Pair Ungkana Kerttongmee
← Research

Flagship research project

Clinical Knowledge-Guided Adaptive AI Interviews for Depression Screening

Can explicit, auditable clinical knowledge help an AI-assisted interview ask better follow-up questions, while staying transparent, safe and under clinician oversight?

Ongoing researchNon-diagnostic research

Motivation

Structured questionnaires are consistent and well understood, but they cannot adapt to what a person has already said. Free-flowing AI conversation adapts naturally, but it can be hard to explain why a particular question was asked, or whether enough evidence has been gathered.

Specialist time is a scarce resource in many health systems, and any tool that supports early conversations must earn trust through transparency and safety, not only fluency.

This project explores whether the two strengths can be combined: the adaptability of conversation, governed by knowledge that clinicians can inspect.

Research question

Can follow-up questions chosen by explicit, auditable clinical knowledge help an AI-assisted interview gather the evidence it needs more completely and more transparently than a fixed-question baseline, while remaining safe and acceptable to participants?

Related questions

  • Completeness: does adaptive selection cover the relevant evidence more fully than a fixed script?
  • Explainability: can a clinician follow why each question was asked?
  • Safety: are risk signals detected and escalated reliably?
  • Experience: do participants find the interview acceptable, and does optional self-reflection add value?
  • Context: does the approach fit the Thai language and screening pathway?

These are proposed research questions, not established findings.

Conceptual framework

The design separates what to ask from how to say it, so that the first can be inspected and the second can be humane.

  1. Knowledge framework

    A structured knowledge base informed by PHQ-9 and DSM-5, defining what an interview needs to cover.

  2. Evidence tracking

    A running record of what has been covered, what is still missing and what is unclear.

  3. Adaptive question selection

    Rules choose the next question from the gaps in evidence. They are inspectable, not hidden in a model.

  4. Reflective listening

    A conversational layer phrases each question and response warmly and attentively.

  5. Safety check

    Risk signals are screened throughout and trigger escalation.

  6. Clinician handoff

    A transparent summary, with evidence traced back to what the participant said, is passed to a clinician.

The cycle repeats until the evidence is sufficient or the interview is escalated or stopped.

Knowledge-guided adaptive interviewing

Instead of following a fixed script, the interview chooses each follow-up from what is still missing or unclear. Selection is governed by explicit rules derived from the knowledge framework, so the reason for any question can be shown.

A key design principle is that the language model does not decide what to ask. It helps express the question naturally, while selection stays in inspectable rules.

AI reflective listening

Reflective listening means paraphrasing and acknowledging what someone has said before moving on. In this design it is a conversational skill the system performs, not a claim that it feels or understands as a person does.

It is intended to make the interview more humane without influencing which questions are asked.

Participant self-reflection

Participants may optionally reflect on their own responses, for example by reviewing a plain-language summary of what they said.

The aim is to support agency and engagement. Whether it adds value is itself something to evaluate.

Safety, explainability and clinician oversight

  • Safety escalation

    Risk signals are screened throughout, and the interview is designed to escalate to a person rather than continue alone.

  • Explainability

    Each question can be traced to the evidence gap that prompted it, so reviewers can follow the reasoning.

  • Clinician oversight

    The system hands over to a clinician. It does not act as one, and it makes no diagnosis.

Proposed evaluation

Proposed offer

All steps below are proposed and subject to supervision and ethical approval.

  1. Simulated interviews

    Test the approach with varied simulated profiles and communication styles, comparing it with a fixed-question baseline on evidence completeness and number of questions.

  2. Expert review

    Clinicians review the appropriateness of questions, the quality of explanations and safety behaviour.

  3. Safety testing

    Scripted risk scenarios check that escalation behaves as designed.

  4. Participant studies

    Only after appropriate ethical approval and with clinical oversight, studies of acceptability and experience.

Current development status

Status as of November 2026. No empirical findings have been reported.

  • In design

    Research framing and design

    Defining the research question, framework and evaluation plan.

  • Early development

    Knowledge framework

    Provisional. Thai clinical content still needs confirmation with clinicians.

  • Early development

    Software prototype

    An early prototype is under development, with testing based on simulated interviews.

  • Planned

    Expert review

    Planned once the prototype is stable enough to review.

  • Not started

    Studies with participants

    Not started. Would require appropriate ethical approval and clinical oversight.

  • Not started

    Publications and results

    None yet.

Interested in this research?

Pair welcomes conversations with researchers, clinicians and partners who care about responsible AI in mental health.

Start an enquiry