Deep learning for detecting depression in individuals with and without alexithymia.

Area of Science:

• Artificial Intelligence in Mental Health
• Computational Psychiatry
• Machine Learning for Clinical Diagnosis

Background:

• Accurate mental health detection requires understanding how personal characteristics influence outcomes.
• Alexithymia, difficulty recognizing/articulating emotions, impacts depression detection.
• This study investigates AI's potential to improve depression detection in alexithymia.

Purpose of the Study:

• To evaluate if deep learning models enhance depression detection accuracy in individuals with alexithymia.
• To compare the performance of large language models (LLMs) against self-report scales.

Main Methods:

• Analysis of data from 194 major depressive disorder patients and 105 controls.
• Utilized eight large language models (LLMs) trained on structured interview transcripts.
• Employed Hamilton Depression Rating Scale (HAMD) for data collection and as a gold standard.

Main Results:

• Generalized logistic regression showed a positive link between alexithymia and depression.
• LLMs (AUCs=0.87-0.89) outperformed the HADS-D scale (AUC=0.79) in depression detection.
• For individuals with alexithymia, LLMs achieved AUCs of 0.79-0.96, while HADS-D reached only 0.35.

Conclusions:

• LLMs demonstrate superior performance over self-report scales for depression detection, particularly in alexithymic individuals.
• Patient characteristics like alexithymia are critical for accurate depression assessment.
• Deep learning offers enhanced accuracy for clinical depression assessment and potentially other mental health conditions.