[Paper Review] Predicting Opioid Relapse Using Social Media Data

Zhou Yang, Long Nguyen+1 more|arXiv (Cornell University)|Nov 14, 2018

Mental Health via Writing25 references4 citations

TL;DR

This study proposes a Generative Adversarial Network (GAN) model that predicts opioid relapse by analyzing sentiment and social influence patterns in Reddit posts from recovering addicts. The model achieves 83.9% accuracy and 85.47% F1-score, outperforming logistic regression, SVM, and KNN, with generated sentiment images revealing 'joy' and 'negative' emotions as key relapse indicators.

ABSTRACT

Opioid addiction is a severe public health threat in the U.S, causing massive deaths and many social problems. Accurate relapse prediction is of practical importance for recovering patients since relapse prediction promotes timely relapse preventions that help patients stay clean. In this paper, we introduce a Generative Adversarial Networks (GAN) model to predict the addiction relapses based on sentiment images and social influences. Experimental results on real social media data from Reddit.com demonstrate that the GAN model delivers a better performance than comparable alternative techniques. The sentiment images generated by the model show that relapse is closely connected with two emotions `joy' and `negative'. This work is one of the first attempts to predict relapses using massive social media data and generative adversarial nets. The proposed method, combined with knowledge of social media mining, has the potential to revolutionize the practice of opioid addiction prevention and treatment.

Motivation & Objective

• To address the critical need for early opioid relapse prediction to improve prevention and treatment outcomes.
• To leverage social media data from Reddit as a real-time, scalable source of behavioral and emotional indicators of relapse.
• To develop a novel deep learning framework using Generative Adversarial Networks (GANs) that integrates sentiment and social influence features for relapse prediction.
• To validate the model’s performance against traditional machine learning methods like logistic regression, SVM, and KNN on a manually labeled dataset of 1,000 users.

Proposed method

• The model uses a GAN architecture with a generator and discriminator trained on two-channel input images representing sentiment and social influence layers.
• Sentiment and social influence features were extracted from Reddit posts and transformed into 2-channel image representations for each user.
• The generator produces synthetic sentiment images from random noise, while the discriminator distinguishes between real and generated images to refine the data distribution learning.
• Training used Adam optimizer with a learning rate of 0.0001 over 7,000 epochs, with convergence observed around 2,100 epochs.
• The model was evaluated using 50%, 60%, 70%, 80%, and 90% of the training data to assess performance under varying data availability.
• TensorFlow was used for implementation, with tensor dimensions calculated using standard convolutional neural network principles.

Experimental results

Research questions

• RQ1 Can social media data from Reddit be effectively used to predict opioid relapse within the next week?
• RQ2 How does a GAN-based model compare in performance to traditional supervised methods like logistic regression, SVM, and KNN for relapse prediction?
• RQ3 Which emotional states, as captured in social media posts, are most strongly associated with impending relapse?
• RQ4 To what extent do social influence patterns and sentiment dynamics co-vary with relapse risk in recovering opioid users?

Key findings

• The GAN model achieved an accuracy of 83.90% and an F1-score of 85.47%, significantly outperforming logistic regression, SVM, and KNN across all data split sizes.
• The generator's learned latent space produced sentiment images that consistently highlighted 'joy' and 'negative' emotions as strongly associated with relapse.
• Loss curves for both generator and discriminator stabilized around 2,100 epochs, indicating model convergence after approximately 13 hours of training.
• The model maintained strong performance even with reduced training data, achieving 71.40% accuracy and 77.16% F1-score at 50% of the training set.
• Sentiment images with multiple white spaces (indicating mixed emotions) were linked to longer, emotionally complex posts, suggesting deeper psychological states preceding relapse.
• The study demonstrates that generative modeling of social media content can uncover meaningful, interpretable emotional patterns predictive of relapse.