| Popis: |
Due to the high economic and public health burden of chronic pain, and the risk of public health consequences of opioid-based treatments, there is a need to identify effective alternative therapies. The evidence basis for many alternative therapies is weak or nonexistent. Social media presents a unique opportunity to gather large-scale knowledge about such therapies self-reported by sufferers themselves. We attempted to (i) verify the presence of largescale chronic pain-related chatter on Twitter, (ii) develop natural language processing (NLP) and machine learning for automatically detecting chronic pain sufferers, and (iii) identify the types of chronic pain-related information reported by them. We collected data from Twitter using chronic pain-related hashtags and keywords. We manually performed binary annotation of a sample of 4998 posts to indicate if they were self-reports of chronic pain experiences or not, and obtained inter-annotator agreement of 0.82 (Cohen’s kappa). We trained and evaluated several state-of-the-art transformer-based text classification models using the annotated data. The RoBERTa model outperformed all others (F1 score = 0.84; 95% CI: 0.80-0.89), and we used this model to classify a large number of unlabeled posts. We identified 22,795 self-reported chronic pain sufferers and collected their past posted data. Via manual and NLP-driven analyses, we found information about but not limited to alternative treatments, sufferers’ sentiments about treatments, side effects, and self-management strategies. Our social media-based approach will result in an automatically growing massive cohort over time, and the data can be leveraged to identify self-reported effective alternative therapies for diverse chronic pain types. |
