For this observational retrospective study, we collected user-generated posts and comments from 27 subreddits(/r/) related to cannabis use (Figure 1; Multimedia Appendix 1). A set of keywords for adverse experiences SRRMA was used for data mining and manual coding to generate a sample of “signal” Reddit data (ie, posts about cannabis product adverse experiences requiring medical attention: hospitalization, or visiting a health care professional, emergency room, or urgent care). This initial sample of SRMMA prelabeled “signal” was joined with data from the full Reddit data corpus.

Natural language processing was applied to cluster Reddit posts and comments for further manual annotation. We used the bidirectional encoder representations from transformers (BERT) topic model with the Python (version 3.7) package BERTopic (version 0.6.0), which is a pretrained, self-supervised transformer-based algorithm that embeds texts, extracts topics, and then clusters texts.

BERT was used to cluster groups of nonlabeled data with SRRMA prelabeled signal data into k topic clusters, where the value of k is determined according to the dataset size (5≤k≥20). In each generated cluster, we calculated the percentages of SRRMA prelabeled signal data in that topic cluster to estimate how well that cluster was related to our SRRMA prelabeled data (% of potential SRRMA prelabeled signal data in a cluster that could be appended to the SRRMA prelabeled dataset = amount of SRRMA prelabeled training signal data in a cluster/amount of total data in a cluster). The higher the percentage, the more likely the uncoded data were related to our SRRMA prelabeled topic. Cluster selection was determined based on the highest percentages. The new signals were then added to our coded signal pool for the next iteration of BERT modeling: we filtered the uncoded dataset with the keywords repeatedly, combined the filtered uncoded data with all prelabeled signal data, and then performed the next iteration of BERT to find the best clusters to manually annotate for identifying additional signals. Posts and comments from these clusters were extracted, and an inductive coding approach was used to label and characterize cannabis-related adverse experiences and adverse experiences that met the SRRMA criteria. Human ethics and consent to participate were not applicable, as all information was from the public domain and did not involve user interaction, and any identifiable information was aggregated and removed from the results.