An interpretable connectivity-based decoding model for classification of chronic marijuana use

Kaustubh Kulkarni, Matthew Schafer, Laura A. Berner, Vincenzo G. Fiore, Matthew Heflin, Kent E. Hutchison, Vince D. Calhoun, Francesca M. Filbey, Gaurav Pandey, Daniela Schiller, Xiaosi GubioRxiv 2021 · DOI

TL;DR

This paper introduces a novel machine learning approach that combines accurate prediction of chronic marijuana use with interpretable insights into the underlying brain connectivity patterns. By leveraging whole-brain functional connectivity data from fMRI, the model achieves high classification accuracy and identifies specific brain regions and networks crucial for distinguishing users from non-users, offering a new tool for understanding substance use disorders.

What Problem Does This Paper Address?

Psychiatric neuroimaging often focuses on either understanding neural mechanisms (encoding models) or predicting clinical outcomes (decoding models), but rarely bridges the gap between the two. This disconnect limits our ability to gain both accurate predictions and meaningful insights into the neural basis of disorders. Specifically, for substance use disorders like chronic marijuana use, reliable biomarkers are lacking, hindering research and clinical applications. This paper aims to address this by developing a decoding model that is not only predictive but also interpretable, providing a deeper understanding of the neural differences associated with chronic marijuana use.

Methods

The researchers employed a machine learning framework that integrates functional connectivity data with graph theory.

  • Decoding Model Training: Linear classification models (logistic regression and support vector machines) were trained on whole-brain functional connectivity data derived from fMRI scans. The input features were pairwise correlation values between 90 brain regions.
  • Hyperparameter Tuning: L1 and L2 regularization were used to optimize model performance, with regularization strength (alpha) tuned via 10-fold cross-validation on the training set.
  • Model Evaluation: The best-performing models were evaluated on an independent, held-out test set using metrics like accuracy, Area Under the Curve (AUC), precision, and recall.
  • Interpretability via Network Analysis: After training the best predictive model, its learned weights were used to create "weighted connectivity matrices." These matrices highlight the importance of specific functional connections for classification.
  • Predictive Importance of Regions: The mean weighted connectivity for each brain region was calculated to identify regions with high predictive importance.
  • Graph Theoretic Analysis: Network analysis was applied to these weighted connectivity matrices to identify modular subnetworks (communities) that were most predictive of chronic marijuana use. This involved calculating degree centrality (a measure of a region's importance in the network) and using community detection algorithms.

Datasets

The study utilized a combined dataset from two pre-existing fMRI studies:

  • Filbey 2009 dataset: n=125 participants.
  • Filbey 2016 dataset: n=198 participants.

The combined dataset included:

  • Chronic marijuana users (MJ): n=195
  • Healthy controls (HC): n=128

This combined sample of 323 participants is noted as the largest fMRI sample used to date for classifying substance use disorders.

Key Insights and Ideas

  • Connectivity is Key: Functional connectivity between brain regions, rather than just the activity of individual regions, carries significantly more information for distinguishing chronic marijuana users from non-users.
  • Interpretable Models are Possible: It's feasible to build highly accurate predictive models that also offer clear insights into the underlying neural mechanisms, bridging the gap between prediction and explanation.
  • Network Structure Matters: Specific subnetworks (communities) within the brain's functional connectome are disproportionately important for classifying chronic marijuana use, suggesting that disruptions in these networks are central to the disorder.
  • Novel Hypotheses Generation: The approach not only validates existing findings about brain regions involved in craving but also highlights potentially novel regions and network configurations that may be important for understanding marijuana use disorder.

Key Results

  • High Classification Accuracy: The models achieved high out-of-sample accuracy, around 80%, in classifying chronic marijuana users from healthy controls. This performance significantly outperformed chance and models based on regional activity alone.
  • Predictive Regions Identified: Key brain regions implicated in substance use disorders, such as the anterior cingulate cortex (ACC), dorsolateral prefrontal cortex (PFC), and precuneus, were identified as highly predictive. However, the analysis also revealed some exceptions and novel regions.
  • Predictive Communities Found: The graph theoretic analysis identified four core communities of brain regions that were crucial for successful classification. Including these top four communities in the prediction model achieved 80% accuracy, outperforming the full model and demonstrating their significant predictive power.
  • Model Weight Consistency: Different linear classification algorithms produced highly similar model weights, suggesting a robust and consistent representation of the underlying connectivity patterns.

Limitations and Future Work

  • Clinical Deployment Threshold: The reported classification accuracies, while high, may not yet be sufficient for direct clinical deployment.
  • Categorical vs. Continuous Use: The study classified users into a binary category (chronic users vs. non-users) and did not disentangle continuous effects of marijuana use.
  • Self-Reported Behavior: The prediction was based on self-reported behavior, which is subject to its own limitations.
  • Specificity of Effects: The study focused on marijuana use and did not explore the specificity of these neural signatures compared to other substance use disorders or psychiatric conditions.
  • Future Directions:
    • Longitudinal Studies: Predicting future risk of chronic use, especially in adolescents, would be valuable.
  • Data-Driven Parcellation: Using methods like independent component analysis or multimodal data could refine neural signatures.
  • Regression Models: Employing regression-oriented models could allow for stronger inferences about continuous clinical features (e.g., symptom severity) rather than just group differentiation.
  • Voxel-Level Analysis: Refining network analysis at the voxel level could provide finer-grained insights.
  • Comparison with Other Drugs: Comparing marijuana users to users of other substances would help establish marijuana-specific neural signatures.

Summary updated

Abstract

ABSTRACT Background Psychiatric neuroimaging typically proceeds with one of two approaches: encoding models, which aim to model neural mechanisms, or decoding models, which aim to predict behavioral or clinical characteristics from brain imaging data. In this study, we seek to combine these aims by developing interpretable decoding models that offer both accurate prediction and novel neural insights. We demonstrate the effectiveness of this combined approach in a case study of chronic marijuana use. Methods Chronic marijuana (MJ) users (n=195) and non-using healthy controls (n=128) completed a cue-elicited craving task during functional magnetic resonance imaging. Linear machine learning methods were used to classify individuals into chronic MJ users and non-users based on task-evoked, whole-brain functional connectivity. We then used graph theoretic analyses to identify ‘predictive functional connectivities’ among brain regions that contributed most substantially to the classification of chronic marijuana use. Results We obtained high (~80% out-of-sample) accuracy across four different classification models, demonstrating that task-evoked, whole-brain functional connectivity can successfully differentiate chronic marijuana users from non-users. Subsequent network analyses revealed key predictive regions (e.g., anterior cingulate cortex, dorsolateral prefrontal cortex, and precuneus) that are often implicated in neuroimaging studies of substance use disorders, as well as some key exceptions. We also identified a core set of networks of brain regions that contributed to successful classification, comprised of many of the same predictive regions. Conclusions Our dual aims of accurate prediction and interpretability were successful, producing a predictive model that also provides interpretability at the neural level. This novel approach may complement other predictive-exploratory approaches for a more complete understanding of neural mechanisms in drug use and other neuropsychiatric disorders.
BibTeX
@misc{kulkarni2021interpretable,
  author = {Kaustubh Kulkarni and Matthew Schafer and Laura A. Berner and Vincenzo G. Fiore and Matthew Heflin and Kent E. Hutchison and Vince D. Calhoun and Francesca M. Filbey and Gaurav Pandey and Daniela Schiller and Xiaosi Gu},
  title = {{An interpretable connectivity-based decoding model for classification of chronic marijuana use}},
  year = {2021},
  doi = {10.1101/2021.05.04.442433},
  howpublished = {bioRxiv},
}
Figure