Comprehensive Summary
Hossain et al aimed to obtain disease-specific symptoms shared on social media discourse as a method of studying prevalence and occurrence patterns of polycystic ovary syndrome (PCOS). With assistance of LLMs to identify symptom-related key words, the authors leveraged a lexicon-based symptom extraction (LSE) method to group symptoms. Amongst BERT-Base, BioBERT and Phrase-BERT-based embeddings used for studying the symptoms, Bio-BERT-based k-means clustering was found most effective. Analyses showed the LSE approach outperformed automatic medical extraction tools and LLMs in identifying 64 PCOS symptoms that have been validated by eHealth forums. This model was also able to identify 28 potentially emerging symptoms of PCOS and 8 comorbidities associated with PCOS.
Outcomes and Implications
Polycystic ovary syndrome is both highly prevalent and underdiagnosed due to its variable presentation. This study shows how patient-reported symptoms can be leveraged to build a clinically relevant picture of PCOS, improving diagnostic accuracy and earlier detection. While immediate clinical implementation requires validation, the findings suggest that AI-based public health surveillance could help standardize recognition and management of PCOS across diverse populations.