The research, led by Mahrukh Shakoor and Tamim Ahmed Khan of the Department of Software Engineering at Bahria University in Islamabad, together with Hina Ghafoor of the university’s Department of Professional Psychology, describes a system that integrates reflective listening, large language models, knowledge graph reasoning, and clinical guidelines drawn from the DSM-5-TR, the current edition of the Diagnostic and Statistical Manual of Mental Disorders. The goal is not to replace clinicians but to give them a structured, interpretable tool for the initial assessment process, where human resources are scarce and early signals are easy to miss.

At the heart of the framework is a two-stage screening workflow designed to mimic, in a limited way, the flow of a real clinical intake. In the first stage, a reflective-listening conversational agent engages the user in empathetic dialogue. Reflective listening is a counseling technique in which the listener mirrors and restates what the speaker has said, validating their experience and encouraging deeper disclosure. By building this technique into the language model, the researchers aimed to create an agent that does not simply interrogate the user with checklist questions but instead draws out clinically relevant symptoms through conversation that feels supportive and engaging.

During this first-stage dialogue, the system performs contextual reasoning over what the user says, extracting symptom features and representing them within a knowledge graph, a structured network of entities and relationships that allows the AI to connect scattered conversational cues into a coherent clinical picture. Rather than treating each mention of sleeplessness, worry, or low mood as an isolated data point, the knowledge graph links symptoms to one another and to diagnostic categories, enabling an initial mental health risk screening that can be traced and explained. This emphasis on explainability is central to the framework: clinicians can see how the system arrived at its conclusions rather than receiving an opaque risk score from a black-box model.

The second stage brings quantitative rigor to the conversation. The framework incorporates standardized psychometric instruments, including the Penn State Worry Questionnaire, a widely used measure of pathological worry; the Beck Depression Inventory-II, one of the most established tools for assessing the severity of depressive symptoms; and the McLean Screening Instrument for Borderline Personality Disorder, a validated screening tool for that condition. By embedding these instruments into the workflow, the system grounds its conversational impressions in validated, disorder-specific measures, producing a quantitative assessment that supports more fine-grained risk estimation for particular conditions.

To train and evaluate the system, the team constructed a structured dataset that is notable for its clinical grounding. It comprises AI-generated reflective conversations, screening labels informed by DSM-5-TR criteria, extracted symptom features, knowledge graph representations, and, critically, annotations reviewed by clinicians. This dataset was then used to fine-tune the proposed language model, aligning its behavior not just with general linguistic patterns but with the specific vocabulary and logic of mental health assessment. The clinician-reviewed component of the dataset provides a human benchmark against which the AI’s outputs can be measured.

Experimental results reported in the study show encouraging agreement between the framework’s AI-assisted screening outcomes and assessments reviewed by clinicians. The authors highlight that integrating reflective listening, psychometric evaluation, knowledge graph reasoning, and generative AI improved not only screening performance but also interpretability and user engagement, two factors that often pull in opposite directions in automated health tools. Systems that are highly engaging are frequently opaque, while systems that are transparent can feel clinical and cold. The reflective-listening approach appears to soften that trade-off, keeping users in conversation while the underlying reasoning remains inspectable.

The ethical scaffolding around the study is unusually detailed for work in this space. The research was reviewed and approved by the Ethics Review Committee of Bahria School of Professional Psychology at Bahria University, with approval dated 26 January 2026, and all procedures were conducted in accordance with the Declaration of Helsinki. Participation was voluntary and informed consent was obtained from all participants, with data anonymized before analysis and personally identifiable information removed. Importantly, the framework was evaluated solely as an AI-assisted screening and decision-support prototype and was not used for clinical diagnosis or treatment. Participants identified as potentially high-risk, including those expressing suicidal ideation, self-harm, psychotic symptoms, or abuse-related concerns, were advised to discontinue the assessment and seek immediate evaluation from qualified mental health professionals or emergency services, a safeguard that acknowledges the real dangers of deploying conversational AI in sensitive domains.

The significance of the work lies in how it addresses the structural bottlenecks of mental healthcare. Screening is the first and often most neglected step in the care pathway: a person must be identified as needing help before any treatment can begin. In many parts of the world, the ratio of mental health professionals to population is so low that systematic early screening is simply impossible at scale. A conversational agent that can conduct an empathetic intake, administer validated psychometric measures, and hand clinicians a structured, explainable summary of risk could extend the reach of overstretched services, flagging individuals who would otherwise remain invisible until crisis point.

The researchers are careful to position the framework as a support tool rather than an autonomous diagnostician. The system’s outputs are intended to assist healthcare professionals during initial assessment, with final clinical judgment remaining in human hands. That framing, combined with the framework’s explainability features and its grounding in DSM-5-TR criteria and validated instruments, reflects a growing consensus in medical AI research: the most promising systems are not those that attempt to replace clinical expertise, but those that structure information, surface early signals, and make their reasoning transparent enough for professionals to trust, verify, and act upon. As generative AI continues to mature, studies like this one suggest a future in which the first conversation a struggling person has about their mental health, often the hardest one to start, might be made a little easier, and a great deal more likely to lead to help.

Subject of Research: An explainable AI framework for early mental health screening and clinical decision support using reflective listening and generative AI

Article Title: An explainable AI framework for early mental health screening and clinical decision support using reflective listening and generative AI

Article References: Shakoor, M., Khan, T. A., & Ghafoor, H. (2026). An explainable AI framework for early mental health screening and clinical decision support using reflective listening and generative AI. Discover Psychology. https://doi.org/10.1007/s44202-026-00877-3

Image Credits: AI Generated

DOI: 10.1007/s44202-026-00877-3

Keywords: mental health screening, explainable AI, reflective listening, generative AI, large language models, knowledge graph reasoning, DSM-5-TR, clinical decision support, psychometric assessment, psychological disorders, early identification, conversational agents