A Hybrid Cognitive-Domain Framework for Intent Recognition and Action in Medical AI

Abstract

The efficacy and safety of AI-based medical assistants depend critically on their ability to accurately interpret the latent intent behind a user’s query and respond appropriately. Existing frameworks are often either too rigid, failing to capture the user’s nuanced cognitive state, or too general, lacking the structured communication paradigms essential for medical discourse. This paper introduces a novel hybrid framework that synthesizes two models: the Mutually Exclusive, Collectively Exhaustive (MECE) model for structured domain classification and the theoretical principles of a Universal Discourse Core (UDC), a cognitive model grounded in established theories of communication. We propose a multi-layer architecture where a “Cognitive State Encoder,” based on UDC principles (Epistemic, Intentional, Affective), first analyzes the underlying communicative goal. This cognitive assessment then enriches the input for a “MECE Domain Router” and, crucially, selects an appropriate structured response paradigm. By mapping cognitive states to clinically-aligned communication models (e.g., SBAR, SOAP, Teach-Back), our framework moves beyond simple classification to intent-driven action. This MECE-UDC model aims to improve discourse alignment, reduce misinterpretation risk, and provide a robust foundation for building trustworthy AI assistants in high-stakes environments like healthcare. We conclude by proposing a validation protocol, including novel metrics to evaluate both intent recognition and response quality.

1. Introduction

The integration of Artificial Intelligence (AI) into clinical workflows, particularly through conversational assistants, promises to revolutionize how medical professionals access information, support decisions, and manage patient care. However, the unique nature of medical discourse — characterized by high stakes, specialized terminology, and nuanced, often implicit, user intent — presents a formidable challenge. A query like “information on aspirin” can represent vastly different needs: a request for general knowledge, a specific dosage question, or an inquiry into clinical trials. Misinterpreting this latent intent can lead to inefficient, irrelevant, or, in the worst case, clinically unsafe outcomes.

To address this, developers often employ structured classification systems, such as the Mutually Exclusive, Collectively Exhaustive (MECE) framework, to categorize query types. However, while effective for structuring domains of knowledge, MECE by itself is insufficient as it primarily classifies the topic of a query, not the user’s underlying cognitive or communicative goal.

This paper identifies this critical gap and proposes a solution by integrating the domain-specific structure of MECE with a cognitive-centric model, which we term the Universal Discourse Core (UDC). Rooted in established research on Relevance Theory (Sperber & Wilson, 1986), Joint Intentionality (Tomasello, 2008), and Affective Computing (Picard, 1997), the UDC provides a way to model the user’s underlying “why.”

Our central thesis is that a truly effective medical AI assistant requires not just a two-layer interpretation process, but a direct link from that interpretation to a corresponding communicative action. It must first understand the cognitive state behind an utterance before classifying its domain, and then use that cognitive understanding to formulate its response in a structured, professionally recognized format. This paper proposes a formal architecture for this hybrid MECE-UDC framework, detailing its components, its workflow from intent recognition to responsive action, and a methodology for its empirical validation.

2. Foundational Frameworks & Literature Review

2.1 The MECE Framework for Medical Discourse

Originating in management consulting, the MECE principle requires that information be organized into categories that are mutually exclusive (no overlaps) and collectively exhaustive (no gaps). As applied to a medical AI assistant, this framework provides a robust, top-down classification of professional medical discourse into five primary domains:

1. Clinical Formulation & Management: Patient-specific diagnosis and treatment.
2. Knowledge Synthesis & Education: Abstract medical knowledge retrieval.
3. Research & Investigation: Clinical trial design and evidence synthesis.
4. Population & System-Level Health: Public health and administrative tasks.
5. System & Meta-Interaction: User interaction with the AI system itself.

The strength of this approach lies in its structure, but its limitation is its rigidity in the face of ambiguity where the intent, not the topic, is the key differentiator.

2.2 The Universal Discourse Core (UDC) as a Cognitive Model

To address this limitation, we introduce the Universal Discourse Core (UDC), a concept further detailed by Soshnin (2025) as a cognitive infrastructure. This model posits that any utterance can be decomposed into a vector of three primitives representing the communicator’s cognitive state:

• Epistemic Primitive (Pe): Manages the exchange of knowledge. Grounded in Relevance Theory (Sperber & Wilson, 1986), it reflects an intent to change the knowledge state of the listener (e.g., asking a question, providing an explanation).
• Intentional Primitive (Pi): Manages the alignment of goals and plans. Based on Joint Intentionality (Tomasello, 2008), it reflects an intent to influence behavior or coordinate action (e.g., giving a command, making a recommendation).
• Affective Primitive (Pa): Manages the emotional and social dimension. Drawing from Affective Computing (Picard, 1997), it reflects the presence of significant emotional valence in an utterance (e.g., expressing anxiety, frustration).

