MDT-mimic MAS for Clinical Decision Support

• MDT-mimic MAS is a class of multi-agent systems that replicates multidisciplinary teams by mapping specialist roles to LLM agents for clinical decision support.
• These systems structure agent interactions through round-based debates, role-specific messaging, and consensus matrices to ensure inspectable and auditable deliberation.
• Evaluation studies demonstrate improved evidence tracking and conflict resolution with enhanced clinician control, though performance gains vary with task complexity.

MDT-mimic MAS denotes a class of multi-agent systems that reproduces the role structure, information flow, and deliberative dynamics of a multidisciplinary team rather than attempting an autonomous replacement. In the recent clinical literature, the term is used most directly for large-language-model systems in which role-specialized agents interpret a shared case, exchange evidence-backed opinions, surface conflicts, and produce a consensus or decision summary that remains inspectable and steerable by clinicians (Kuai et al., 30 Mar 2026, Zhang et al., 14 Feb 2026). Related systems extend the same design logic to oncology consensus formation, multimorbidity therapy planning, and emergency medical dispatch, but they differ sharply in how they formalize expertise, organize interaction, and evaluate collaboration (Han et al., 16 Dec 2025, Wu et al., 15 Jul 2025, Li et al., 24 Oct 2025).

1. Conceptual framing and scope

In the clinical AI literature represented here, MDT-mimic MAS is grounded in the observation that many difficult decisions are inherently distributed across specialties. Rare-disease diagnosis is described as requiring a multidisciplinary team because such diseases have heterogeneous, multisystem symptoms and fragmented evidence across specialties; ovarian tumour management is described as relying on multidisciplinary tumour board deliberation across the care continuum; multimorbidity therapy planning is motivated by the way general practitioners occasionally convene multidisciplinary team collaboration to resolve treatment conflicts (Kuai et al., 30 Mar 2026, Zhang et al., 14 Feb 2026, Wu et al., 15 Jul 2025). In these settings, the multi-agent system maps specialist roles to LLM agents, and maps team deliberation to structured rounds, targeted consultations, or consensus procedures.

A defining theme is that the virtual MDT is usually framed as decision support rather than as a substitute for clinical authority. MDTRoom states that fully automated MDT-like diagnosis is unrealistic and therefore treats the system as a shared workspace between human clinicians and MAS agents (Kuai et al., 30 Mar 2026). OMGs is positioned similarly: it is intended to provide MDT-style, evidence-backed decision support where MDT resources are scarce or unavailable, while preserving explicit rationales and reassessment triggers (Zhang et al., 14 Feb 2026). The safer therapy recommendation study adopts a GP-centred orchestration in which a GP agent frames the case, detects conflicts, convenes a minimal specialist set, and integrates recommendations into a final plan (Wu et al., 15 Jul 2025).

The same underlying pattern also appears in protocolized operational workflows. DispatchMAS is described as a “medical dispatch team–mimic MAS”: instead of a hospital MDT, it models the structured interaction between Caller and Dispatcher Agents within a six-phase emergency medical dispatch protocol grounded in a clinical taxonomy and a fact commons (Li et al., 24 Oct 2025). This suggests that “MDT-mimic” can denote not only replication of a tumour board or diagnostic conference, but more generally the reproduction of expert role allocation, protocolized sequencing, and deliberative control in high-stakes medical workflows.

2. Core architecture and formal components

Most MDT-mimic MAS instantiate a specialist pool, a structured case representation, and a coordination mechanism that converts free-form language generation into auditable objects. In MDTRoom, a clinician enters a free-text case narrative, Qwen3-Max parses it into editable structured items,

$$\text{Case} = \{ \text{Demographics}, \text{History}, \text{Exam}, \text{Labs}, \text{Imaging}, \dots \},$$

and the clinician selects specialty roles such as Infectious Disease, Cardiology, Gastroenterology, Rheumatology, Neurology, Hematology, and Differential Diagnosis (Kuai et al., 30 Mar 2026). Debate proceeds in rounds $R_1,\dots,R_T$; at round $r$, each agent outputs a leading hypothesis $h_i^{(r)}$, a reasoning chain, and cited evidence,

$$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$

The group state is summarized by support counts and support proportions,

$$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$

which function as a group-level MDT vote (Kuai et al., 30 Mar 2026).

OMGs adopts a stricter tumour-board decomposition. Its five specialty agents are Chair, Medical oncologist, Radiologist, Pathologist, and Nuclear medicine physician, coordinated by an agent orchestrator and shared agent servers for structuring, context assembly, evidence retrieval, report selection, and provenance tracking (Zhang et al., 14 Feb 2026). Raw EHR is transformed by

$X = f_{\text{struct}}(X_{\text{raw}})$

into schema-conformant JSON with explicit fields and document-level provenance. The chair’s final output is constrained to

$$Y = (\text{Final Assessment}, \text{Core Treatment Strategy}, \text{Change Triggers}),$$

with citations to patient reports, guidelines, trials, and PubMed records (Zhang et al., 14 Feb 2026).

