Machine-learning of medical cannabis chemical profiles reveals analgesia beyond placebo expectations.

Abstract

Background: The efficacy of medical cannabis in alleviating pain has been demonstrated in clinical trials, yet questions remain regarding the extent to which specific chemical compounds contribute to analgesia versus expectation-based (placebo) responses. Effective blinding is notoriously difficult in cannabis trials, complicating the identification of compound-specific effects.

Methods: In a prospective study of 329 chronic pain patients (40% females; aged 48.9 ± 15.5) prescribed medical cannabis, we examined whether the chemical composition of cannabis cultivars could predict treatment outcomes. We used a Random Forest classifier with nested cross-validation to assess the predictive value of demographics, clinical features, and approximately 200 chemical compounds. Model robustness was evaluated using six additional machine learning algorithms.

Results: Here we show that incorporating chemical composition markedly improves the prediction of pain relief (AUC = 0.63 ± 0.10) compared to models using only demographic and clinical features (AUC = 0.52 ± 0.09; p < 0.001). This result is consistent across all models tested. While well-known cannabinoids such as THC and CBD provide limited predictive value, specific terpenoids, particularly α-Bisabolol and eucalyptol, emerge as key predictors of treatment response.

Conclusions: Our findings demonstrate that pain relief can be predicted from cannabis chemical profiles that are unknown to patients, providing evidence for compound-specific therapeutic effects. These results highlight the importance of considering the full range of cannabis compounds when developing more precise and effective cannabis-based therapies for pain management.