Background
Substance use disorders (SUDs) pose significant societal challenges, and underlying mechanisms remain poorly understood. Work within the growing field of computational psychiatry has begun to offer novel insights into these underlying mechanisms, including impairments in learning from negative outcomes and less deterministic decision-making, among others. However, the longitudinal stability and predictive utility of these computational measures remain underexplored, limiting their clinical applicability.
Methods
A confirmatory longitudinal study was conducted with 144 participants (75 with SUDs and 69 healthy comparisons [HCs]) from the Tulsa 1000 project. Participants completed a three-armed bandit task at baseline and 1-year follow-up. Computational modeling assessed parameters including learning rates and action precision, among others. Bayesian and frequentist approaches tested group differences, stability, and associations with symptom severity (Drug Abuse Screening Test [DAST] scores). Machine learning analyses also evaluated out-of-sample predictive accuracy when combining this sample with an earlier exploratory dataset (83 SUDs, 48 HCs).
Results
Computational measures showed moderate stability over 1 year (ICC range: 0.4–0.58). Learning rates for losses were consistently lower in individuals with SUDs than HCs (posterior probability > 0.99), replicating prior findings. Baseline computational parameters did not significantly predict follow-up DAST scores. Out-of-sample classification achieved modest accuracy (59 %, AUC = 0.62).
Conclusion
Findings confirm moderate longitudinal stability and group differences in computational parameters, supporting their mechanistic relevance but raising questions about their predictive value. This highlights the need for experimental designs and enhanced reliability in computational psychiatry. Future work should integrate neurophysiological measures and dimensional approaches to improve clinical relevance.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号