Post-traumatic stress comorbidity in substance use disorder: machine learning analyses of phenotypic drivers

David C. Houghton,Heidi M. Spratt

DOI: https://doi.org/10.1038/s41386-024-01969-1

2024-08-22

Neuropsychopharmacology

Abstract:Substance use disorders (SUDs) rarely exist in isolation, but rather tend to associate with psychiatric comorbidities that complicate prognosis. Indeed, SUD co-occurs frequently with post-traumatic stress disorder (PTSD), which correlates with increased SUD severity and poorer treatment outcomes. Theoretical accounts of PTSD/SUD comorbidity have been put forth, including the self-medication (i.e., management of emotions through intoxication) and the shared vulnerability (i.e., common etiological mechanisms) hypotheses. Research investigating these possibilities has traditionally occurred by testing for statistical differences in trait levels between comorbid and non-comorbid samples. Unfortunately, the population-level inferences drawn from these comparisons replicate infrequently in external samples [1], making it difficult to identify the core etiological mechanisms underlying comorbid SUD-PTSD. Recent studies have demonstrated the potential of machine learning (ML) approaches in identifying the mechanistic biosignatures of psychiatric sub-groups, including within SUDs [2]. The primary advantage of ML over classical statistics is their relative suitability for high-dimensional data (i.e., numerous independent variables), allowing researchers to find hidden patterns among noisy datasets. As such, our study [3] utilized classification and regression trees (CART) with the goal of extracting significant predictors of comorbid SUD-PTSD as compared to SUD alone. A sample of adults with active SUD ( n = 102; Mean Age = 44.38; 40.20% female; 19.61% white) were recruited at a SUD clinic and completed a cross-sectional, standardized, multidimensional phenotyping assessment battery (PhAB), constructed in collaboration with the National Institute of Drug Abuse and based on Research Domain Criteria principles [4]. Most participants were primary opioid users ( n = 55), while marijuana ( n = 25) and cocaine users ( n = 22) made up the minority. Domains of predictors included negative emotionality, metacognition, reward, interoception, and sleep. After training, the test of the CART algorithm achieved high (86.67%) predictive accuracy (Area Under the Curve = 0.81; Sensitivity = 66.67%; Specificity = 75.00%), with comorbidity status predicted by depressive symptoms and fear-related attentional bias.

pharmacology & pharmacy,neurosciences,psychiatry