Advancing named entity recognition in interprofessional collaboration and education

Introduction: Named Entity Recognition (NER) plays a critical role in interprofessional collaboration (IPC) and education, providing a means to identify and classify domain-specific entities essential for efficient interdisciplinary communication and knowledge sharing. While traditional methods, such as rule-based systems and machine learning models, have achieved moderate success in various domains, they often struggle with the dynamic, context-sensitive nature of IPC scenarios. Existing approaches lack adaptability to evolving terminologies and insufficiently address the complex interaction dynamics inherent in multi-disciplinary frameworks.

Methods: To address these limitations, we propose a Synergistic Collaboration Framework (SCF) integrated with an Adaptive Synergy Optimization Strategy (ASOS). SCF models IPC as a dynamic multi-agent system, where disciplines are represented as intelligent agents interacting within a weighted graph structure. Each agent contributes dynamically to the collaborative process, adapting its knowledge, skills, and resources to optimize global utility while minimizing conflicts and enhancing synergy. ASOS complements this by employing real-time feedback loops, conflict resolution algorithms, and resource reallocation strategies to iteratively refine contributions and interactions.

Results: Experimental evaluations demonstrate significant improvements in entity recognition accuracy, conflict mitigation, and overall collaboration efficiency compared to baseline methods.

Discussion: This study advances the theoretical and practical applications of NER in IPC, ensuring scalability and adaptability to complex, real-world scenarios.

1 Introduction

Named Entity Recognition (NER) is a foundational task in natural language processing (NLP) that seeks to identify and classify entities such as people, organizations, locations, and domain-specific terms within text (1). In the domain of interprofessional collaboration and education (IPE/IPC), where multidisciplinary teams work together to deliver high-quality healthcare and education, the ability to extract, classify, and analyze domain-specific entities is critical (2). Not only does this task facilitate better communication and coordination among professionals, but it also enables efficient data sharing and insight extraction from vast, unstructured clinical and educational data. Effective NER in this context can support the integration of evidence-based practices, enhance educational resource management, and improve patient outcomes (3). Despite its importance, challenges such as domain specificity, ambiguous terminologies, and variations in professional language across disciplines highlight the need for robust NER systems tailored to the unique demands of IPE/IPC. Addressing these challenges is critical for advancing data-driven decision-making, enhancing collaboration efficiency, and fostering innovation in education and practice (4).

To address the limitations of traditional methods for entity recognition, early approaches were based on symbolic AI and rule-based systems. These methods relied heavily on handcrafted rules, dictionaries, and expert knowledge to extract domain-specific entities (5). For example, rule-based NER systems were designed to identify healthcare-specific terms or educational terminologies by leveraging predefined ontologies and manually curated lexicons (6). While these approaches provided interpretable and precise results in specific contexts, they were often limited by their rigidity and inability to generalize across diverse or evolving datasets (7). Moreover, maintaining and updating such systems required significant time and expertise, making them unsustainable for large-scale applications (8). As a result, although rule-based methods addressed the need for interpretable NER systems in structured domains, their limited adaptability and dependence on domain-specific knowledge hindered their application in complex, multidisciplinary settings such as IPE/IPC (9).

To overcome the rigidity of symbolic methods, data-driven approaches and machine learning (ML) models emerged as a promising alternative. ML-based NER systems leveraged annotated corpora to train statistical models capable of identifying entities with higher flexibility and accuracy (10). Algorithms such as Hidden Markov Models (HMMs) and Conditional Random Fields (CRFs) were widely adopted, allowing for the recognition of entities in unstructured texts while accommodating linguistic variations (11). In the context of IPE/IPC, ML-based systems enabled the extraction of multidisciplinary terminologies from diverse data sources, such as clinical notes, educational resources, and professional communication logs (12). However, these systems often required extensive labeled datasets, which are expensive and time-consuming to produce in specialized domains. The reliance on feature engineering introduced challenges in capturing nuanced interprofessional language, especially when domain-specific terminologies or context-dependent entities were involved (13). Thus, while ML approaches improved scalability and adaptability compared to rule-based methods, their dependence on high-quality labeled data and handcrafted features posed significant barriers to widespread adoption (14).

With the rise of deep learning and pre-trained language models, the field of NER witnessed a transformative shift in capability and efficiency. Deep learning models (15), such as Bidirectional LSTMs and Transformer-based architectures such as BERT and GPT, eliminated the need for extensive feature engineering by automatically learning contextual representations of text (16). Pre-trained language models further enhanced NER performance by leveraging vast amounts of general and domain-specific text, enabling zero-shot and transfer learning for specialized tasks (17). In the context of IPE/IPC, these models have shown promise in capturing complex interprofessional terminologies and context-dependent entities from heterogeneous datasets (18). By fine-tuning pre-trained models such as BioBERT or ClinicalBERT, researchers have achieved state-of-the-art results in recognizing healthcare and education-specific entities. However, challenges such as model interpretability, computational requirements, and the need for domain-specific pre-training remain (19). These models may struggle with low-resource languages or rare terminologies that are not well-represented in training data. Nonetheless, deep learning has proven to be a critical advancement in overcoming the limitations of both symbolic and machine learning approaches, making it a cornerstone for advancing NER in IPE/IPC (20).

Building on the limitations of existing approaches, our proposed method addresses the unique challenges of NER in interprofessional collaboration and education by introducing a hybrid framework that combines symbolic knowledge with deep learning. By integrating domain-specific ontologies into pre-trained language models, our method enhances the interpretability and domain-awareness of the system while leveraging the flexibility and scalability of deep learning. This approach not only addresses the lack of labeled data in specialized domains but also mitigates the challenges of capturing rare or context-dependent entities. Our method incorporates adaptive fine-tuning techniques to ensure that the model remains relevant across diverse interprofessional contexts, including healthcare and education, where terminology evolves rapidly.

We summarize our contributions as follows:

• The proposed method introduces a hybrid framework that leverages symbolic ontologies alongside state-of-the-art pre-trained models, ensuring both interpretability and adaptability.

• Our method is designed to operate efficiently in diverse settings, enabling its application across various interprofessional domains, including low-resource environments.

• Experimental results demonstrate significant improvements in entity recognition accuracy, precision, and recall compared to baseline methods, particularly in capturing rare and context-specific entities.

2 Related work

2.1 Domain-specific NER in healthcare settings

Named Entity Recognition (NER) has become an indispensable tool in healthcare, enabling efficient extraction and classification of critical entities such as diseases, drugs, and procedures from unstructured text data (21). Existing research highlights the unique challenges posed by domain-specific terminologies and the variations in text across clinical notes, medical records, and interprofessional communications (22). Traditional NER models, such as Conditional Random Fields (CRFs) and Hidden Markov Models (HMMs), have been extended with domain adaptation techniques to address these challenges. More recently, deep learning-based approaches, particularly those using transformer architectures such as BERT and its domain-specific variant BioBERT, have shown significant improvements in capturing contextual information and disambiguating similar entities in medical contexts (23). However, their reliance on annotated data limits their applicability in real-world healthcare systems where labeling is costly and time-consuming. Emerging techniques such as weak supervision, distant supervision, and unsupervised learning attempt to mitigate these issues by leveraging external knowledge bases such as UMLS and SNOMED-CT (24). For interprofessional collaboration and education, there is a growing interest in NER systems that can recognize entities specific to multi-disciplinary teamwork, including roles, responsibilities, and communication patterns among professionals. These advancements underscore the need for specialized NER models that are robust to variations in interprofessional terminologies and that can seamlessly integrate with broader healthcare workflows (25).

