Human-in-the-loop Evaluation of Assisted Depression Screening (HEADS)

We develop and evaluate HEADS (Human-in-the-loop Evaluation of Assisted Depression Screening), a multilingual, modular, interpretable, HIL, LLM-based automated pipeline for clinical depression diagnosis (DSM-5) and severity (PHQ-9) in diverse clinical populations. Psychiatrists will collect data via QuickSCID-5 and PHQ-9, with adaptive sampling across languages, genders, comorbidity. HEADS has individually trainable modules: (i) Automatic Speech Recognition (ASR) supporting five languages (Kannada, Hindi, Bengali, Assamese, English); (ii) Domain-sensitive translation (Indic to English); (iii) Inference module providing contextual summaries, DSM-5 diagnoses, PHQ-9 score-bands, reasoning and confidence scores. Key deliverables: (a) HEADS solution; (b) de-identified datasets: 4500 Indic language clinical interviews; native-language transcripts and English translations; DSM-5 diagnosis, PHQ-9 scores, clinical reasoning by psychiatrists; 6000 synthetic English transcripts simulating edge cases; (c) well-documented pipelines for eliciting clinical reasoning from psychiatrists in Chain-of-Thought format and synthetic data generation for training diagnostic models. The key goal is to create safe (HIL), scalable, non-intrusive automated system to help non-specialists measure depression in culturally diverse populations, potentially addressing mental health gaps due to shortage of psychiatrists, underdiagnosis or misdiagnosis, with lived experience experts embedded at all stages of study, who will also co-design bias and safety analysis protocols. Keywords: Depression, Culturally-sensitive care, Human-in-the-loop, Generative-AI, Automatic Speech Recognition