Categorising patient concerns using natural language processing techniques.

Area of Science:

• Healthcare Management
• Computational Linguistics
• Patient Experience Research

Background:

• Manual analysis of large patient feedback volumes is challenging.
• Existing frameworks for categorizing patient concerns can be limited.
• Automated methods are needed to efficiently process unstructured patient data.

Purpose of the Study:

• To apply a semi-automated, computational topic-modeling approach to analyze a large corpus of patient feedback.
• To identify and categorize patient concerns in a healthcare system using Latent Dirichlet Allocation (LDA).
• To compare the effectiveness of LDA-generated topics with existing manual categorization frameworks.

Main Methods:

• Utilized Latent Dirichlet Allocation (LDA), an automated topic-modeling algorithm.
• Analyzed 76,163 patient concerns from Alberta Health Services (2011-2018).
• Reduced 40 initial LDA topics to 28 coherent topics after manual interpretation.

Main Results:

• The LDA model identified key patient concerns including communication issues causing delays (10.58%), elderly community care (8.82%), nurse interactions (8.80%), and emergency department care (7.52%).
• LDA-generated topics offered more nuanced categorizations than manual methods.
• Some LDA topics were novel, such as 'smoking in inappropriate settings', not captured by the existing framework.

Conclusions:

• Latent Dirichlet Allocation (LDA) provides a more detailed understanding of patient concerns compared to traditional manual categorization.
• This computational approach is valuable for identifying actionable insights within large datasets of patient feedback.
• The study demonstrates the utility of LDA for discovering emergent themes in healthcare system feedback.