Multidisciplinary Team Debates

• Multidisciplinary Team debates are structured exchanges among diverse professionals that use defined phases and specialized roles to drive complex decision-making.
• They integrate multi-agent systems with LLMs and VLMs to simulate human expert debate, ensuring robust conflict resolution and evidence-based outcomes.
• These debates leverage shared mental models and SSRL frameworks to enhance communication, trust, and team adaptability in both clinical and AI settings.

Multidisciplinary Team (MDT) debates are structured, goal-oriented exchanges among professionals from diverse fields convened to solve complex problems—most notably in clinical, AI development, and human–computer interaction (HCI) domains. The MDT debate mechanism is typified by rigorous consensus-building, targeted conflict resolution, and the use of specialized artifacts and roles to facilitate collaboration. Underpinning MDT debates are both organizational constructs (such as shared mental models and facilitated workflows) and computational advances, particularly multi-agent system (MAS) frameworks leveraging LLMs, vision–LLMs (VLMs), and optimization-based team assembly systems.

1. Foundational Structures and Roles in MDT Debates

MDT debates are grounded in explicit phase structures and well-defined roles, as evidenced in ten-year ethnographies of clinical MDT meetings (Kane et al., 2019). Core stages include: pre-meeting data preparation, succinct case presentations, group consensus discussions, formal decision documentation, and post-meeting action allocation. The interaction model prioritizes round-robin specialist reporting, open floor discussion for conflicting views, and decisive leadership from the meeting chair. Key roles span clinical domains (e.g., radiologist, pathologist, surgeon, oncologist, nurse, data manager) and, in AI contexts, extend to data scientists, domain experts, and product stakeholders (Piorkowski et al., 2021).

The integration of multi-agent architectures further formalizes specialist roles in computational MDTs. In therapy planning MAS (Wu et al., 15 Jul 2025), a general practitioner agent coordinates conflict detection and delegates to specialist agents, with consensus facilitation managed via mediator agents in case of deadlock. Hierarchical agent tiers in VLM frameworks (e.g., UCAgents (Feng et al., 2 Dec 2025)) mirror real-world clinical verification chains—junior specialists provide independent reads, senior experts audit for evidence alignment, and critical analysis is conducted by adversarial agents whose positions are fixed to suppress rhetorical drift.

2. Theoretical Models: Shared Mental Models and SSRL

The stability and efficiency of MDT debates depend on the construction and regimented use of shared mental models (SMMs) (Piorkowski et al., 2021), defined as team members’ common understanding of task responsibilities, information needs, and role boundaries. SMMs are operationalized through components such as the task model (e.g., procedures and algorithmic states), team model (communication channels, domain expertise), and principled interaction mechanisms: consistency, reactivity, proactivity, coordination, and knowledge stability [Scheutz et al. 2017].

The Socially Shared Regulation of Learning (SSRL) framework extends SMM principles to dynamical team processes, where regulation is codified as a time-evolving vector across cognitive, metacognitive, emotional, and coordination domains (Huang et al., 2 May 2025). SSRL supports real-time monitoring and responsive adaptation, and is quantifiable as $$\mathrm{Coord}(t) = \|\,[x_c(t),\,x_m(t),\,x_e(t),\,x_r(t)]\,\|_2,$$ with higher norms correlating with adaptive decision performance.

3. Computational Frameworks and Algorithmic MDT Debate Models

LLM-based and VLM-based multi-agent frameworks have been extensively developed to simulate MDT debates computationally. The MAS implementation for therapy recommendations structures agent–agent message passing, conflict-targeted specialist allocation, iterative consensus mechanisms, and final plan integration (Wu et al., 15 Jul 2025). Simulation protocols are formalized, e.g., agent decisions and consensus are driven by

$$C = \{C_1, \dots, C_K\}\;\text{detected conflicts},\quad MDT_k \subseteq \{\text{specialists}\}$$

with multi-round evaluation and a strict integration function for revised prescriptions.

MDTeamGPT (Chen et al., 18 Mar 2025) organizes debate in multi-round cycles, utilizing a lead physician for consensus aggregation, residual discussion summaries for memory control, and self-evolving knowledge bases—CorrectKB for validated answers, ChainKB for error reflection. Consensus is distilled into four dimensions: consistency, conflict, independence, integration, sharply reducing context drift, lowering token usage, and enabling continual learning via retrieval of closest past cases.

UCAgents (Feng et al., 2 Dec 2025) enforces hierarchical, unidirectional convergence using structured evidence auditing—agents maintain fixed initial positions, verifying against image data in sequential tiers, culminating in chair-led arbitration. One-round inquiry protocols are used rather than open-ended dialogues, quantitatively constraining both textual and visual noise (dual-noise bottleneck), and maximizing mutual information along the chain $I(Y;\mathcal{I}) = I(Y;V) + I(Y;T|V) + I(Y;M|V,T)$ under entropy constraints.

