Deep learning for detecting depression in individuals with and without alexithymia

BACKGROUND: To accurately detect individuals' mental health issues using artificial intelligence and self-report scales, it is crucial to recognize how personal characteristics can affect the detection. This study focuses on the role of alexithymia-a condition where individuals struggle to recognize and articulate emotions and symptoms-in the detection of depression. We aimed to determine whether deep learning models could enhance the accuracy of depression detection in people with alexithymia compared to self-report scales. METHODS: We analyzed data from 194 patients with major depressive disorder and 105 community controls, employing eight large language models (LLMs) trained on transcript text from clinician-administered structured interviews using the Hamilton Depression Rating Scale (HAMD). RESULTS: Here we show that generalized logistic regression analysis indicates a positive relationship between alexithymia and depression. Using the HAMD as the gold standard, individuals with alexithymia show poorer performance on the self-reported Hospital Anxiety and Depression Scale-Depression Subscale (HADS-D) in identifying depression (b = -0.37, p = .002). Four of the eight LLMs (AUCs=0.87-0.89) significantly outperform the HADS-D (AUC = 0.79) in depression detection (p