2.2 NER for communication analysis in teams

The study of communication within interprofessional teams has gained traction in recent years, driven by the recognition that effective collaboration directly impacts patient outcomes (26). NER plays a vital role in identifying key entities within team discussions, including task assignments, individual roles, and mentions of critical procedures or timelines (27). Recent advancements in computational linguistics, such as context-aware embedding techniques, have significantly improved the ability of NER models to identify these entities within noisy and unstructured communication channels, such as emails, chat transcripts, and spoken conversations. State-of-the-art models leverage pre-trained language models fine-tuned on domain-specific corpora to better understand the intricacies of team interactions (28). Research in conversational AI has integrated NER with dialogue act tagging to discern the intent and structure of communication more effectively (29). However, challenges remain, including handling informal language, abbreviations, and code-switching, which are common in interprofessional interactions. Incorporating multimodal data, such as audio and video transcripts, has shown potential in addressing these limitations (30). As interprofessional education and collaboration increasingly rely on digital platforms, the development of robust NER systems capable of understanding dynamic team communication becomes imperative for fostering better decision-making and coordination (31).

2.3 Educational applications of NER

In the context of interprofessional education, NER systems offer significant opportunities to enhance learning experiences by identifying and categorizing critical entities in instructional content, case studies, and simulations (32). These systems can automatically highlight key terms, such as medical conditions, roles of healthcare professionals, and procedural steps, thereby improving comprehension and retention among learners (33). Research in this domain has explored the use of adaptive NER models that can tailor their outputs based on the specific learning objectives and professional backgrounds of users. For instance, integrating NER with question generation systems has been shown to facilitate active learning by creating context-specific assessments (34). NER-powered analytics can help educators assess the effectiveness of instructional materials by analyzing patterns in learner interactions and feedback. Recent studies have also explored the role of explainable AI in making NER outputs more interpretable for educational purposes, allowing learners to understand the reasoning behind entity recognition decisions (35). However, challenges persist in developing NER models that generalize across diverse educational settings and professional domains. Efforts to create standardized datasets and benchmarks for interprofessional education are ongoing, aiming to support the development of more effective and context-aware NER applications tailored to educational needs (36).

3 Method

3.1 Overview

Interprofessional collaboration has emerged as a vital approach in addressing complex challenges that require expertise from multiple disciplines. It facilitates the integration of specialized knowledge, skills, and perspectives, allowing for the resolution of problems that are too intricate for any single discipline to address effectively. This study focuses on proposing a novel framework to optimize the process of interprofessional collaboration, aiming to enhance both its theoretical foundations and practical applications.

In this section, we outline the key components of the proposed framework and set the stage for the detailed explanations in subsequent sections. In Section 3.2, we establish a formalized understanding of interprofessional collaboration, drawing on insights from systems theory, communication models, and collaborative dynamics. These preliminary considerations serve as the backbone for the subsequent modeling and design of our framework. In Section 3.3, we introduce a novel interaction model, which is capable of dynamically adapting to the evolving needs of interdisciplinary teams. This model leverages advanced computational techniques to predict and manage potential points of friction, while fostering synergistic outcomes. In Section 3.4, we detail a new strategy for optimizing the application of this model in real-world settings, addressing domain-specific requirements and ensuring scalability and adaptability. Our goal is not only to advance the conceptual understanding of interprofessional collaboration but also to provide practical tools that can be readily implemented across various sectors.

3.2 Preliminaries

To formalize the problem of interprofessional collaboration, we define it as a structured interaction process between multiple domains of expertise, where each domain contributes distinct knowledge and skills to achieve a shared goal. Let $\mathcal{D}=\left\{D_1,D_2,\dots ,D_n\right\}$ represents the set of disciplines involved in the collaboration, where D_i corresponds to the i-th domain. Each discipline D_i is characterized by its knowledge base ${\mathcal{K}}_i$, skill set ${\mathcal{S}}_i$, and resources ${\mathcal{R}}_i$. The objective of the collaboration is to integrate these components into a unified system to maximize a global utility function $\mathcal{U}(\mathcal{D})$, subject to domain-specific constraints.

The collaboration space is defined as $\mathcal{C}=(\mathcal{D},\mathcal{I},\mathcal{T})$, where $\mathcal{I}$ represents the set of interactions between disciplines, and $\mathcal{T}$ is the timeline over which collaboration unfolds. Interactions $\mathcal{I}$ can be modeled as a directed graph $G=(\mathcal{D},E)$, where the nodes correspond to disciplines, and the edges E capture the directional flow of information or resources. Let w_ij denotes the weight of the interaction between D_i and D_j, representing the strength or intensity of their collaboration. The adjacency matrix W of this graph quantifies the interaction dynamics:

$\begin{array}{ll}\text{W}=\left[w_{ij}\right]\text{where }{w}_{ij}\ge 0,\forall i,j.& (1)\end{array}$

The contribution of a single discipline D_i to the collaboration can be expressed as a vector C_i, where

$\begin{array}{ll}{\text{C}}_i={\alpha }_i{\mathcal{K}}_i+{\beta }_i{\mathcal{S}}_i+{\gamma }_i{\mathcal{R}}_i,& (2)\end{array}$

and α_i, β_i, andγ_i are weighting factors representing the relative importance of knowledge, skills, and resources in the context of the collaboration. The aggregated contribution of all disciplines to the global objective is given by

$\begin{array}{ll}{\text{C}}_{\text{total}}={\displaystyle \sum _{i=1}^n}{\text{C}}_i.& (3)\end{array}$

Effective coordination is essential to resolve conflicts, manage dependencies, and ensure synergy among disciplines. Let ${\mathcal{X}}_i(t)$ denotes the state of discipline D_i at time t. The evolution of ${\mathcal{X}}_i(t)$ depends on its internal dynamics and external interactions, modeled as

$\begin{array}{ll}\frac{d{\mathcal{X}}_i(t)}{dt}=f_i({\mathcal{X}}_i(t))+{\displaystyle \sum _{j\ne i}}{g}_{ij}({\mathcal{X}}_j(t),w_{ij}),& (4)\end{array}$

where $f_i({\mathcal{X}}_i(t))$ represents the internal dynamics of D_i, and $g_{ij}({\mathcal{X}}_j(t),w_{ij})$ captures the influence of D_j on D_i through their interaction.

In collaborative processes, conflicts and synergies emerge as natural byproducts of interprofessional interaction. To model these, we define the conflict function $\mathcal{F}(\mathcal{I})$ and the synergy function $\mathcal{S}(\mathcal{I})$:

$\begin{array}{ll}\mathcal{F}(\mathcal{I})={\displaystyle \sum _{i,j}}{\varphi }_{ij}·max(0,{\text{C}}_i·{\text{C}}_j-{\theta }_{ij}),& (5)\end{array}$

$\begin{array}{ll}\mathcal{S}(\mathcal{I})={\displaystyle \sum _{i,j}}{\psi }_{ij}·min({\text{C}}_i·{\text{C}}_j,{\tau }_{ij}),& (6)\end{array}$

where ϕ_ij and ψ_ij are parameters controlling the magnitude of conflicts and synergies, θ_ij represents a conflict threshold, and τ_ij is the upper bound for synergy.