The oncology “Consensus Matrix System” formalizes the same idea with seven role-specialized agents: oncologist, radiologist, nurse, psychologist, patient advocate, nutritionist, and rehabilitation therapist (Han et al., 16 Dec 2025). Each agent emits a structured opinion

$$o_i = (\mathbf{p}_i, r_i, \kappa_i, \mathcal{Z}_i, \mathcal{E}_i),$$

where $\mathbf{p}_i$ is a treatment preference vector, $R_1,\dots,R_T$0 is reasoning, $R_1,\dots,R_T$1 is confidence, $R_1,\dots,R_T$2 is a concern list, and $R_1,\dots,R_T$3 is an evidence chain (Han et al., 16 Dec 2025). These are aggregated into a consensus matrix $R_1,\dots,R_T$4, and agreement is quantified by Kendall’s coefficient of concordance,

$R_1,\dots,R_T$5

A threshold of $R_1,\dots,R_T$6 is treated as consensus achieved (Han et al., 16 Dec 2025).

System	Roles/domain	Distinguishing mechanism
MDTRoom	Rare-disease diagnosis; specialist LLMs plus clinician	Round-based debate with structured hypotheses, evidence links, and conflict objects
OMGs	Ovarian tumour MDT	Role-scoped evidence retrieval, trigger-controlled deliberation, standardized MDT summary
Consensus Matrix System	Oncology MDT	Seven agents, consensus matrix, Kendall’s $R_1,\dots,R_T$7, RL-optimized interaction
Safer therapy MAS	Multimorbidity therapy planning	GP-led conflict detection, conflict-specific MDT formation, mediator-assisted consensus
DispatchMAS	Emergency medical dispatch	Taxonomy-grounded six-phase protocol, fact commons, Caller and Dispatcher Agents

3. Interaction design, inspectability, and human control

A central distinction within MDT-mimic MAS is whether the system merely simulates a team or makes the team’s state inspectable and steerable. MDTRoom is explicitly built to transform raw multi-agent chat into a structured workspace. It maintains explicit mappings from patient data items $R_1,\dots,R_T$8 to agent usage $R_1,\dots,R_T$9, and marks a datum as conflicting at round $r$0 when

$r$1

Its interface is organized into a Report and Evidence View, an Agent Discussion View, and a Conflict View keyed by round index $r$2, so that patient data, hypothesis trajectories, stance changes, and unresolved disagreements become first-class visual objects rather than latent properties of a transcript (Kuai et al., 30 Mar 2026).

Human intervention is similarly formalized. In MDTRoom, clinicians select one or more data items $r$3, write a natural-language explanation or correction $r$4, choose a target subset of agents $r$5, and trigger a revision round: $r$6 Because unaffected agents retain prior state, clinicians can compare pre- and post-intervention hypotheses, support proportions, and conflict status (Kuai et al., 30 Mar 2026). This makes “steerability” operational rather than rhetorical.

OMGs uses a more tightly gated interaction policy. Specialist contributions are not free-form; the orchestrator permits them only when predefined triggers are met, including inter-role conflicts, safety concerns, missing critical information, and new decision-relevant evidence (Zhang et al., 14 Feb 2026). Contributions are role-directed and tagged, such as [safety], [conflict], [missing], or [new]. The resulting transcript is preserved, but the final chair synthesis remains schema-constrained and citation-backed. The system therefore combines deliberation with provenance and post hoc auditability (Zhang et al., 14 Feb 2026).

The consensus-matrix approach operationalizes inspectability differently. It stores role-specific preferences, confidence, concerns, and evidence chains, then exposes the consensus matrix across up to three discussion rounds, identifies discordant agents by deviation from the group, and applies targeted feedback when $r$7 (Han et al., 16 Dec 2025). This does not externalize patient-chart provenance as richly as MDTRoom, but it provides a mathematically explicit representation of agreement, concern burden, and round-by-round convergence.

