Use of a Machine Learning Framework to Predict Substance Use Disorder Treatment Success

L ACION; D KELMANSKY; M VAN DER LAAN; E SAHKER; D JONES; S ARNDT

PLOS ONE

PUBLIC LIBRARY SCIENCE

Lugar: San Francisco; Año: 2017

1932-6203

There are several methodsfor building prediction models. The wealth of currently available modelingtechniques usually forces the researcher to judge, a priori, what will likelybe the best method. Super learning (SL) is a methodology that facilitates thisdecision by combining all identified prediction algorithms pertinent for aparticular prediction problem. SL generates a final model that is at least asgood as any of the other models considered for predicting the outcome. Theoverarching aim of this work is to introduce SL to analysts and practitioners. Thiswork compares the performance of logistic regression, penalized regression,random forests, deep learning neural networks, and SL to predict successful substanceuse disorders (SUD) treatment. A nationwide database including 99,013 SUDtreatment patients was used. All algorithms were evaluated using the area underthe receiver operating characteristic curve (AUC) in a test sample that was notincluded in the training sample used to fit the prediction models. AUC for themodels ranged between 0.793 and 0.820. SL was superior to all but one of thealgorithms compared. An explanation of SL steps is provided. SL is the firststep in targeted learning, an analytic framework that yields double robusteffect estimation and inference with fewer assumptions than the usualparametric methods. Different aspects of SL depending on the context, itsfunction within the targeted learning framework, and the benefits of thismethodology in the addiction field are discussed.