Schedule:
Tuesday, May 31, 2016
Pacific D/L (Hyatt Regency San Francisco)
* noted as presenting author
Matthew Tyburski, PhD, Researcher, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
Karran A Phillips, MD, Medical Director, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
C. Debra Michelle Furr-Holden, PhD, Associate Professor, Johns Hopkins University, Baltimore, MD
Adam J Milam, PhD, Medical Student, Wayne State University, Detroit, MI
Massoud Vahabzadeh, PhD, Section Chief, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
Mustapha Mezghanni, MS, Researcher, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
Jia-Ling Lin, PhD, Researcher, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
David H. Epstein, PhD, Researcher, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
Kenzie L. Preston, PhD, Branch Chief, National Institute on Drug Abuse Intramural Research Program, Baltimore, MD
Some behaviors are better to prevent than to have to treat in their aftermath. Examples include manic episodes, disordered eating, and suicide attempts—each of which has been the focus of scientific efforts to foresee the future. In our specialty, addiction treatment, we want to prevent relapse, which has long been recognized as problem separate from, and more challenging than, persistence of ongoing drug misuse. Prior attempts to predict imminent lapse have usually focused on time scales of weeks or months, which is problematic when attempting to predict a behavior that occurs in a moment. Predicting behavior in real time would open up a world of live, just-in-time, mHealth interventions for drug lapses to relapse and other behaviors. The present study attempts to take a step towards that goal by making ambulatory predictions of mood and behavior 90 minutes into the future.
In two separate pools of polydrug-using, methadone-maintained outpatients (pilot study, n = 27; main study, n = 81), we collected time-stamped GPS data and ratings of mood, stress, and drug craving over 28 weeks at randomly prompted times via ecological momentary assessment (EMA). We mapped participants’ GPS tracks for the 6.5 hours before each EMA entry. We trained randomForest machine-learning models to predict stress, mood, heroin craving, and cocaine craving 30, 60, and 90 minutes into the future. The main predictor was an independently obtained observer rating of visible neighborhood disorder (NIfETy) along the GPS tracks. In each of the two studies, we randomly reserved part of the data to validate the prediction model.
The models predicted mood, drug craving, and stress 90 minutes into the future in both the pilot and main study. In the main study, using only enviornmental data, agreement was generally at a moderate level (agreement statistics, kappa: Stress .38-0.45, Cocaine Craving .36-.55, Heroin Craving .52-.64; r: Positive Mood .71-.78, negative mood .52-.63) Adding person-level predictors increased the accuracy of the predictions to generally substantial levels (agreement statistics, kappa: Stress .43 -.52, Cocaine Craving .40-.62, Heroin Craving .61-.69; r: Positive Mood .80-.83, negative mood .59 -.68).
Tremendous advances are being made in collection and analysis of environmental and big data. Our goal is to harness these advances for precision medicine to improve the lives of patients and reduce the burdens of behavioral disorders. We succeeded in automated prediction of the immediate behavioral future by a model that could ultimately run on a smartphone. Our models perform well on group-level data—a valuable achievement in itself. But to achieve the goal of a just-in-time mHealth intervention these predictions will need to be extended to the individual level.
![[ Visit Client Website ]](images/banner.gif)