4. Evaluation paradigms and empirical results

MDT-mimic MAS are evaluated on more than raw accuracy. MDTRoom was tested with 12 clinicians from diverse specialties with 1 to 15 years of experience in a within-subjects study. Each participant diagnosed two rare, complex cases, one with MDTRoom and one with a transcript-only baseline using the same multi-agent backend (Kuai et al., 30 Mar 2026). On understanding multi-agent reasoning, MDTRoom improved “Track hypothesis evolution” from $r$8 to $r$9, with $h_i^{(r)}$0; “Understand evidence usage” from $h_i^{(r)}$1 to $h_i^{(r)}$2, with $h_i^{(r)}$3; and “Identify conflicts and resolution” from $h_i^{(r)}$4 to $h_i^{(r)}$5, with $h_i^{(r)}$6 (Kuai et al., 30 Mar 2026). It also improved sense of control, willingness to keep exploring, and feeling of collaboration, while lowering NASA-TLX mental demand from $h_i^{(r)}$7 to $h_i^{(r)}$8 and frustration from $h_i^{(r)}$9 to $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$0 (Kuai et al., 30 Mar 2026).

OMGs evaluates recommendation quality through SPEAR, with Safety, Personalization, Evidence, Actionability, and Robustness each scored on a 1–5 Likert scale. The overall score is safety-gated: $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$1 In multicentre re-evaluation on a scene-balanced FUSCC subset of $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$2, OMGs achieved an overall safety-gated SPEAR of $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$3 versus $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$4 for re-MDT, with higher Evidence scores $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$5 versus $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$6 and higher Robustness $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$7 versus $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$8 (Zhang et al., 14 Feb 2026). In prospective multicentre evaluation on 59 patients, all 95% confidence intervals for overall SPEAR differences lay within the $$\text{Agent}_i: (D, \text{History}, \text{Instructions}_i) \mapsto \left( h_i^{(r)}, \text{reason}_i^{(r)}, E_i^{(r)} \right).$$9 equivalence margin, and paired human–AI studies showed the largest gains in Evidence and Robustness, especially for residents and non-tertiary physicians (Zhang et al., 14 Feb 2026).

The consensus-matrix oncology framework reports broader benchmark-style performance. Across MedQA, PubMedQA, DDXPlus, MedBullets, and SymCat, it achieved an average accuracy of $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$0 compared with $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$1 for the strongest baseline, a consensus achievement rate of $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$2, and a mean Kendall’s $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$3 of $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$4 (Han et al., 16 Dec 2025). Expert reviewers rated the clinical appropriateness of outputs at $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$5, and among the integrated RL methods, PPO yielded $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$6 accuracy, $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$7, and convergence rate $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$8, compared with $$c_j^{(r)} = \bigl|\{ i : h_i^{(r)} = H_j \}\bigr|,\qquad p_j^{(r)} = \frac{c_j^{(r)}}{\sum_{k=1}^K c_k^{(r)}},$$9, $X = f_{\text{struct}}(X_{\text{raw}})$0, and convergence rate $X = f_{\text{struct}}(X_{\text{raw}})$1 for the no-RL condition (Han et al., 16 Dec 2025).

Not all empirical findings favor explicit multi-agent decomposition. In safer therapy recommendation for multimorbidity, the reported result is that, with current LLMs, a single agent GP performs as well as MDTs (Wu et al., 15 Jul 2025). That study introduced clinically oriented metrics beyond precision and recall, including DDI Ratio, Contraindication Ratio, Met Goals Ratio, and Medication Ratio, and found that correctness was often high but completeness remained limited. Omission errors dominated, and some models introduced unnecessary medications, increasing medication burden and new conflicts (Wu et al., 15 Jul 2025). This is an important counterpoint: MDT mimicry may improve explanation structure or conflict handling without necessarily improving final plan quality.

5. Design patterns, benchmarks, and system families

Several recurrent design patterns emerge across these systems. MDTRoom’s design lessons are explicit: use multiple specialist agents with clearly defined specialty roles; structure conversation into rounds with explicit hypothesis commitments; track per-agent hypotheses $X = f_{\text{struct}}(X_{\text{raw}})$2 and group support $X = f_{\text{struct}}(X_{\text{raw}})$3; maintain explicit mappings from reasoning to patient data and literature; elevate conflict to a central object; support targeted, data-grounded clinician interventions; and evaluate understanding, control, engagement, and cognitive load rather than accuracy alone (Kuai et al., 30 Mar 2026). DispatchMAS yields a complementary set of patterns for workflow-constrained settings: start with a domain taxonomy and fact commons, encode phase-based state machines, separate role agents with distinct constraints, apply turn-level chief-complaint classification and retrieval, and combine expert ratings with automated analyses of sentiment, readability, politeness, and operational behavior (Li et al., 24 Oct 2025).

The broader architecture landscape is mapped by MedMASLab, which standardizes 11 heterogeneous MAS architectures across 24 medical modalities, 11 organ systems, and 473 diseases drawn from 11 clinical benchmarks (Qian et al., 10 Mar 2026). Its taxonomy spans debate, discussion, reconcile, hub-and-spoke meta-prompting, dynamic graph methods such as DyLAN, AutoGen-style conversational agents, blackboard EHR systems such as ColaCare, logic-graph designs such as MedLA, mediator-guided systems such as MedOrch, and mixture-of-agents systems such as MoMA (Qian et al., 10 Mar 2026). Medical-specific MDT-like systems integrated in that framework include MDTeamGPT, MDAgents, MedAgents, and ColaCare, all of which are treated as pluggable solvers with common multimodal I/O and common evaluation (Qian et al., 10 Mar 2026).

