JMIR Publications

ABSTRACT

Background:

7-Hydroxymitragynine (7-OH) is a potent μ-opioid receptor agonist found in trace amounts in Mitragyna speciosa (kratom) and increasingly marketed in concentrated, "enhanced" products that may expose consumers to a substantially stronger opioid than conventional botanical kratom. Because such products are emerging faster than traditional surveillance systems can track, naturalistic online discussion offers a potential early-warning window, but manual thematic coding of large post volumes is resource-intensive, and unconstrained use of large language models (LLMs) risks introducing unsupported content.

Objective:

This study applied a schema-guided, LLM-assisted extraction pipeline to characterize patterns of 7-OH use and dependence-relevant signals across dedicated Reddit communities, while directly benchmarking model coding against blinded human coding to determine which variables could be reliably automated.

Methods:

We collected 1,671 publicly available posts from five 7-OH–focused subreddits from community founding (2017–2024) through October 2024. A codebook developed with substance-use experts defined nine coding items spanning dependence indicators (quitting, withdrawal, tolerance, addiction language), consumption practices (route, reason for use), kratom comparisons, and physical and psychological side effects. A 100-post validation subset was coded by two blinded human coders and by GPT-3.5 Turbo. Interrater reliability was computed using Gwet's AC1, and items were retained for full-corpus analysis only when AI–human agreement fell within 0.05 of human–human agreement and exceeded 0.50. Descriptive prevalence and co-occurrence statistics were then computed across the full corpus.

Results:

Eight of nine primary items met reliability criteria; for two (kratom comparisons, physical side-effect presence) AI–human agreement exceeded human–human agreement. Only "reason for use" failed, reflecting low human–human agreement (AC1=0.20) on an item requiring clinical judgment. Tolerance was the most frequently discussed dependence indicator (17.6%, n=294), followed by quitting (6.0%) and withdrawal (5.3%); addiction language appeared in 4.2% of posts. Two or more dependence indicators co-occurred in 7.1% of posts; among posts using addiction language, 70% also mentioned tolerance and 54.3% mentioned quitting. Physical side effects were reported in 9.8% of posts (most commonly insomnia, 6.7%) and psychological side effects in 3.7% (anxiety, 3.2%).

Conclusions:

Across 1,671 posts, frequently co-occurring discussion of tolerance, withdrawal, addiction, and quitting—alongside opioid-consistent side effects—suggests 7-OH may carry meaningful abuse liability distinct from botanical kratom. Methodologically, a hybrid design in which LLM coding is retained only after item-level reliability benchmarking achieved human-comparable performance for most, but not all, variables, supporting validated LLM assistance as a scalable tool for surveillance of emerging substances.