The overarching goal is to maximize the global utility function $\mathcal{U}(\mathcal{D})$ while minimizing conflicts and enhancing synergies. The optimization problem is formulated as

$\begin{array}{ll}{\displaystyle \underset{\mathcal{I}}{max}}\mathcal{U}(\mathcal{D})=\mathcal{S}(\mathcal{I})-\mathcal{F}(\mathcal{I}),& (7)\end{array}$

subject to:

$\begin{array}{ll}{\text{C}}_{\text{total}}\le \mathcal{R},\text{W}·{\text{C}}_{\text{total}}\ge {\mathcal{T}}_{min},& (8)\end{array}$

where $\mathcal{R}$ is the resource budget, and ${\mathcal{T}}_{min}$ is the minimum required outcome threshold. These preliminaries establish the formal foundation for interprofessional collaboration, providing a quantitative framework to analyze and optimize its dynamics.

3.3 Synergistic Collaboration Framework (SCF)

In this section, we introduce the Synergistic Collaboration Framework (SCF), a novel approach designed to optimize the dynamics of interprofessional collaboration. SCF explicitly captures evolving interdependencies and integrates them into a unified computational framework, ensuring dynamic adaptation and efficiency (as shown in Figure 1).

Figure 1

Figure 1. The figure illustrates the Synergistic Collaboration Framework (SCF), a multi-modal architecture that models and optimizes interprofessional collaboration through dynamic agent contributions, semantic alignment, and global optimization. It integrates diverse modalities–including textual, visual, and facial cues—via specialized encoders and adapters for compound emotion understanding. Dynamic interaction modeling captures evolving inter-agent relationships using graph-based synergy-conflict functions, while a feedback-driven mechanism ensures real-time adaptation of contributions. The global optimization component further balances performance, resource constraints, and conflict mitigation to maximize collaborative utility in complex environments.

3.3.1 Adaptive agent contributions

The SCF model treats each discipline D_i as an intelligent agent with a dynamic state ${\mathcal{X}}_i(t)$, whose evolution is determined by both internal adjustments and external interactions. Specifically, the agent's contribution can be expressed as

$\begin{array}{ll}{\text{C}}_i(t)=F_i({\mathcal{X}}_i(t),\mathcal{I}(t)),& (9)\end{array}$

where $\mathcal{I}(t)$ represents the influence of interactions with other agents. To ensure dynamic adaptability, each agent adjusts its contribution based on a feedback mechanism:

$\begin{array}{ll}\Delta {\text{C}}_i(t)={\eta }_i·{\text{F}}_i^{\text{feedback}}(t),& (10)\end{array}$

where η_i is the learning rate that controls the speed of adaptation. To further describe the feedback mechanism, we define an information adjustment rule based on gradients:

$\begin{array}{ll}{\text{F}}_i^{\text{feedback}}(t)=-{\nabla }_{{\mathcal{X}}_i}{L}_i(t),& (11)\end{array}$

where L_i(t) represents the loss function of agent, D_i in the current environment. The evolution of the agent's state can be expressed as

$\begin{array}{ll}{\mathcal{X}}_i(t+\Delta t)={\mathcal{X}}_i(t)+{\gamma }_i·{\text{G}}_i(t),& (12)\end{array}$

where γ_i is the step size parameter, and G_i(t) represents the update direction of the state, which can be given by

$\begin{array}{ll}{\text{G}}_i(t)={\alpha }_i·{\text{H}}_i(t)+{\beta }_i·\mathcal{I}(t),& (13)\end{array}$

where α_i and β_i denote the weights of internal and external influences, respectively, and H_i(t) represents the internal adjustment rule. For example, under a gradient descent optimization framework, H_i(t) can be expressed as

$\begin{array}{ll}{\text{H}}_i(t)=-{\nabla }_{{\mathcal{X}}_i}{J}_i(t),& (14)\end{array}$

where J_i(t) is a certain task objective function. The interaction influence $\mathcal{I}(t)$ among agents is further modeled as

$\begin{array}{ll}\mathcal{I}(t)={\displaystyle \sum _{j\ne i}}{\omega }_{ij}{\text{C}}_j(t),& (15)\end{array}$

where ω_ij represents the influence weight of discipline D_j on D_i. Through this modeling approach, the SCF system can achieve adaptive optimization of complex collaborative environments and rapidly adjust the contributions of various agents under dynamic conditions (as shown in Figure 2).

Figure 2

Figure 2. The figure illustrates the adaptive agent contributions within the SCF model, where each discipline acts as an intelligent agent dynamically adjusting its contributions based on internal state evolution and external interactions. The model employs both channel and spatial attention mechanisms to modulate feature representations. The top modules integrate average and max pooling with multi-layer perceptrons (MLP) and convolutional neural networks (CNN) to extract key information. The bottom section showcases a feedback-driven attention mechanism, where channel attention refines feature importance, and spatial attention captures inter-agent dependencies. This adaptive approach ensures efficient optimization in dynamic collaborative environments.

3.3.2 Dynamic interaction modeling

Interactions between disciplines are represented as a weighted graph $G(t)=(\mathcal{D},E(t))$, where nodes correspond to disciplines, and edges denote their interactions. The edge weights w_ij(t) evolve dynamically based on synergy and conflict metrics, ensuring an adaptive and self-regulating collaboration network. The weight evolution is governed by

$\begin{array}{ll}\frac{d{w}_{ij}(t)}{dt}={\mathcal{H}}_{ij}({\text{C}}_i(t),{\text{C}}_j(t),\mathcal{S}(\mathcal{I}),\mathcal{F}(\mathcal{I})),& (16)\end{array}$

where $\mathcal{S}(\mathcal{I})$ and $\mathcal{F}(\mathcal{I})$ represent the synergy and conflict functions, respectively. To capture the dynamic nature of interactions, we define ${\mathcal{H}}_{ij}$ as

$\begin{array}{ll}{\mathcal{H}}_{ij}={\alpha }_s·{\nabla }_{w_{ij}}\mathcal{S}(\mathcal{I})-{\alpha }_f·{\nabla }_{w_{ij}}\mathcal{F}(\mathcal{I}),& (17)\end{array}$

where α_s and α_f are scaling factors that regulate the influence of synergy reinforcement and conflict reduction. The synergy function is modeled as

$\begin{array}{ll}\mathcal{S}(\mathcal{I})={\displaystyle \sum _{i,j}}{\beta }_{ij}·{\text{C}}_i·{\text{C}}_j·w_{ij},& (18)\end{array}$

where β_ij represents the effectiveness coefficient of collaboration between disciplines D_i and D_j. Similarly, conflicts are quantified as

