantly enhance clinical efficiency.
• Understanding patient-generated text is crucial for developing advanced NLP applications.

Purpose of the Study:

• To characterize the content of patient portal messages using NLP methods.
• To provide descriptive analyses of patient text data.
• To inform the development of sophisticated NLP applications for healthcare.

Main Methods:

• Collected approximately 3,000 patient portal messages from cardiology, dermatology, and gastroenterology departments.
• Classified messages into 'Active Symptom', 'Logistical', 'Prescription', or 'Update' categories.
• Utilized Named Entity Recognition (NER) with the UMLS library to identify medical concepts and analyzed message distributions.

Main Results:

• Active Symptom and Logistical messages constituted about 67% of the corpus.
• 'Findings' was the most frequent medical concept across message types and departments.
• Specific keywords like 'Anatomical Sites' and 'Disorders' were common in 'Active Symptom' messages, while 'Drugs' were prevalent in 'Prescription' messages.

Conclusions:

• Descriptive analysis of patient portal messages reveals key content themes and variations.
• Insights into message content differences can guide the creation of more effective NLP models.
• This study provides a foundation for improving automated message triage and clinical workflow.