📄 medicine

Towards a CONCORD Between Patient Perspective and Ecological Case Formulation: Conversational Large Language Models for Personalised Mental Health

This paper introduces CONCORD, an AI-supported conversational framework that generates transparent, theory-guided, and personalized causal case formulations for mental health by analyzing natural patient conversations, demonstrating superior accuracy and complexity compared to traditional questionnaire-based methods.

Original authors: Nimrod Hertz-Palmor, Amit Oren, George Phillips, Yuxi Wang, Gary Brown, Anna Bevan, Tim Dalgleish, Guy Laban

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nimrod Hertz-Palmor, Amit Oren, George Phillips, Yuxi Wang, Gary Brown, Anna Bevan, Tim Dalgleish, Guy Laban

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For decades, the standard way to track mental health has been to ask people to fill out the same checklists over and over again. Patients rate their symptoms on a scale, and doctors look for patterns in the numbers to decide if a treatment is working. While this method is useful for diagnosis, it often misses the deeper story of why a person feels the way they do. It tells us that someone is struggling, but it rarely explains how their thoughts, memories, and fears connect to keep the struggle going. To truly help someone, clinicians need to understand the specific chain of events that maintains their distress, a process known as case formulation. However, building these personalized maps of the mind is difficult, time-consuming, and often relies on patients forcing their complex inner lives into rigid boxes that don't quite fit.

A new study published by researchers from the University of Cambridge and Ben-Gurion University of the Negev suggests a different path. Instead of forcing patients to fill out surveys, the team developed a system called CONCORD that listens to natural conversations. Using advanced artificial intelligence, the system engages people in brief daily chats about their experiences. Over three weeks, eleven individuals living with post-traumatic stress disorder (PTSD) spoke with a chatbot designed to explore their thoughts and feelings without the pressure of a clinical interview. The researchers found that these simple conversations could reveal the intricate, cause-and-effect relationships that drive a person's mental health, creating a detailed, personalized picture of their condition that was far richer than what traditional questionnaires could capture.

The study involved eleven participants who had experienced significant trauma, ranging from childhood abuse to war and accidents. For three weeks, these individuals completed three different types of assessments. They filled out standard daily surveys about their symptoms, answered structured questions about how they believed their symptoms were linked, and engaged in daily text-based conversations with the CONCORD chatbot. The chatbot, powered by a large language model, asked open-ended questions like, "What happened that led to this feeling?" or "If you hadn't experienced X, do you think Y would still have happened?" This approach allowed the participants to describe their lives in their own words, rather than selecting from a list of pre-written options. The researchers then used the AI to analyze these conversations, identifying specific psychological components—such as triggers, memories, and coping strategies—and mapping out how the participants themselves believed these factors influenced one another.

The results showed that the conversations captured a much more complex and accurate picture of the participants' minds than the statistical models based on the surveys. While the traditional surveys produced sparse networks with very few connections, the conversational approach revealed dense, intricate webs of cause and effect. The system successfully identified that for these individuals, external triggers were the primary starting point of their distress. These triggers led to a sense of current threat, which then drove maladaptive coping strategies, such as avoidance or emotional numbing. This pattern closely matched a leading psychological theory of PTSD, known as the Ehlers and Clark model, which suggests that trauma is maintained by a specific cascade of thoughts and behaviors. Remarkably, the AI discovered this structure without being explicitly programmed to find it; the participants simply described their experiences, and the system organized those descriptions into a coherent map.

What makes this finding particularly significant is that the system did not just guess at these connections. Every link in the resulting map could be traced back to a specific sentence the participant had written in the chat. This means the AI did not act as a black box; instead, it acted as a translator, turning natural language into a structured clinical tool that doctors could actually understand and verify. The study demonstrated that people are naturally capable of explaining the causal chains of their own suffering when given the freedom to speak rather than the burden of rating. The conversations revealed that for most participants, the flow of distress was largely one-way: triggers led to threats, which led to coping strategies, with very little feedback looping back. This directional clarity is something that standard statistical methods, which often only show that two things happen at the same time, struggle to detect.

The researchers were careful to note that this was a proof-of-concept study with a small group of English-speaking participants, so the findings cannot yet be applied to everyone or every mental health condition. The system was designed specifically to listen and map, not to provide therapy or medical advice. However, the success of the approach suggests a promising future for mental health monitoring. If these conversational tools can be integrated into everyday digital platforms, they could offer clinicians a way to continuously update their understanding of a patient's condition without the fatigue of repetitive questionnaires. By listening to the stories people tell rather than just the numbers they provide, this method offers a way to see the unique architecture of an individual's mind, turning the chaotic noise of daily life into a clear, actionable guide for healing.

Technical Summary: CONCORD – Conversational Large Language Models for Personalised Mental Health

Problem Statement

Routine outcome monitoring in evidence-based psychology traditionally relies on standardized symptom questionnaires. While indispensable for diagnosis, these methods quantify distress without elucidating the psychological mechanisms that generate and maintain it. Effective intervention requires understanding these causal processes rather than merely measuring consequences.