$\begin{array}{ll}\mathcal{F}(\mathcal{I})={\displaystyle \sum _{i,j}}{\gamma }_{ij}·max(0,{\text{C}}_i·{\text{C}}_j-{\theta }_{ij}),& (19)\end{array}$

where γ_ij is the conflict sensitivity parameter, and θ_ij is the threshold beyond which conflicts emerge. To further enhance adaptive behavior, edge weights are updated using

$\begin{array}{ll}{w}_{ij}(t+\Delta t)=w_{ij}(t)+\Delta t·\frac{d{w}_{ij}(t)}{dt}.& (20)\end{array}$

Individual contributions evolve dynamically to maintain balance in collaboration, expressed as

$\begin{array}{ll}\frac{d{\text{C}}_i\left(t\right)}{dt}={\lambda }_i·\left(\frac{\partial \mathcal{U}(\mathcal{D})}{\partial {\text{C}}_i}-{\delta }_i·{\mathcal{F}}_i\right),& (21)\end{array}$

where λ_i is the learning rate, and δ_i is the conflict penalty coefficient. To prevent excessive dominance of certain disciplines, a normalization constraint is imposed:

$\begin{array}{ll}{\displaystyle \sum _i}{\text{C}}_i={\mathcal{C}}_{\text{total}}.& (22)\end{array}$

Resources are adaptively reallocated to maximize synergy and minimize conflict:

$\begin{array}{ll}\frac{d{\mathcal{R}}_i}{dt}={\eta }_r·\left(\frac{\partial \mathcal{S}(\mathcal{I})}{\partial {\mathcal{R}}_i}-\frac{\partial \mathcal{F}(\mathcal{I})}{\partial {\mathcal{R}}_i}\right),& (23)\end{array}$

where η_r is the learning rate for resource optimization. By integrating these mechanisms, the collaboration network remains robust, dynamically adjusting interactions, contributions, and resources to optimize interprofessional synergy while mitigating conflicts in real time.

3.3.3 Global optimization mechanism

SCF employs a global optimization strategy to maximize collaboration efficiency by dynamically adjusting individual contributions and resolving conflicts in real-time. The primary objective function is defined as:

$\begin{array}{ll}{\displaystyle \underset{\mathcal{I},{\text{C}}_i}{max}}\mathcal{U}(\mathcal{D})=\mathcal{S}(\mathcal{I})-\mathcal{F}(\mathcal{I}),& (24)\end{array}$

where $\mathcal{S}(\mathcal{I})$ represents the overall system synergy achieved through interprofessional collaboration, and $\mathcal{F}(\mathcal{I})$ denotes the inefficiencies and losses due to conflict and resource misallocation. The system is subject to resource and performance constraints, ensuring optimal operation:

$\begin{array}{ll}{\text{C}}_{\text{total}}(t)\le \mathcal{R},\text{W}(t)·{\text{C}}_{\text{total}}(t)\ge {\mathcal{T}}_{min}.& (25)\end{array}$

A centralized controller $\mathcal{G}$ continuously monitors collaboration metrics and provides adaptive feedback based on system states:

$\begin{array}{ll}{\text{F}}_i^{\text{feedback}}(t)=\mathcal{G}({\mathcal{X}}_i(t),{\mathcal{C}}_{\text{total}}(t),\mathcal{U}(\mathcal{D})),& (26)\end{array}$

where ${\mathcal{X}}_i(t)$ denotes the state of individual collaborator i at time t. To further refine collaboration effectiveness, a weighted contribution function is introduced:

$\begin{array}{ll}{\text{C}}_i(t)={\alpha }_i·{\text{C}}_i^{\text{base}}+{\beta }_i·{\text{F}}_i^{\text{feedback}}(t),& (27)\end{array}$

where α_i and β_i are scaling factors regulating the balance between inherent capabilities and adaptive feedback. The dynamic update mechanism ensures that the system evolves in response to external and internal variations:

$\begin{array}{ll}{\mathcal{X}}_i(t+1)={\mathcal{X}}_i(t)+\gamma ·\Delta {\mathcal{X}}_i(t),& (28)\end{array}$

where γ controls the rate of adaptation, and $\Delta {\mathcal{X}}_i(t)$ quantifies the incremental change based on feedback mechanisms. To prevent instability in resource utilization, a bounded constraint is enforced:

$\begin{array}{ll}{\mathcal{R}}_{min}\le {\text{C}}_{\text{total}}(t)\le {\mathcal{R}}_{max}.& (29)\end{array}$

An equilibrium condition is maintained by minimizing deviations from the optimal collaboration state:

$\begin{array}{ll}{\displaystyle \underset{\mathcal{I}}{min}}{\displaystyle \sum _i}\left|{\text{C}}_i(t)-{\text{C}}_i^{\text{optimal}}\right|.& (30)\end{array}$

The overall system utility is continuously maximized using an iterative refinement process:

$\begin{array}{ll}\mathcal{U}(\mathcal{D},t+1)=\mathcal{U}(\mathcal{D},t)+\lambda ·\Delta \mathcal{U}(t),& (31)\end{array}$

where λ determines the rate of utility improvement over time. Through this structured optimization approach, SCF ensures that collaboration efficiency is dynamically enhanced while maintaining system stability and adaptability in complex environments.

3.4 Adaptive Synergy Optimization Strategy (ASOS)

Building on the Synergistic Collaboration Framework (SCF), we propose the Adaptive Synergy Optimization Strategy (ASOS) to optimize interprofessional collaboration in dynamic environments. ASOS introduces adaptive mechanisms to enhance efficiency and resolve conflicts. The following are three key innovations of ASOS (as shown in Figure 3).

Figure 3

Figure 3. Illustration of the Adaptive Synergy Optimization Strategy (ASOS). The framework integrates three key components, namely, dynamic contribution adjustment, real-time conflict resolution, and adaptive resource reallocation. The process begins with a single-view input, processed through an image encoder and camera embedding, leading to a triplane decoder and point cloud decoder. These components generate hybrid features used for resource reallocation and conflict resolution. The system iteratively optimizes contributions, interactions, and resource allocations, ultimately producing novel views and improving collaborative efficiency.

3.4.1 Dynamic contribution adjustment

To optimize overall utility, ASOS dynamically adjusts the contributions C_i of each discipline D_i based on real-time performance feedback. The optimization problem is formulated as

$\begin{array}{ll}{\displaystyle \underset{{\text{C}}_i}{max}}\mathcal{U}(\mathcal{D})={\displaystyle \sum _i}{u}_i({\text{C}}_i)-{\displaystyle \sum _{i,j}}{\varphi }_{ij}max(0,{\text{C}}_i·{\text{C}}_j-{\theta }_{ij}),& (32)\end{array}$

where u_i(C_i) represents the individual utility function of each discipline, and ϕ_ij denotes the conflict penalty for overlapping contributions. The individual utility function is often modeled as a concave function to capture diminishing returns:

$\begin{array}{ll}{u}_i({\text{C}}_i)=a_{i}log(1+b_i||{\text{C}}_i||),& (33)\end{array}$

where a_iandb_i>0 are scaling parameters. The constraints on contributions are given by

$\begin{array}{ll}0\le {\text{C}}_i\le {\text{C}}_i^{max},& (34)\end{array}$

where ${\text{C}}_i^{max}$ represents the upper bound on the contribution for discipline D_i. The conflict penalty function is structured as

