Detecting clinically related content in online patient posts.

Area of Science:

• Health Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Online health communities are vital for patients managing chronic conditions.
• Concerns exist regarding misinformation and the challenge of moderating large volumes of posts.
• Efficiently identifying clinically relevant content is crucial for providing accurate patient support.

Purpose of the Study:

• To develop and evaluate text classification models for automatically identifying clinical topics in online health forum posts.
• To assist online health communities in managing content and directing patients to validated clinical resources.
• To improve the efficiency and accuracy of content moderation in digital health spaces.

Main Methods:

• Annotated 1817 posts (4966 sentences) from an online diabetes community.
• Tested various text classification models, including Naïve Bayes, with unigrams, bigrams, trigrams, and MetaMap Semantic Types.
• Evaluated model performance using F-measure, Precision, and Recall.

Main Results:

• The Naïve Bayes model achieved the best performance (F-measure: 0.83, Precision: 0.79, Recall: 0.86) on the initial dataset.
• The classifier demonstrated feasibility when applied to a second online diabetes community (F-measure: 0.63, Precision: 0.57, Recall: 0.71).
• Model training was rapid (5 seconds), and classification was near-instantaneous (less than 1 second).

Conclusions:

• Automated identification of clinical topics in online health forums is feasible and effective.
• The developed model can help manage content exchange and connect patients with necessary clinical expertise.
• The approach shows potential for scalability to other online health communities.