Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2016 Apr 1;161:247-57.
doi: 10.1016/j.drugalcdep.2016.02.008. Epub 2016 Feb 15.

Machine-learning Identifies Substance-Specific Behavioral Markers for Opiate and Stimulant Dependence

Affiliations
Free PMC article

Machine-learning Identifies Substance-Specific Behavioral Markers for Opiate and Stimulant Dependence

Woo-Young Ahn et al. Drug Alcohol Depend. .
Free PMC article

Abstract

Background: Recent animal and human studies reveal distinct cognitive and neurobiological differences between opiate and stimulant addictions; however, our understanding of the common and specific effects of these two classes of drugs remains limited due to the high rates of polysubstance-dependence among drug users.

Methods: The goal of the current study was to identify multivariate substance-specific markers classifying heroin dependence (HD) and amphetamine dependence (AD), by using machine-learning approaches. Participants included 39 amphetamine mono-dependent, 44 heroin mono-dependent, 58 polysubstance dependent, and 81 non-substance dependent individuals. The majority of substance dependent participants were in protracted abstinence. We used demographic, personality (trait impulsivity, trait psychopathy, aggression, sensation seeking), psychiatric (attention deficit hyperactivity disorder, conduct disorder, antisocial personality disorder, psychopathy, anxiety, depression), and neurocognitive impulsivity measures (Delay Discounting, Go/No-Go, Stop Signal, Immediate Memory, Balloon Analogue Risk, Cambridge Gambling, and Iowa Gambling tasks) as predictors in a machine-learning algorithm.

Results: The machine-learning approach revealed substance-specific multivariate profiles that classified HD and AD in new samples with high degree of accuracy. Out of 54 predictors, psychopathy was the only classifier common to both types of addiction. Important dissociations emerged between factors classifying HD and AD, which often showed opposite patterns among individuals with HD and AD.

Conclusions: These results suggest that different mechanisms may underlie HD and AD, challenging the unitary account of drug addiction. This line of work may shed light on the development of standardized and cost-efficient clinical diagnostic tests and facilitate the development of individualized prevention and intervention programs for HD and AD.

Keywords: Addiction; Amphetamines; Heroin; Impulsivity; Machine-learning; Protracted abstinence.

Conflict of interest statement

Conflicts of Interest

No conflict declared.

Figures

Figure 1
Figure 1
Multivariate patterns of demographic, psychiatric, personality, and neurocognitive measures classifying individuals with past heroin- or amphetamine-dependence. CD = Conduct Disorder; ASPD = Antisocial Personality Disorder; BDI = Beck Depression Inventory; Anx = Anxiety; Anx Sens = Anxiety Sensitivity; LRSP = Levenson’s Self-Report Psychopathy Scale; PCL = Psychopathy Checklist: Screening Version; WURS = Wender Utah Rating Scale for ADHD; BIS = Barratt Impulsiveness Scale; SSS = Sensation Seeking Scale; IGT = Iowa Gambling Task; SST = Stop Signal Task; IMT = Immediate Memory Task; DRD = Delayed Reward Discounting; BART = Balloon Analogue Risk Task; GNGT = Go/Nogo Task; CGT = Cambridge Gambling Task; DA = Delay Aversion; DT = Decision Time; QDM = Quality Decision-Making; RA = Risk Adjustment; RT = Risk Taking.
Figure 2
Figure 2
Classification accuracy of past heroin dependence as indexed by the Receiver-Operating-Characteristic (ROC) curves and their area under the curve (AUC) of (A) the training set and (B) the test set. Panels C and D represent the histograms of AUCs when we randomly selected training and test sets 1,000 times. Dashed black lines indicate the mean values of histograms.
Figure 3
Figure 3
Classification accuracy of past amphetamine dependence as indexed by the Receiver-Operating-Characteristic (ROC) curves and their area under the curve (AUC) of (A) the training set and (B) the test set. Panels C and D represent the histograms of AUCs when we randomly selected training and test sets 1,000 times. Dashed black lines indicate the mean values of histograms.

Similar articles

See all similar articles

Cited by 17 articles

See all "Cited by" articles

Publication types

Feedback