2.3 The Gap: From Domain and Cognition to Action

The MECE framework provides the map of medical discourse, while the UDC provides the compass. However, simply knowing the location and the bearing is not enough; the system must know how to act at the destination. The critical missing link is the set of instructions that translate a recognized intent into a safe and effective communicative action.

While sophisticated intent-recognition models exist in non-medical domains (e.g., in customer service platforms like Google Dialogflow or Rasa), they are primarily optimized for task completion and topic routing. Their architecture does not inherently account for the high-stakes nature of professional discourse, where the manner and structure of a response are as critical as its factual content. Our proposed framework is novel in its assertion that for a domain like medicine, recognizing the cognitive intent is not merely a means to better classification, but a necessary prerequisite for selecting a safe and professionally aligned communicative behavior.

A hybrid approach is therefore necessary not only to bridge the Domain-Cognition gap, but to connect them both to structured, clinically-relevant response generation.

3. The Hybrid MECE-UDC Architecture: From Intent to Action

We propose an architecture that leverages both frameworks to process a query, classify its intent and domain, and generate a structured, context-aware response.

3.1 Architectural Overview

1. Layer 1: Cognitive State Encoder (CSE): An input utterance is first passed to the CSE. This module analyzes its linguistic features to determine its underlying cognitive composition, outputting a cognitive state vector (CSV), ($\vec{v} = [a_E, a_I, a_A]$), where each component represents the activation weight of the Epistemic, Intentional, and Affective primitives.
2. Layer 2: Enhanced MECE Router: The original utterance, now augmented with its cognitive state vector ($\vec{v}$), is passed to the MECE Router for robust, context-aware domain classification.
3. Layer 3: Structured Response Generator: Based on the dominant primitive in the CSV, the system selects a pre-defined, structured communication paradigm to formulate its response. This ensures the AI’s behavior aligns with the user’s latent communicative goal.

3.2 Layer 1 & 2 in Action: An Example

• Query: “This patient has a history of GI bleeds, what about aspirin for them?”
• CSE Output (Layer 1): High Intentional activation ($\vec{v} = [0.3, 0.8, 0.4]$) due to the patient-specific context and implicit goal of making a clinical decision.
• Router Decision (Layer 2): The high intentional score directs the query to Domain 1: Clinical Formulation & Management.

3.3 Layer 3: From Cognitive State to Communicative Action

This stage is the primary contribution of our framework. It operationalizes the insights from the CSE by assigning a specific “persona” or behavioral model to the AI based on the user’s cognitive state. The practical realization of this Structured Response Generator can be achieved in Large Language Models through targeted prompt engineering, a methodology analogous to Constitutional AI principles (Bai et al., 2022), as detailed in the following section.

3.4 Detailing the Cognitive State Encoder (CSE)

The efficacy of the entire framework hinges on the CSE’s ability to accurately translate a user’s utterance into a cognitive state vector ($\vec{v}$). This module is envisioned not as a simple classifier but as a sophisticated multi-headed model, likely built upon a transformer architecture fine-tuned on medical corpora (e.g., BioBERT, ClinicalBERT). Its task is to analyze the query for a rich set of linguistic features indicative of underlying intent. Key features for analysis would include:

• Pragmatic and Syntactic Features: The model would be trained to recognize sentence moods (e.g., interrogative for Epistemic intent, imperative for Intentional) and the presence of modal verbs that signal speculation or obligation (“could,” “should,” “must”).
• Lexical and Semantic Features: The CSE would analyze word choice, identifying sentiment-bearing words (e.g., “worried,” “confused”) to gauge Affective activation ($a_A$). The presence of patient-specific context versus abstract terminology would help differentiate Intentional from Epistemic queries.
• Attention-Based Analysis: Using attention mechanisms, the model can determine which parts of the query are most salient to the user’s goal, allowing it to weigh the different cognitive primitives appropriately. For example, in “I’m worried this patient’s history of GI bleeds means we can’t use aspirin,” the model would learn to place high weight on both “worried” (Affective) and “can’t use” (Intentional).

The output would be a continuous vector (e.g., [0.3, 0.8, 0.4]) rather than a single class, allowing the system to handle the blended intents common in real-world discourse.

