OBJECTIVES: Anhedonia, the reduced ability to experience pleasure, typically presents in patients with disorders including schizophrenia and major depressive disorder. However, due to inconsistencies in anhedonia definitions and assessment, proper diagnosis is challenging. Several studies have implemented natural language processing (NLP) to identify key expressions around an illness among social media users; however, it is unknown which words are most significantly associated with anhedonia.

METHODS: All posts within the Reddit anhedonia forum between 2017 and 2022 were downloaded. Existing code was adapted to prepare the data for NLP through the processes of tokenization, stop-word removal, and lemmatization. Six unsupervised machine learning (ML) algorithms (including sentiment analyses, clustering algorithms, and topic models) were used to identify patterns associated with anhedonia among 9,887 posts, which were then interpreted by a team of humans.

RESULTS: Among the top 30 identified two-word phrases, term frequency-inverse document frequency (TF-IDF) was highest for phrases related to causation (average score: 12.70), potentially the frustrations of users not knowing the cause of anhedonia Following this were phrases describing anhedonia, resolution of the illness, and other types of phrases (average scores of 8.51, 8.46 and 7.59, respectively). Both clustering and topic model algorithms identified similar words related to anhedonia, including “dopamine”, “low”, “pleasure”, “help”, and “depress” each of which clustered with anhedonia in two of three algorithms tested. While sentiment analyses suggested a neutral sentiment of users toward anhedonia, the severity of positive and negative sentiments was vague.

CONCLUSIONS: Application of NLP-based methods to social media revealed anhedonia-associated language. Words strongly linked to the topic of anhedonia could help to identify anhedonia-related content in other datasets. Future analyses may help identify undiagnosed individuals using informed models and large standardized datasets such as electronic health records and patient health questionnaires.