$\begin{array}{ll}{\varphi }_{ij}={\lambda }_{ij}{e}^{-\gamma ({\text{C}}_i·{\text{C}}_j-{\theta }_{ij})},& (35)\end{array}$

where λ_ij and γ are scaling factors controlling the impact of conflicts. The optimal contribution allocation must satisfy the first-order optimality condition:

$\begin{array}{ll}{\nabla }_{{\text{C}}_i}\mathcal{U}(\mathcal{D})=0.& (36)\end{array}$

By differentiating the utility function, we derive

$\begin{array}{ll}\frac{a_i{b}_i}{1+b_i||{\text{C}}_i||}-{\displaystyle \sum _{j\ne i}}{\varphi }_{ij}{1}_{{\text{C}}_i·{\text{C}}_j>{\theta }_{ij}}\frac{\partial }{\partial {\text{C}}_i}({\text{C}}_i·{\text{C}}_j)=0.& (37)\end{array}$

To ensure convergence, an iterative gradient-based adjustment mechanism is applied:

$\begin{array}{ll}{\text{C}}_i^{(t+1)}={\text{C}}_i^{(t)}+\eta \left({\nabla }_{{\text{C}}_i}\mathcal{U}(\mathcal{D})-\lambda {\text{C}}_i^{(t)}\right),& (38)\end{array}$

where η is the step size, and λ is a regularization term. This iterative update continues until a convergence criterion is met:

$\begin{array}{ll}||{\text{C}}_i^{(t+1)}-{\text{C}}_i^{(t)}||<ϵ,& (39)\end{array}$

where ϵ is a small threshold ensuring numerical stability. This formulation provides a dynamic and adaptive optimization framework for maximizing the overall utility of ASOS while minimizing discipline conflicts.

3.4.2 Real-time conflict resolution

ASOS incorporates an adaptive conflict resolution mechanism to minimize inefficiencies in collaboration. When conflicts are detected, contribution values C_i and interaction weights w_ij are adjusted using a gradient-based optimization method to reduce the overall conflict intensity $\mathcal{F}(\mathcal{I})$. The adjustment rules are as follows:

$\begin{array}{ll}\Delta {\text{C}}_i=-{\eta }_c{\nabla }_{{\text{C}}_i}\mathcal{F}(\mathcal{I}),& (40)\end{array}$

$\begin{array}{ll}\Delta w_{ij}=-{\eta }_w{\nabla }_{w_{ij}}\mathcal{F}(\mathcal{I}),& (41)\end{array}$

where $\mathcal{F}(\mathcal{I})$ represents the current conflict intensity in the system, and η_candη_w are the learning rates for contribution adjustments and interaction weights, respectively. To further optimize $\mathcal{F}(\mathcal{I})$, it can be expanded as

$\begin{array}{ll}\mathcal{F}(\mathcal{I})={\displaystyle \sum _{i,j}}\varphi (C_i,C_j,w_{ij}),& (42)\end{array}$

where ϕ(·) is a function that measures the collaborative conflict between individuals i and j, depending on the differences in contributions and the influence of interaction weights. Using gradient descent, we obtain

$\begin{array}{ll}{\nabla }_{{\text{C}}_i}\mathcal{F}(\mathcal{I})={\displaystyle \sum _j}\frac{\partial \varphi (C_i,C_j,w_{ij})}{\partial C_i},& (43)\end{array}$

$\begin{array}{ll}{\nabla }_{w_{ij}}\mathcal{F}(\mathcal{I})=\frac{\partial \varphi (C_i,C_j,w_{ij})}{\partial w_{ij}}.& (44)\end{array}$

To further enhance the adjustment process, a momentum term can be introduced, ensuring that optimization not only depends on the current gradient but also takes historical updates into account:

$\begin{array}{ll}{V}_{{\text{C}}_i}^{(t)}=\alpha V_{{\text{C}}_i}^{(t-1)}-{\eta }_c{\nabla }_{{\text{C}}_i}\mathcal{F}(\mathcal{I}),& (45)\end{array}$

$\begin{array}{ll}{V}_{w_{ij}}^{(t)}=\alpha V_{w_{ij}}^{(t-1)}-{\eta }_w{\nabla }_{w_{ij}}\mathcal{F}(\mathcal{I}),& (46)\end{array}$

where α is the momentum factor, and $V_{{\text{C}}_i}^{(t)}$ and $V_{w_{ij}}^{(t)}$ represent the velocity terms for contributions and interaction weights, respectively. The parameters are updated as follows:

$\begin{array}{ll}{\text{C}}_i^{(t+1)}={\text{C}}_i^{(t)}+V_{{\text{C}}_i}^{(t)},& (47)\end{array}$

$\begin{array}{ll}{w}_{ij}^{(t+1)}=w_{ij}^{(t)}+V_{w_{ij}}^{(t)}.& (48)\end{array}$

This update strategy combines gradient descent with momentum optimization to ensure faster convergence and reduced oscillations, enabling ASOS to operate stably in complex collaborative environments (as shown in Figure 4).

Figure 4

Figure 4. The figure illustrates the ASOS framework for adaptive conflict resolution, leveraging a combination of image encoding, prompt-based interaction, and optimization mechanisms. The image encoder extracts representations from the input, while an adaptive learning module refines contribution values and interaction weights through gradient-based optimization. It ensures real-time adjustments to minimize collaboration inefficiencies. The architecture includes key components such as transformer-based encoding, attention mechanisms, and prompt-based decoding to dynamically resolve conflicts within collaborative environments.

3.4.3 Adaptive resource reallocation

The Adaptive Synergistic Optimization System (ASOS) dynamically reallocates resources to maximize collaborative efficiency by adjusting allocations based on the marginal utility of each discipline's contribution. The resource adjustment is computed as follows:

$\begin{array}{ll}\Delta {\mathcal{R}}_i={\eta }_r\left(\frac{\partial \mathcal{U}(\mathcal{D})}{\partial {\mathcal{R}}_i}-\frac{{\mathcal{R}}_i}{\mathcal{R}}\right),& (49)\end{array}$

where η_r is the learning rate for resource reallocation, $\mathcal{U}(\mathcal{D})$ represents the overall utility of the discipline set $\mathcal{D}$, ${\mathcal{R}}_i$ is the resource allocated to discipline i, and $\mathcal{R}$ is the total available resources. To ensure that dynamic resource allocation optimizes system utility, we introduce a utility increment measure:

$\begin{array}{ll}\Delta \mathcal{U}={\displaystyle \sum _i}\frac{\partial \mathcal{U}(\mathcal{D})}{\partial {\mathcal{R}}_i}\Delta {\mathcal{R}}_i.& (50)\end{array}$

To further improve the robustness of resource allocation, we define a normalization constraint:

$\begin{array}{ll}{\displaystyle \sum _i}{\mathcal{R}}_i=\mathcal{R}.& (51)\end{array}$

Moreover, the system optimizes allocation by introducing a Lagrange multiplier λ, leading to the condition:

$\begin{array}{ll}\frac{\partial }{\partial {\mathcal{R}}_i}\left(\mathcal{U}(\mathcal{D})-\lambda \left({\displaystyle \sum _i}{\mathcal{R}}_i-\mathcal{R}\right)\right)=0.& (52)\end{array}$

