Capturing additional variance in impact of adolescent depression from free text: proof of concept and preliminary results

Depression in adolescents has grown in prevalence in the past decade. Despite indications of differences in symptom presentation, measurement properties, and treatment response between adolescents and adults, research is dominated by adult participants. Qualitative work has identified symptomatic and functional domains important to depressed adolescents and those around them; however, there remains little quantitative evidence of how these domains relate to the functional impact of the disorder. In this preregistered study, we use adolescent self-report (n = 153) and parent-report (n = 106) responses in the depression section of the Development and Well-Being Assessment (DAWBA) to examine what information relevant to reported depression impact may be missing from its structured symptom-scale. We apply Large Language Models (LLMs) to participant free-text answers to extract information on a taxonomy of 80 outcomes. Through a sequential residual stacking framework that estimates incremental information in symptom sum-scores (S ), item-level symptom-scale answers (Q), and LLM-derived text representations (T ), we find that parent text contains information relevant to their report of depression impact that is not available in the structured symptom-scale, with the largest contributions concentrated in functioning-related domains (e.g., professional responsibilities, social exclusion, autonomy, coping). No such evidence was found in adolescent self-reports. Given the moderate sample size, these results are preliminary. Nonetheless, we provide a proof-of-concept framework for extracting information from open-ended free text in large existing datasets, and estimating incremental information across sources, which could be applied to uncover specific aspects of adolescent mental disorders across populations.