Existing attempts to operationalize "case formulation" (an individualized account of underlying mechanisms) via idiographic network models face significant limitations:

Statistical Inference vs. Causal Perception: Traditional network models (e.g., Graphical Gaussian Models, Ising models) infer structure from statistical associations in Ecological Momentary Assessment (EMA) data. These associations do not establish causal influence and often require burdensome, repetitive data collection to achieve statistical power, resulting in sparse networks.
Rigidity of Elicitation: Direct causal elicitation methods, such as Momentary Structured Networks (MSN), improve interpretability by asking patients to report perceived causal links. However, these still rely on structured questionnaires and explicit causal judgments, which can feel unnatural and may miss mechanisms outside pre-defined theoretical frameworks.
The Gap: There is a need for a method that captures perceived causal relationships naturally through dialogue, reducing participant burden while maintaining clinical interpretability and alignment with established psychological theories.

Methodology

The study introduces CONCORD (Conversational Network Case Organization and Relational Detection), an AI-supported framework that constructs individualized causal case formulations from natural language conversations.

Study Design and Participants

Participants: 11 adults (aged 20–77) with significant PTSD symptomatology, recruited in the UK. Diagnoses were confirmed via SCID or PCL-5/Life Events Checklist.
Duration: Three weeks of daily monitoring.
Data Collection: Participants engaged in three parallel assessment streams:
1. CONCORD: Daily brief natural-language conversations with a chatbot (powered by OpenAI's GPT-4o-mini via the LEXI platform). The agent used theory-guided prompts (based on the Ehlers and Clark Cognitive Model of PTSD) and counterfactual reasoning (e.g., "If you had not experienced X, would Y still have happened?") to explore causal links.
2. MSN: Theory-guided causal surveys where participants selected items and explicitly rated causal influences between components.
3. EMA: Standardized symptom ratings on a 0–100 scale.
Total Data: 1,629 observations across methods, including 467 conversational exchanges comprising 5,099 messages.

Technical Implementation

Annotation Pipeline: Conversational exchanges were annotated using GPT-5 (temperature = 0) in a two-stage process:
1. Identification of the five theoretical components (Triggers, Trauma Memory, Negative Appraisals, Sense of Current Threat, Maladaptive Strategies).
2. Evaluation of directed relationships between identified components.
- Annotations included binary classifications, justifications, and supporting quotes from the client's text to ensure traceability.
Network Construction:
- Elicitation-based Networks: Derived from CONCORD and MSN data using conditional probabilities (Parent-Conditioned Probability and Child-Conditioned Probability) to quantify perceived causal strength.
- Association-based Networks: Derived from EMA and CONCORD data using Graphical Gaussian Models (GGM) and Ising models for comparison.
Validation Metrics: Network complexity, edge overlap (sensitivity, specificity, precision, F1 score), correlation of edge weights, and alignment with the Ehlers and Clark theoretical model (using Jaccard Index and odds ratios).

Key Contributions

Conversational Case Formulation: Demonstrates that natural language conversations, processed by LLMs, can extract individualized causal structures without requiring explicit causal judgments or repetitive rating scales.
Traceable AI Formulation: Addresses the "black box" concern of LLMs by linking every inferred relationship to specific excerpts from the original conversation, ensuring the formulation is auditable and clinically interpretable.
Methodological Comparison: Provides a rigorous empirical comparison between conversational elicitation (CONCORD), structured elicitation (MSN), and statistical association models (GGM/Ising).

Results

Network Complexity and Density: CONCORD networks were significantly denser and more complex than association-based networks (GGM/Ising). CONCORD showed a trend toward greater complexity than MSN networks, capturing richer, non-uniform patterns.
Convergence with MSN: CONCORD networks showed high convergence with MSN networks (correlation $b=0.55$ , $p<.0001$ ) and high overlap in edge detection (F1 score = 0.82). This suggests conversational methods can replicate the causal structures identified by structured surveys.
Theoretical Alignment: The formulations closely mirrored the Ehlers and Clark Cognitive Model:
- Causal Sources: "Triggers" emerged as the dominant causal source (Source-Sink Index = 1.48), acting as the primary driver.
- Causal Sinks: "Strategies" (maladaptive coping) functioned as the terminal sink (SSI = -1.25).
- Pathways: The strongest edges followed the theoretical cascade: Triggers $\rightarrow$ Threat $\rightarrow$ Strategies. Theory-congruent edges were nearly three times stronger than theory-incongruent edges.
- Architecture: Networks exhibited a predominantly feed-forward, hierarchical architecture (Hierarchy score = 0.76) with low reciprocity (14%), indicating a sequential causal flow rather than a recurrent feedback loop.
Validation against Association Models: While CONCORD recovered almost all edges present in association-based networks (high sensitivity), it generated many additional edges, demonstrating that natural conversation captures causal mechanisms that statistical correlations miss.

Significance and Claims

The paper claims that CONCORD offers a practical route to continuous, theory-guided case formulation that is less burdensome than traditional questionnaires. By shifting the unit of measurement from fixed questions to natural dialogue, the framework:

Reduces Burden: Eliminates the need for repetitive, structured rating scales that can reduce ecological validity.
Enhances Generativity: Allows participants to introduce novel themes and mechanisms not predefined by the researcher, while still anchoring the analysis in established theory.
Improves Interpretability: Unlike statistical models that infer direction from temporal precedence, CONCORD recovers the patient's perceived causal mechanisms directly from their narrative.
Scalability: The framework is designed to be adaptable to different theoretical models and clinical populations using widely available conversational agents.

The authors conclude that conversational ecological case formulation represents a significant step toward scalable, personalized precision mental health, enabling clinicians to monitor evolving causal architectures rather than static symptom severity. However, they note limitations regarding the small sample size, the specific focus on PTSD, and the need for future research to evaluate safety, clinical decision-making utility, and generalizability across disorders.