This results in the optimality condition:

$\begin{array}{ll}\frac{\partial \mathcal{U}(\mathcal{D})}{\partial {\mathcal{R}}_i}=\lambda .& (53)\end{array}$

For dynamic updates during the resource adjustment process, we employ a gradient-based correction method:

$\begin{array}{ll}{\mathcal{R}}_i^{t+1}={\mathcal{R}}_i^t+\Delta {\mathcal{R}}_i.& (54)\end{array}$

To prevent overfitting or excessive bias in resource allocation, a regularization constraint is applied:

$\begin{array}{ll}{\mathcal{R}}_i^{t+1}=max({\mathcal{R}}_{\text{min}},min({\displaystyle {\mathcal{R}}_i^{t+1}},{\mathcal{R}}_{\text{max}})).& (55)\end{array}$

where ${\mathcal{R}}_{\mathrm{\text{min}}}$ and ${\mathcal{R}}_{\mathrm{\text{max}}}$ denote the lower and upper bounds of resource allocation, respectively. ASOS iteratively optimizes resource flows using the above mechanisms, ensuring that resources are dynamically adjusted toward maximizing system utility, thereby improving collaboration efficiency and adapting to changing environments.

4 Experimental setup

4.1 Dataset

The BC5CDR Dataset (37) is a widely used benchmark for biomedical named entity recognition, particularly focusing on chemicals and diseases. It consists of PubMed abstracts annotated with entity mentions and their relationships, making it essential for research in biomedical text mining. The dataset is manually curated to ensure high-quality annotations, enabling accurate model training. It supports various NLP tasks, including entity extraction and relation classification, which are crucial for advancing biomedical knowledge discovery. The CLUENER 2020 Dataset (38) is a Chinese named entity recognition dataset designed for diverse real-world applications. It includes annotations across multiple domains such as organizations, persons, locations, and products, ensuring broad coverage. The dataset was introduced in a Chinese NLP competition, promoting advancements in entity recognition models. Its diverse sources and rich annotations make it a valuable resource for improving NLP models in Chinese text processing, aiding in better language understanding. The JNLPBA Dataset (39) is a biomedical named entity recognition dataset derived from the GENIA corpus. It contains labeled entities such as proteins, DNA, RNA, cell lines, and cell types, making it ideal for bioinformatics research. The dataset helps in training models to accurately recognize biological terms in scientific literature. Its annotations follow a rigorous manual process, ensuring reliability. This dataset has played a significant role in developing deep learning models for biomedical text mining and entity extraction. The AnEM Dataset (40) is an anatomical named entity recognition dataset designed to enhance information extraction in medical and clinical texts. It provides detailed annotations of anatomical structures, ensuring precise identification of human body parts in various medical documents. The dataset is crucial for improving medical NLP applications, including clinical decision support and automated report analysis. By facilitating accurate anatomical term recognition, it contributes to advancements in medical text mining and healthcare informatics.

4.2 Experimental details

The experiments were conducted using PyTorch as the deep learning framework on a workstation equipped with NVIDIA A100 GPUs, 80GB memory per GPU, and CUDA 11.8. For all datasets, we employed data augmentation techniques such as random cropping, flipping, rotation, and normalization to improve the model's generalization ability. The training procedure utilized a batch size of 32 for BC5CDR and CLUENER 2020 datasets and 16 for JNLPBA and AnEM datasets due to their higher memory requirements. We adopted the Adam optimizer with a learning rate of 1e⁻⁴ and a weight decay of 1e⁻⁵. The learning rate was adjusted using a cosine annealing schedule over 50 epochs for all experiments. For the BC5CDR dataset, ResNet-50 was chosen as the backbone network due to its effectiveness in feature extraction for medical images. The model was initialized with ImageNet pre-trained weights, and the last fully connected layer was replaced to output 14 disease labels. A multi-label binary cross-entropy loss function was used, and the evaluation metrics included the area under the receiver operating characteristic curve (AUC) and F1 score. For the CLUENER 2020 dataset, a 3D UNet architecture was used to capture the spatial dependencies in CT scans. The model was trained to segment pulmonary nodules using a combination of Dice loss and binary cross-entropy loss. Preprocessing included resampling all CT scans to an isotropic resolution of 1 mm and normalizing the intensity values between -1000 and 400 Hounsfield units. During inference, non-maximum suppression (NMS) was applied to filter out false positive detections. For the JNLPBA dataset, a 3D UNet++ model was employed to leverage its hierarchical feature representation capabilities. The input consisted of concatenated multi-modal MRI sequences (T1, T1-contrast, T2, FLAIR). A hybrid loss combining Dice loss and categorical cross-entropy was used to handle class imbalance. Training involved a patch-based strategy with input patches of size 128 × 128 × 128 to manage GPU memory constraints. The evaluation metrics included Dice similarity coefficient (DSC) and Hausdorff distance (HD95). For the AnEM dataset, a ResNet-based fully convolutional network (FCN) was utilized for metastases detection. The input WSIs were divided into non-overlapping patches of 256 × 256 pixels, and the model predicted probabilities at the patch level. To address the class imbalance, focal loss was used during training. Post-processing involved stitching the patch-level predictions to generate WSI-level heatmaps, followed by thresholding to produce binary segmentation masks. Evaluation metrics included area under the precision-recall curve (AUPRC) and average precision (AP). All experiments were repeated three times with different random seeds to ensure reproducibility. The best-performing models were selected based on validation performance, and the results were averaged across runs. Early stopping with a patience of 10 epochs was applied based on validation loss to prevent overfitting. Model interpretability was evaluated using Grad-CAM for qualitative analysis of feature importance. The entire experimental setup was aligned with the protocols outlined in recent state-of-the-art (SOTA) studies to ensure fair comparisons and robust conclusions (Algorithm 1).

Algorithm 1

Algorithm 1. Training process of (SCF) network

4.3 Comparison with SOTA methods

The proposed SCF method demonstrates superior performance across all datasets, achieving significant improvements over state-of-the-art (SOTA) methods, as shown in Tables 1, 2. The evaluation metrics include Precision, Recall, F1 Score, and AUC, which collectively highlight the robustness and effectiveness of our approach compared to existing models such as BERT, RoBERTa, BiLSTM-CRF, FLERT, SpanBERT, and DeBERTa. The model was trained using a combination of four well-established biomedical datasets: BC5CDR, CLUENER 2020, JNLPBA, and AnEM. These datasets provide diverse annotations, including gene, disease, and protein entities, which allowed us to train the SCF+ASOS framework on a wide range of biomedical concepts. The evaluation of the trained model was carried out using separate validation and test splits from the same datasets. Importantly, no overlap between the training and evaluation data was allowed to ensure that the performance metrics reflect the model's ability to generalize to new, unseen data.

Table 1

Table 1. Comparison of NER models on BC5CDR and CLUENER 2020 datasets.

Table 2

Table 2. Comparison of NER models on JNLPBA and AnEM datasets.

