Deep learning models (DL) have gained considerable attention in the domain of Artificial Intelligence because of their ability to achieve high levels of abstraction through consecutive nonlinear transformations resulting in learning the optimal representation from high-dimensional data. Convolutional Neural Networks (CNNs) are a subtype of DL that show promise in the field of medical image analysis, allowing for the classification of psychiatric disorders based on subtle and diffuse structural brain alterations. 

This project seeks to highlight the merit of state-of-the-art DL approaches, with the primary goal of training a CNN architecture to identify substance use disorder (SUD) in comparison to healthy controls (HCs) based on structural Magnetic Resonance Imaging (sMRI) data and to compare the performance of CNNs with traditional machine learning models. The secondary goal is to perform a series of interpretability analysis on the final networks including Deep Local Interpretable Model-Agnostic Explanations (LIME) and theory-based masking to provide general SUD and substance-specific importance feature measures.

A recently published study entitled “Machine Learning Prediction of Early-Onset Alcohol Use: A Cross-Study, Cross-Sample Validation” (Afzali et al., 2019) focuses on the prediction modeling of early-onset alcohol use in an attempt to highlight important methodological issues and to start a dialogue around best practices in prediction psychiatry. This study addressed five issues in prediction modeling, 1) examination of the comparative performance of different machine-learning algorithms and the potential use of super-learners, 2) feature clustering and domain-contribution analysis 3) coefficient extraction and interpretability, 4) limitations of k-fold cross-validation and necessity of an independent test sample, and 5) ethical concerns regarding the concept of “prediction”.

The computational modelling of decision-making tasks is a powerful method of identifying complex cognitive processes that can be used to create a more detailed picture of the underlying learning mechanisms. In this work, we showed how the combination of biases and sensitivity to reinforcements constitutes a reliable predictor of behavior in a decision-making paradigm. We presented a Bayesian hierarchical computational model which attempts to capture increasingly more complex learning mechanisms on the Passive Avoidance Learning Paradigm (PALP) task, using a reinforcement learning framework.

Model selection was done on the basis of predictive accuracy, leave-one-out cross-validation, and Bayes factor. We illustrated the usefulness of the selected model through the identification of developmental changes in subpopulations, showcasing how it can be used for classification. We identified developmental changes in action bias and sensitivity to reward in adolescents, from ages 12-13 to 17 years old. Furthermore, we conducted two experiments to evaluate the developmental trajectories for gender and for a group of high-risk cannabis users. In this work, we've successfully distinguished developmental changes which can be used to assess divergent maturation trajectories and help understand the cognitive evolution during adolescence. The significance of these findings extends to the design of new standards of cognitive maturity that could lead to better diagnoses of mental health disorders, such as depression and addiction.

In an innovative project, Agent-Based Models (ABM) are utilized to simulate the future of substance use of participants by knowing the initial structure of the environment (i.e. number of initial agents and early onset substance users, rules of interaction, risk level and network of each agent). Models are calibrated by using the initial information of recruited agents in each study and adjusting hyper parameters.

This type of prediction can be used not only for substance use forecast, but also for investigating the effect of an intervention during a certain period of time. Therefore, the model enables us to compare the future of substance use in two scenarios of receiving versus not receiving any intervention. This model can be used to provide a more accurate cost benefit analysis, to compare the effectiveness of an intervention in different scenarios, and to estimate the effect of different aspects of the intervention along the period of time that agents interact in the same environment.

The artificial intelligence/machine learning community’s interest in causality has significantly increased in recent years. Causal inference, as pioneered by Judea Pearl, focuses on the multiple aspects of causal dynamics such as temporal precedence, association, and counterfactuals, which can be quantitatively conceptualized through graphs and structural equations. 

The exceptional opportunity of the access to several large, multimodal, and longitudinal data in the Venture Lab has allowed us to follow a strand of research focusing on quantitative models of causal dynamics between risk factors (e.g. substance use, screen time) and psychiatric outcomes (e.g. depression, psychotic like experiences) resulting in six publications in highly ranked journals (e.g. JAMA psychiatry, JAMA pediatrics). In this strand of research, we focused on quantitative modelling of causal dynamics through associations with temporal precedence and mediator processes using structural equation modelling and Bayesian multi-level models. This quantitative approach to causal dynamics paves the way for an ambitious project on inferential causal modelling within the context of machine learning.