MedMASLab’s empirical findings are especially important for interpreting MDT-mimic MAS as a research field. It reports that MAS often improve over single agents on some tasks, but that no single MAS dominates across all 11 benchmarks; it identifies a domain-specific “specialization penalty,” meaning that architectures tuned for one sub-domain often degrade markedly when moved to another (Qian et al., 10 Mar 2026). It also shows that rule-based evaluation can catastrophically under-score verbose multi-agent outputs: for example, MDTeamGPT on PubMedQA falls from $X = f_{\text{struct}}(X_{\text{raw}})$4 under VLM-SJ to $X = f_{\text{struct}}(X_{\text{raw}})$5 under Rule-MR, and DyLAN drops from $X = f_{\text{struct}}(X_{\text{raw}})$6 to $X = f_{\text{struct}}(X_{\text{raw}})$7 under Rule-EM (Qian et al., 10 Mar 2026). This has direct implications for the encyclopedic meaning of MDT-mimic MAS: the category is defined as much by orchestration, communication protocol, and evaluation alignment as by nominal role count.

6. Limitations, controversies, and terminological ambiguity

The principal limitations are now consistent across systems. MDTRoom emphasizes hallucinations, overconfidence, misinterpretation of labs and imaging, the long-tail nature of rare diseases, and the possibility that MAS may amplify errors if multiple agents latch onto the same wrong evidence (Kuai et al., 30 Mar 2026). OMGs notes dependence on EHR quality, a fixed evidence bank, offline non-interventional evaluation, and untested generalizability outside Chinese healthcare settings (Zhang et al., 14 Feb 2026). DispatchMAS identifies “question overload” and “premature misclassification” as specific failure patterns, while noting that its evaluation involved only physicians and simulated calls without background noise or extreme caller behaviors (Li et al., 24 Oct 2025). MedMASLab documents architecture fragility under backbone changes, format failures in structured protocols, convergence failures, and large token inflation when interaction management is weak (Qian et al., 10 Mar 2026).

A central controversy concerns whether explicit MDT simulation is always necessary. The safer therapy recommendation study is the clearest negative result: a carefully prompted single GP-like agent with explicit intermediate reasoning performed as well as the multi-agent MDT simulation on the benchmarked multimorbidity tasks (Wu et al., 15 Jul 2025). The same study also reports that “conflict-targeted MDTs are better than global MDTs,” because early global consultation produced oversized teams, overlapping proposals, and increased medication burden (Wu et al., 15 Jul 2025). MedMASLab reaches a related conclusion at scale: increasing agents and rounds does not monotonically improve performance, and fixed or dynamic expert-playing can explode token cost without consistent gain (Qian et al., 10 Mar 2026). A plausible implication is that the value of MDT mimicry is conditional on task structure, heterogeneity of evidence, and the need for provenance, steerability, or explicit conflict representation rather than on multi-agent multiplicity alone.

The term is also acronymically ambiguous outside clinical LLM systems. In high-energy physics, the phrase is used in an adjacent sense for an “MDT-mimic” muon algorithm/selection built around precision Muon Drift Tube chambers in the ATLAS first-level trigger, where MAS refers to a muon algorithm/selection rather than a medical multi-agent system (Kroha et al., 2016). In cellular networking, “MDT-driven” denotes Minimization of Drive Tests data used to construct a centralized but MAS-like deep-reinforcement-learning environment for joint antenna-tilt optimization (Skocaj et al., 2022). In medical imaging, MASD is a model-agnostic 1-class saliency detector for weakly supervised lesion detection from breast DCE-MRI (Maicas et al., 2018). In causal inference for multidrug-resistant tuberculosis, an “MDT-mimic MAS” framing is applied to marginal structural models and targeted maximum likelihood estimation for multidrug treatment effect modification in an individual-patient-data meta-analysis (Liu et al., 2021). These uses do not describe the same object, but they show that the paired abbreviations “MDT” and “MAS” are domain-dependent and cannot be interpreted reliably without context.

Across the clinical literature, the most stable meaning remains the multidisciplinary-team-mimicking multi-agent system: a deliberative, role-specialized, evidence-linked, and increasingly benchmarked class of systems whose central technical questions are no longer only whether multiple agents can be made to talk, but whether their interaction can be structured, measured, audited, and productively coupled to human expertise (Kuai et al., 30 Mar 2026, Qian et al., 10 Mar 2026).