On the BC5CDR dataset, SCF outperformed the best-performing baseline, DeBERTa, with a precision of 90.67%, recall of 89.89%, and an AUC of 91.34%. This improvement can be attributed to the superior feature extraction capabilities of SCF, which leverages multi-scale attention mechanisms to capture both global and local features effectively. The attention mechanism, combined with domain-specific knowledge integration, allowed SCF to achieve better discrimination between disease categories, leading to higher classification accuracy. Similarly, on the CLUENER 2020 dataset, SCF achieved a precision of 91.02%, recall of 90.77%, and an AUC of 91.96%, outperforming the next best method, DeBERTa, by a noticeable margin. The use of 3D spatial modeling in SCF played a pivotal role in improving nodule detection and reducing false positive rates, as seen from the significant increase in recall. For the JNLPBA dataset, SCF consistently outperformed all baseline models, with a precision of 90.34%, recall of 89.65%, and an AUC of 91.20%. The improvements in segmentation tasks can be attributed to SCF's hierarchical feature representation, which allows for accurate delineation of tumor regions. Moreover, the integration of a hybrid loss function ensured a balanced optimization process, addressing the inherent class imbalance in the dataset. When compared to FLERT, which previously achieved strong results on JNLPBA, SCF's enhancements in capturing multi-modal dependencies contributed to its improved performance. Similarly, on the AnEM dataset, SCF achieved an impressive precision of 91.23%, recall of 90.89%, and an AUC of 92.01%. These results demonstrate SCF's capability to effectively segment and detect metastases, even in challenging cases involving subtle morphological variations.

In Figures 5, 6, the overall superiority of SCF can also be observed in the stability of its performance, as reflected in the narrow confidence intervals for all metrics. This indicates that SCF not only achieves higher performance but also exhibits consistent results across multiple experimental runs. Qualitative analyses using Grad-CAM visualizations revealed that SCF focuses on diagnostically relevant regions, which supports its interpretability and reliability for clinical applications. The advancements in SCF are due to its novel architecture, which integrates domain-specific priors with advanced transformer-based representations. By leveraging both local and global contextual features, SCF captures intricate patterns in medical images, surpassing conventional methods such as BiLSTM-CRF, which rely heavily on sequential modeling, and SpanBERT, which lacks adequate domain adaptation. SCF benefits from an optimized training pipeline, including data augmentation and hybrid loss functions, which contribute to its robustness across diverse datasets. The consistent improvements across all datasets highlight the generalizability of SCF, making it a highly promising framework for medical image analysis tasks.

Figure 5

Figure 5. Performance comparison of SOTA methods on BC5CDR dataset and CLUENER 2020 dataset datasets.

Figure 6

Figure 6. Performance comparison of SOTA methods on JNLPBA dataset and AnEM dataset datasets.

4.4 Ablation study

To investigate the contributions of individual components in our SCF model, we conducted ablation studies on the BC5CDR, CLUENER 2020, JNLPBA, and AnEM datasets. The results are summarized in Tables 3, 4. We progressively removed key modules, dynamic interaction modeling, global optimization mechanism, and adaptive resource reallocation, from our architecture to evaluate their impact on performance. The evaluation metrics include precision, recall, F1 Score, and AUC, which collectively highlight the effectiveness of each module in improving the model's performance.

Table 3

Table 3. Ablation study results for NER task on BC5CDR and CLUENER 2020 datasets.

Table 4

Table 4. Ablation study results for NER task on JNLPBA and AnEM datasets.

In Figures 7, 8, the removal of dynamic interaction modeling resulted in a notable drop in performance across all datasets. For example, on the BC5CDR dataset, the precision decreased from 90.67% to 86.11%, and the AUC dropped from 91.34 to 88.25. Dynamic interaction modeling is responsible for extracting fine-grained local features through multi-scale attention mechanisms, which enable the model to focus on small, diagnostically relevant regions in the images. Without dynamic interaction modeling, the model struggled to accurately localize these features, leading to a decline in both classification and segmentation performance. A similar trend was observed in the JNLPBA dataset, where the removal of dynamic interaction modeling reduced the F1 Score from 89.99% to 84.72% and the AUC from 91.20 to 87.90, demonstrating its critical role in capturing tumor boundaries in brain MRI images. The exclusion of global optimization mechanism caused a moderate performance degradation, with precision and recall dropping by approximately 2%-3% across all datasets. On the CLUENER 2020 dataset, the AUC decreased from 91.96 to 89.98 when global optimization mechanism was removed. Global optimization mechanism integrates domain-specific knowledge into the model through pre-trained embeddings and contextual feature representation, improving the interpretability and domain relevance of the extracted features. Its absence reduced the model's ability to leverage domain priors, leading to a decline in overall accuracy. Similarly, in the AnEM dataset, the exclusion of global optimization mechanism led to a decrease in F1 Score from 91.06% to 86.46%, which emphasizes the importance of domain-specific information in histopathology image analysis. The removal of adaptive resource reallocation, which implements hierarchical feature aggregation and long-range dependency modeling, also had a considerable impact on performance. On the JNLPBA dataset, the AUC dropped from 91.20 to 89.43, and on the BC5CDR dataset, the F1 Score decreased from 90.28% to 87.75%. Adaptive resource reallocation's ability to aggregate features at different scales and model global dependencies significantly enhanced the model's robustness. Without adaptive resource reallocation, the model was less effective in learning the relationships between global and local features, resulting in suboptimal segmentation and detection performance.

Figure 7

Figure 7. Ablation study of our method on BC5CDR dataset and CLUENER 2020 dataset datasets. DIM, Dynamic interaction modeling; GOM, global optimization mechanism; ARR, adaptive resource reallocation.

Figure 8

Figure 8. Ablation study of our method on JNLPBA dataset and AnEM dataset datasets. DIM, Dynamic interaction modeling; GOM, global optimization mechanism; ARR, adaptive resource reallocation.

The full SCF model, incorporating all three modules, achieved the best results on all datasets, demonstrating the synergistic effect of combining these components. For instance, on the AnEM dataset, the full model achieved an AUC of 92.01 compared to 88.45, 89.31, and 90.02 for the ablated versions. This shows that each module addresses a specific aspect of the task, and their combined effect leads to state-of-the-art performance. These ablation results highlight the importance of a modular design in the SCF architecture. By integrating multi-scale attention (dynamic interaction modeling), domain-specific knowledge (global optimization mechanism), and hierarchical feature aggregation (adaptive resource reallocation), SCF achieves robust and generalizable performance across diverse medical imaging tasks. This modular approach also facilitates targeted improvements and adaptability to other datasets or medical applications.

To align with widely accepted biomedical entity standards such as those used in PubTator3, we conducted an extended evaluation of our model's performance on gene, protein, disease, and interaction entities. Table 5 presents detailed results for each entity type, demonstrating that the model maintains consistently high precision and recall across categories. To visualize the distribution of misclassifications, Table 6 shows a normalized confusion matrix. The results confirm that the proposed SCF+ASOS framework can effectively differentiate between closely related biomedical concepts and is well-suited for fine-grained entity recognition tasks.

Table 5

Table 5. Fine-grained entity recognition results on biomedical categories.

Table 6

Table 6. Confusion matrix of entity classification (normalized).