4. Practical Implementation via Prompt Engineering

The mapping of a cognitive state to a communicative action can be implemented by dynamically loading a “constitutional” system prompt that governs the LLM’s behavior. This prompt defines the AI’s persona, rules, and output format for the duration of the interaction.

4.1 Responding to Epistemic Intent: The “Educator” Persona

When high Epistemic activation ($a_E$) is detected, the system loads the “Educator” persona. The goal is to maximize comprehension and retention.

System-Level Constitutional Prompt: You are a medical educator AI. Your primary goal is to explain complex medical topics with maximum clarity. 1. Break down information into simple, logical chunks. 2. Define all medical jargon in plain language. 3. Use analogies where helpful. 4. After presenting the core information, you MUST ask the user to explain it back in their own words to verify understanding (The “Teach-Back” method). You will then confirm or gently correct their understanding.

Example Interaction:

User Query: “Can you give me information on metformin?”

LLM Response (as Educator): “Of course. Metformin is a first-line medication for managing type 2 diabetes. Let’s break down how it works. 1. It reduces glucose production: Think of your liver as a sugar factory. Metformin tells this factory to slow down its production of glucose, the sugar that enters your bloodstream. 2. It improves insulin sensitivity: It makes your body’s cells more responsive to insulin, the hormone that helps cells absorb sugar for energy. To make sure I’ve been clear, could you tell me in your own words the two main ways metformin helps control blood sugar?”

4.2 Responding to Intentional Intent: The “Collaborator” Persona

For queries with high Intentional activation ($a_I$), the AI adopts a “Collaborator” persona, designed to structure information for clinical decision-making.

System-Level Constitutional Prompt: You are a clinical decision support AI. Your role is to act as a structured collaborator. 1. Your task is to synthesize clinical information into a professionally recognized format. 2. For patient-specific queries about assessment and planning, you MUST format your response using the SOAP (Subjective, Objective, Assessment, Plan) note structure. 3. Your output must be objective and based only on the information provided and established clinical knowledge. State any assumptions. 4. Always conclude with a disclaimer advising consultation with clinical guidelines and a human expert.

Example Interaction:

User Query: “This patient has a history of GI bleeds, what about aspirin for them?”

LLM Response (as Collaborator): “Synthesizing this query into a SOAP format for review: S (Subjective): The patient has a known history of gastrointestinal (GI) bleeds. O (Objective): The clinical question is regarding the suitability of prescribing aspirin. A (Assessment): Aspirin is an antiplatelet agent that impairs clotting. In a patient with a history of GI bleeding, aspirin presents a significant risk of precipitating or exacerbating a bleed. The risk versus benefit must be carefully evaluated based on the indication for aspirin (e.g., primary vs. secondary prevention of cardiovascular events). P (Plan): 1. Recommend assessing the indication for antiplatelet therapy. 2. Recommend considering alternative agents with a lower GI risk profile or the addition of a proton pump inhibitor (PPI) for gastric protection. 3. Strongly advise consulting up-to-date clinical guidelines and a specialist (e.g., cardiologist, gastroenterologist) before making a decision. Disclaimer: This is an AI-generated synthesis for informational purposes and is not a substitute for professional medical judgment.”

4.3 Responding to Affective Intent: The “Supporter” Persona

When high Affective activation ($a_A$) is detected, the AI prioritizes de-escalation and empathy before providing information, adopting the “Supporter” persona.

System-Level Constitutional Prompt: You are a supportive assistant AI. Your primary directive is to handle emotionally charged queries with care. 1. Acknowledge and validate the user’s expressed emotion (e.g., worry, fear, frustration) with a brief, empathetic statement. 2. Do NOT offer diagnoses or false reassurances. 3. After the initial empathetic response, cautiously and clearly address the user’s informational request. 4. Maintain a calm, non-judgmental, and supportive tone throughout.

Example Interaction:

User Query: “I just got my lab results back and the numbers are all over the place, I’m so scared. What does elevated ALT mean?”

LLM Response (as Supporter): “I understand that getting unexpected lab results can be a very worrying and stressful experience. It’s completely normal to feel scared when you see numbers you don’t understand. Let’s address your question. ALT stands for Alanine Aminotransferase. It is an enzyme found mostly in the liver. An ‘elevated ALT’ result means the level of this enzyme in your blood is higher than the typical range. This can be an indicator of liver inflammation or damage, but it can also be caused by many other factors. It is very important to discuss these results with the doctor who ordered them, as they can interpret them in the context of your full medical history and determine what they mean for you.”

