Explainable Artificial Intelligence Opens New Avenues for Depression Detection from Language

🧠🤖 In mental health, obtaining a prediction is not enough: we also need to understand how it was generated.

From the ALBA group at the University of León, in collaboration with Xeridia’s Artificial Intelligence Department, we have developed a method to estimate the severity of depression from clinical interview transcripts.

In our work, we use large language models to transform the content of interviews into ten interpretable variables, linked to clinical dimensions of the PHQ-8. From these, a regression model calculates the estimated score, while SHAP allows us to understand the influence of each variable on the result. 🔍

📊 The combination of GPT-4o-mini and Ridge reduced the error by up to 30% compared to traditional representations like TF-IDF and Word2Vec. Additionally, we identified feelings of guilt and suicidal ideation among the factors with the highest predictive relevance.

Our study combines language models, predictive capability, and clinically understandable explanations to advance towards more transparent artificial intelligence tools in the field of mental health.

📄 Published in Applied Soft Computing
🔗 https://lnkd.in/euZ4B6hU

🤝 ALBA Research Group (University of León) · Xeridia