Social Media Images Can Predict Suicide Risk Using Interpretable Large Language-Vision Models.

Area of Science:

• Artificial Intelligence in Mental Health
• Computational Social Science
• Digital Psychiatry

Background:

• Suicide is a major public health issue, exacerbated by social media and the COVID-19 pandemic.
• Current AI methods for suicide risk prediction from social media have limitations, including "black box" models and underutilization of image data.
• There is a need for interpretable AI models that can predict suicide risk using non-verbal social media inputs like images.

Purpose of the Study:

• To develop an interpretable artificial intelligence (AI) model for predicting clinically valid suicide risk from images.
• To address the limitations of existing AI methods in suicidology, such as lack of interpretability and focus on textual data.

Main Methods:

• Utilized a dataset of 177,220 images from 841 Facebook users who completed a suicide risk scale.
• Employed CLIP (Contrastive Language-Image Pre-training) to extract interpretable features from images.
• Developed a hybrid model integrating theory-driven features with a simple logistic regression classifier.

Main Results:

• The hybrid model achieved high prediction performance (AUC = 0.720, Cohen's d = 0.82), surpassing common deep learning algorithms.
• Results supported the hypothesis that users at risk of suicide tend to post images with more negative emotions and less sense of belonging.
• Demonstrated that publicly available images can be effectively used to predict validated suicide risk.

Conclusions:

• This study is the first to demonstrate the potential of using publicly available images for validated suicide risk prediction.
• The developed interpretable model offers a promising strategy for real-life suicide monitoring tools.
• Findings highlight the importance of incorporating image analysis into digital mental health surveillance for suicide prevention.