Machine-learning prediction of adolescent alcohol use: a cross-study, cross-cultural validation

Background and aims The experience of alcohol use among adolescents is complex, with international differences in age of purchase and individual differences in consumption and consequences. This latter underlines the importance of prediction modeling of adolescent alcohol use. The current study (a) compared the performance of seven machine-learning algorithms to predict different levels of alcohol use in mid-adolescence and (b) used a cross-cultural cross-study scheme in the training-validation-test process to display the predictive power of the best performing machine-learning algorithm. Design A comparison of seven machine-learning algorithms: logistic regression, support vector machines, random forest, neural network, lasso regression, ridge regression and elastic-net. Setting Canada and Australia. Participants The Canadian sample is part of a 4-year follow-up (2012-16) of the Co-Venture cohort (n = 3826, baseline age 12.8 +/- 0.4, 49.2% girls). The Australian sample is part of a 3-year follow-up (2012-15) of the Climate Schools and Preventure (CAP) cohort (n = 2190, baseline age 13.3 +/- 0.3, 43.7% girls). Measurements The algorithms used several prediction indices, such as F-1 prediction score, accuracy, precision, recall, negative predictive value and area under the curve (AUC). Findings Based on prediction indices, the elastic-net machine-learning algorithm showed the best predictive performance in both Canadian (AUC = 0.869 +/- 0.066) and Australian (AUC = 0.855 +/- 0.072) samples. Domain contribution analysis showed that the highest prediction accuracy indices yielded from models with only psychopathology (AUC = 0.816 +/- 0.044/0.790 +/- 0.071 in Canada/Australia) and only personality clusters (AUC = 0.776 +/- 0.063/0.796 +/- 0.066 in Canada/Australia). Similarly, regardless of the level of alcohol use, in both samples, externalizing psychopathologies, alcohol use at baseline and the sensation-seeking personality profile contributed to the prediction. Conclusions Computerized screening software shows promise in predicting the risk of alcohol use among adolescents.