Utility of Machine-Learning Approaches to Identify Behavioral Markers for Substance Use Disorders: Impulsivity Dimensions as Predictors of Current Cocaine Dependence

Abstract

Background: Identifying objective and accurate markers of cocaine dependence (CD) can innovate its prevention and treatment. Existing evidence suggests that CD is characterized by a wide range of cognitive deficits, most notably by increased impulsivity. Impulsivity is multidimensional and it is unclear which of its various dimensions would have the highest predictive utility for CD. The machine-learning approach is highly promising for discovering predictive markers of disease. Here, we used machine learning to identify multivariate predictive patterns of impulsivity phenotypes that can accurately classify individuals with CD.

Methods: Current cocaine-dependent users (N = 31) and healthy controls (N = 23) completed the self-report Barratt Impulsiveness Scale-11 and five neurocognitive tasks indexing different dimensions of impulsivity: (1) Immediate Memory Task (IMT), (2) Stop-Signal Task, (3) Delay-Discounting Task (DDT), (4) Iowa Gambling Task (IGT), and (5) Probabilistic Reversal-Learning task. We applied a machine-learning algorithm to all impulsivity measures.

Results: Machine learning accurately classified individuals with CD and predictions were generalizable to new samples (area under the curve of the receiver-operating characteristic curve was 0.912 in the test set). CD membership was predicted by higher scores on motor and non-planning trait impulsivity, poor response inhibition, and discriminability on the IMT, higher delay discounting on the DDT, and poor decision making on the IGT.

Conclusion: Our results suggest that multivariate behavioral impulsivity phenotypes can predict CD with high degree of accuracy, which can potentially be used to assess individuals' vulnerability to CD in clinical settings.

Keywords: LASSO; addiction; cocaine; impulsivity; machine learning; substance dependence.

Figure 1

Multivariate patterns of impulsivity measures predicting cocaine dependence. BIS, Barratt Impulsiveness Scale; SSRT, stop-signal reaction time; IMT, Immediate Memory Task; FP, false-positive (commission) errors; FN, false-negative (omission) errors; ln(k), natural log of delay-discounting rate; PRL, Probabilistic Reversal Learning; IGT, Iowa Gambling Task.

Figure 2

Classification accuracy as indexed by the receiver-operating characteristic (ROC) curves and their area under the curve (AUC), separately on the (A) training and (B) test sets.

Figure 3

Classification accuracy as indexed by the receiver-operating characteristic (ROC) curve and their area under the curve (AUC) on the test set, after removing the “age” regressor from the LASSO model.

Figure 4

Distribution of the area under the curve (AUC) of the receiver-operating characteristic (ROC) curve on (A) training and (B) test sets when we permutated the selection of training and test sets over 1,000 repetitions. The black dotted line indicates the mean AUC of 1,000 repetitions.

Figure 5

Classification accuracy as indexed by the receiver-operating characteristic (ROC) curves and their area under the curve (AUC) using leave-one-out cross-validation.

Figure 6

A correlation matrix between all impulsivity measures. Numbers in cells and color bars indicate Pearson correlation coefficients. Significant correlations (p < 0.05, uncorrected) are filled with blue (positive correlation) or red (negative correlation) colors. BIS, Barratt Impulsiveness Scale; SSRT, stop-signal reaction time; IMT, Immediate Memory Task; FP, false-positive (commission) errors; FN, false-negative (omission) errors; ln(k), natural log of delay-discounting rate; PRL, Probabilistic Reversal Learning; IGT, Iowa Gambling Task.