5. Proposed Validation Methodology

Empirically validating this framework requires evaluating both the accuracy of the classification and the appropriateness of the subsequent action.

5.1 Evaluation Metrics

1. Discourse Task Success Rate (DTSR): A holistic measure of whether the system’s complete response successfully fulfills the user’s goal.
2. Intentional Alignment Score (IAS): An expanded metric where expert evaluators rate, on a Likert scale, both if the system correctly identified the primary cognitive intent and if it selected the correct structured response format (e.g., was a SOAP format appropriate for the query?).
3. Framework Adherence Score (FAS): A new metric to assess how faithfully the AI’s output adheres to the principles of the chosen communicative framework (e.g., “Did the SBAR response clearly and correctly delineate all four components?”).
4. Clinical Safety Metric (CSM): A binary metric where clinicians evaluate if a misinterpretation or inappropriate response could lead to a clinically adverse event.

5.2 Benchmark Dataset

Validation requires the creation of a new Medical Intent & Cognition (MIC) dataset. Queries would be annotated with:

• The ground-truth MECE Domain.
• A ground-truth Cognitive State Vector ($\vec{v}$).
• The optimal response framework (e.g., SOAP, Teach-Back).
• A Clinical Risk Score associated with misinterpretation.

A significant challenge in creating the MIC dataset is ensuring high inter-annotator reliability, especially for abstract labels like the Cognitive State Vector. To address this, we propose a preliminary pilot study to develop a robust annotation protocol. This would involve a panel of clinical and linguistic experts who would collaboratively create a detailed “annotation codex” with explicit rules and examples. Rater agreement would be iteratively tested and refined using a rigorous statistical measure appropriate for subjective data, such as Krippendorff’s alpha, before commencing the full-scale dataset annotation. This ensures the ground-truth data is consistent and reliable.

6. Discussion and Future Work

The proposed MECE-UDC framework represents a significant step towards creating safer medical AI assistants. By explicitly modeling the cognitive dimension of discourse and linking it directly to structured, clinically aligned communicative actions, the system can better navigate ambiguity and align with user intent. This contribution bridges the critical gap between passive intent recognition and active, useful response generation.

The potential impact of this cognitive-action mapping extends beyond conversational assistants. The core principles could be applied to analyze physician notes within Electronic Health Records (EHRs) to identify moments of diagnostic uncertainty or to flag potential misalignments between documented assessments and plans. Furthermore, it could power highly adaptive medical training software that adjusts its teaching style and content structure based on a student’s pattern of inquiry, recognizing whether they are in a knowledge-gathering (Epistemic) or a decision-making (Intentional) mode.

The primary limitation remains the theoretical nature of the full system. Future work must focus on the practical implementation and validation outlined in Section 5, including the significant undertaking of creating the MIC dataset. Further research could also explore more granular mappings between cognitive state mixtures and blended response strategies.

7. Conclusion

Standard intent classification models are insufficient for the complexity of medical discourse. We have argued that a hybrid architecture, combining the domain mapping of MECE with the cognitive insight of a UDC model, provides a more robust path forward. Crucially, we have extended this by introducing a methodology for translating cognitive states into validated medical communication paradigms, practically achievable through dynamic prompt engineering. By enabling an AI to understand not just what the user is asking about, but what the user is doing with their words, and then responding in kind, this framework lays the groundwork for the next generation of AI tools that can earn the trust of medical professionals and be safely integrated into the fabric of healthcare.

Acknowledgments

The conceptualization, drafting, and revision of this article were performed by the AI model Gemini, under the guidance of the author through prompting.

8. References

• Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv. https://arxiv.org/abs/2212.08073
• Picard, R. W. (1997). Affective Computing. MIT Press. https://mitpress.mit.edu/9780262161701/affective-computing/
• Soshnin, A. (2025). A Universal Discourse Core: A Proposed Cognitive Infrastructure for AGI. Medium. https://medium.com/@alexeysoshnin/a-universal-discourse-core-a-proposed-cognitive-infrastructure-for-agi-fbd34b4a5742
• Sperber, D., & Wilson, D. (1986). Relevance: Communication and Cognition. Blackwell. https://www.wiley.com/en-us/Relevance%3A+Communication+and+Cognition%2C+2nd+Edition-p-9780631198789
• Tomasello, M. (2008). Origins of Human Communication. MIT Press. https://mitpress.mit.edu/9780262201773/origins-of-human-communication/