An Exploratory Study of Social Media Analysis for Rare Diseases Using Machine Learning Algorithms: A Case Study of Trigeminal Neuralgia

Abstract

Rare diseases, affecting approximately 30 million Americans, are often poorly understood by clinicians due to lack of familiarity with the disease and proper research. Patients with rare diseases are often unfavorably treated, especially those with extremely painful chronic orofacial rare disorders. In the absence of structured knowledge, such patients often choose social media to seek help from peers within patient-oriented social media communities thereby generating tremendous amounts of unstructured data daily. We investigate whether we can organize this unstructured data using machine learning to help members of rare communities find relevant information more efficiently in real-time. We chose Trigeminal Neuralgia (TN), an extremely painful rare disorder, as our case study and collected 20,000 social media TN posts. We categorized TN posts into Twitter (very short), and Facebook (short, medium, long) datasets based on message length and performed three clustering experiments. Results revealed GSDMM outperformed both K-means and Spherical K-means in clustering Facebook especially for short messages in terms of speed. For long messages, MDS reduction outperformed the PCA when both were used with K-means and Spherical K-means. Our study demonstrated the need for further topic modeling to utilize among high level clusters based on semantic analysis of posts within each cluster.