4. Communication, Conflict, and Trust Management

MDT debates routinely encounter knowledge gaps (disciplinary vocabulary mismatches), trust gaps (skepticism of opaque systems), and expectation gaps (difference between deterministic and probabilistic reasoning) (Piorkowski et al., 2021). Practices to surmount these include:

• Proactive education and alignment sessions (fundamental workshops, domain briefings)
• Pilot deployments and rapid demos showcasing interpretability
• Shared documentation (living FAQs, model overview documents) fostering consistency
• Tailored visualizations (e.g., confusion matrices, feature-importance charts) mapped to stakeholder metrics
• Storytelling and analogical explanation to bridge domain abstraction

Trust is maintained via persistent transparency in metrics and failure cases, regular sync-ups to reduce surprises, calibrated expectations regarding metric plateaus, and iterative clarification of goal definitions.

5. Evaluation Metrics and Quantitative Benchmarks

MDT debate frameworks are assessed on multidimensional clinical and technical metrics. In LLM therapy MAS, precision ($\frac{TP}{TP + FP_w}$), recall ($\frac{TP}{TP+FN}$), DDI ratio ($$\frac{\#\,\text{DDIs revised}}{\#\,\text{original}}$$), contraindications ratio, met-goals ratio ($GR$), medication ratio ($MR$), and preferred option coverage are tabulated across baseline systems. MAS architectures achieved precision/recall (0.90/0.74) with substantive reduction in drug–drug interactions and contraindications compared to single-agent and pure baselines, though completeness remains a challenge (Wu et al., 15 Jul 2025). Similarly, MDTeamGPT delivered accuracy of 90.1% (MedQA), 83.9% (PubMedQA), with ablations revealing marked dependence on structured consensus and knowledge base evolution (Chen et al., 18 Mar 2025).

The UCAgents architecture demonstrated 71.3% accuracy (+6.0 pp over prior multi-agent state-of-the-art) and an 87.7% reduction in token budget due to debate constraints, supporting higher reliability and lower computational cost (Feng et al., 2 Dec 2025).

6. Team Formation, Diversity, and Human Factors

Team assembly in MDT debates is governed by both expertise diversity and interpersonal familiarity, which are quantified and optimized via multi-objective evolutionary algorithms (NSGA-II) and contextual bandit models (Almutairi et al., 5 Jun 2025). Formally,

$$D(T) = \frac{1}{|T|(|T|-1)}\sum_{i\neq j} \operatorname{Ham}(e_i,e_j),\quad F(T) = \frac{1}{|T|(|T|-1)}\sum_{i\neq j} w_{ij}$$

where $D(T)$ measures diversity (via Hamming distance of skill vectors) and $F(T)$ quantifies prior collaboration.

Guidelines for MDT debate optimization recommend small, cross-cutting teams (3–5 members), explicit facilitation, rapid expertise mapping, and alternating diversity/familiarity boosters. Reflection stages are built in to detect process misalignment and surface latent challenges. The SSRL framework (Huang et al., 2 May 2025) incorporates multimodal sentiment and physiological analyses to monitor and tune emotional climate and motivation, impacting knowledge exchange rates and decision quality.

7. Opportunities, Pitfalls, and Future Directions

Machine learning systems integrated into MDTMs present significant opportunities for rapid information access (e.g., audio-indexed case retrieval dropping from 15–20 min to ~2 min), semi-automated documentation, and guideline adherence support (Kane et al., 2019). Key pitfalls include limited generalizability of ML classifiers, privacy and security risks, cognitive overload from over-automation, and complex regulatory landscapes for clinical deployment.

Emerging design principles advocate: conflict-targeted MDT assembly, structured multi-round workflows, hybrid retrieval for guideline grounding, and dynamic model selection. Continual learning frameworks (CorrectKB/ChainKB) and audit-traceable reasoning logs enable robust error analysis and meta-reflection. A plausible implication is that fully autonomous MDT debate models will require further augmentation for deep guideline citation and more comprehensive therapeutic option enumeration.

In summary, MDT debates—whether human-driven or computationally simulated—rely on a layered system of structured interaction protocols, principled consensus mechanisms, diversity optimization, and adaptive communication practice to address the intrinsic complexity of multidisciplinary collective reasoning. These approaches are evidenced to deliver superior safety, transparency, and collaborative efficiency, provided they are meticulously architected and continually validated in real-world settings (Piorkowski et al., 2021, Wu et al., 15 Jul 2025, Chen et al., 18 Mar 2025, Feng et al., 2 Dec 2025, Huang et al., 2 May 2025, Almutairi et al., 5 Jun 2025, Kane et al., 2019).