| Depression remains a widespread concern, underscoring the need for accessible and scalable screening tools. We present Ember, a conversational AI system for preliminary, clinician‑supervised depression screening through structured dialogue. Building on prior conversational depression screening systems such as DEPRA, and PHQ-9-based approaches including Perla and the LLM-based HopeBot, Ember introduces a prompt-constrained framework that preserves structured item delivery, maps open-text responses to PHQ-9 frequency categories, and supports DSM-5-aligned severity categorisation. Powered by GPT-4o, Ember integrates the clinically validated PHQ-9 for symptom screening, with model behaviour guided by structured prompting aligned with DSM-5 domains. Its conversational approach supports open-text responses before mapping them to PHQ-9 frequency options, while real‑time scoring computes the PHQ‑9 total and corresponding severity category. Given the sensitive domain of depression screening, Ember is explicitly non‑diagnostic and not intended to function as a standalone tool. It is designed to support clinicians through structured, reviewable outputs. In a non‑clinical pilot, 30 adult participants evaluated usability, perceived interaction quality and completion of PHQ‑9 delivery. Participants reported the interaction felt supportive and natural, though such impressions are subjective and non‑clinical. Qualitative feedback suggested that Ember could handle off-topic and sensitive disclosures, using clarification and guided redirection to maintain assessment flow. These findings highlight the feasibility of prompt‑constrained conversational PHQ‑9 delivery and motivate future work on clinical validation, safety evaluation and comparison with baseline methods. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.