Predicting quantity of cannabis smoked in daily life: An exploratory study using machine learning

Abstract

Background: Cannabis use is prevalent in the United States and is associated with a host of negative consequences. Importantly, a robust indicator of negative consequences is the amount of cannabis consumed.

Methods: Data were obtained from fifty-two adult, regular cannabis flower users (3+ times per week) recruited from the community; participants completed multiple ecological momentary assessment (EMA) surveys each day for 14 days. In this exploratory study, we used various machine learning algorithms to build models to predict the amount of cannabis smoked since participants' last report including forty-three EMA measures of mood, impulsivity, pain, alcohol use, cigarette use, craving, cannabis potency, cannabis use motivation, subjective effects of cannabis, social context, and location in daily life.

Results: Our best-fitting model (Gradient Boosted Trees; 71.15% accuracy, 72.46% precision) found that affects, subjective effects of cannabis, and cannabis use motives were among the best predictors of cannabis use amount in daily life. The social context of being with others, and particularly with a partner or friend, was moderately weighted in the final prediction model, but contextual items reflecting location were not strongly weighted in the final prediction model, the one exception being not at work.

Conclusions: Machine learning approaches can help identify additional environmental and psychological phenomena that may be clinically-relevant to cannabis use.

Keywords: Cannabis; Digital health; Ecological momentary assessment; Machine learning; Mobile survey.

Figure 1

The attribute weights for Gradient Boosted Trees Note: see Table 2 for definition of abbreviations.