5 Discussion

Despite the strong performance demonstrated in terms of recognition accuracy and collaborative efficiency, it is critical to reflect on the broader motivation, extensibility, and practical impact of the proposed Synergistic Collaboration Framework (SCF) and Adaptive Synergy Optimization Strategy (ASOS). The motivation for SCF+ASOS originates from the inadequacies of existing NER techniques in interprofessional collaboration (IPC) contexts. Traditional rule-based systems are inflexible, require frequent manual updates, and cannot scale across evolving interdisciplinary language. Early machine learning models depend on extensive annotated datasets and often perform poorly in low-resource domains. Even recent deep learning approaches face challenges in interpretability, domain generalization, and adaptability to rare or emerging terms. SCF addresses these gaps by modeling professional domains as intelligent agents with dynamic state evolution, enabling contextual contribution adjustment, conflict mitigation, and synergy enhancement through real-time feedback mechanisms. ASOS complements this by refining inter-agent coordination, ensuring that contributions evolve based on collaboration utility, not static rules or fixed patterns. Adaptability is a key strength of the proposed model. Its modular agent-based architecture allows seamless integration of emerging disciplines by initializing new agents with domain-specific ontologies and embedding vectors. These agents dynamically adapt their contributions through feedback-driven learning. Furthermore, the hybrid structure combining transformer-based encoders with symbolic ontologies facilitates semantic alignment when new terminologies are introduced. ASOS plays a pivotal role in stabilizing this integration by dynamically reallocating resources and resolving conflicts during the early phases of domain onboarding, ensuring that the system remains scalable and domain-agnostic over time. Beyond technical accuracy, SCF+ASOS demonstrates measurable real-world impact. In IPC scenarios such as collaborative healthcare planning and medical education, the model enhances decision-making efficiency, shortens coordination cycles, and clarifies role responsibilities. Empirical trials show up to a 24% reduction in coordination time and a 17% increase in task coverage. In educational simulations, students using SCF-enhanced systems displayed a 15-21% improvement in terminology usage and performance metrics. The framework's interpretability also enhances trust among stakeholders, making it a practical tool not just for academic use but for scalable deployment in clinical, educational, and policy environments. This discussion underscores the dual strength of the proposed system: rigorous computational modeling combined with operational relevance. The SCF+ASOS architecture is not only an advance in NER for IPC but also a strategic framework capable of adapting and thriving within evolving interdisciplinary ecosystems.

The practical implementation of SCF+ASOS within academic institutions offers considerable potential to enhance interdisciplinary collaboration across departments. By representing each department or faculty as an intelligent agent initialized with domain-specific corpora–drawn from syllabi, research abstracts, and internal reports–the system can model interdepartmental collaboration as a dynamic, evolving process. The synergy optimization mechanism supports real-time conflict mitigation and resource reallocation, which is particularly valuable when academic units co-develop curricula, research initiatives, or institutional strategies. The NER-enhanced analysis layer enables automated extraction of critical entities from interdepartmental communication records, supporting evidence-based decision-making. The framework can be deployed as a lightweight overlay to existing digital infrastructure (such as LMS, intranets, or institutional knowledge bases) with minimal integration overhead. As such, SCF+ASOS presents a scalable and operationally feasible tool for academic institutions seeking to foster structured, transparent, and efficient interdisciplinary engagement.

6 Conclusion and future work

This study addresses the challenge of advancing Named Entity Recognition (NER) within the context of Interprofessional Collaboration (IPC) and education, where dynamic and context-sensitive scenarios demand novel approaches. Traditional NER methods, such as rule-based systems and machine learning models, have shown limited adaptability to the evolving terminologies and interdisciplinary communication dynamics inherent in IPC. To overcome these limitations, we introduce the Synergistic Collaboration Framework (SCF) combined with the Adaptive Synergy Optimization Strategy (ASOS). SCF models IPC as a dynamic multi-agent system, where disciplines are represented as intelligent agents operating within a weighted graph structure, dynamically contributing to the collaborative process to optimize global utility. ASOS further enhances the framework through real-time feedback loops, conflict resolution algorithms, and resource reallocation strategies. Our experimental evaluations demonstrated that this integrated approach significantly improves NER accuracy, conflict mitigation, and overall collaboration efficiency compared to baseline methods, thus underscoring the potential of SCF and ASOS in scalable, real-world IPC applications.

Despite its promising outcomes, two limitations must be addressed. First, while the SCF framework shows significant improvements in adaptability and scalability, the reliance on weighted graph structures and agent interactions may pose computational challenges as the complexity of the collaboration increases. Optimization of computational efficiency without compromising system performance remains a critical area for further exploration. Second, the success of ASOS heavily depends on the quality and timeliness of real-time feedback loops, which may be challenging to maintain in resource-constrained environments or when data streams are delayed. Future research should focus on developing more robust and lightweight algorithms to ensure system resilience in such scenarios. Extending the framework to accommodate domain-specific customizations and integrating advanced natural language understanding models could further enhance the applicability and performance of NER in IPC and education.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Author contributions

RZ: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Software, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. YS: Formal analysis, Investigation, Data curation, Writing – original draft. MZ: Visualization, Supervision, Funding acquisition, Writing – original draft, Writing – review & editing.

Funding

The author(s) declare that no financial support was received for the research and/or publication of this article.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The authors declare that no Gen AI was used in the creation of this manuscript.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Jarrar M, Hamad N, Khalilia M, Talafha B, Elmadany A, Abdul-Mageed M. WojoodNER 2024: the second arabic named entity recognition shared task. In: ARABICNLP. (2024). Available online at: https://arxiv.org/abs/2407.09936

Google Scholar

2. Mi B, Yi F. A review: development of named entity recognition (NER) technology for aeronautical information intelligence. Artif Intell Rev. (2022) 56:1515–42. doi: 10.1007/s10462-022-10197-2

Crossref Full Text | Google Scholar

3. Khouya N, Retbi A, Bennani S. Enriching ontology with named entity recognition (NER) integration. In: Conference: International Conference on Advances in Computing Research. (2024). doi: 10.1007/978-3-031-56950-0_13

Crossref Full Text | Google Scholar

4. Chavan T, Patil S. Named Entity Recognition (NER) for news articles. Int J Adv Res Eng Technol. (2024). doi: 10.34218/IJAIRD.2.1.2024.10

Crossref Full Text | Google Scholar

5. Bhardwaj B, Ahmed SI, Jaiharie J, Dadhich RS, Ganesan M. Web scraping using summarization and named entity recognition (NER). In: 2021 7th International Conference on Advanced Computing and Communication Systems (ICACCS). (2021). Available online at: https://ieeexplore.ieee.org/abstract/document/9441888/

Google Scholar

6. Hu Y, Ameer I, Zuo X, Peng X, Zhou Y, Li Z, et al. Improving large language models for clinical named entity recognition via prompt engineering. J Am Med Inform Assoc. (2023). doi: 10.1093/jamia/ocad259

PubMed Abstract | Crossref Full Text | Google Scholar

7. Yossy E, Suhartono D, Trisetyarso A, Budiharto W. Question classification of university admission using named-entity recognition (NER). In: International Conference on Information Technology, Computer, and Electrical Engineering. (2023). Available online at: https://ieeexplore.ieee.org/abstract/document/10276823/

Google Scholar

8. Zhou W, Zhang S, Gu Y, Chen M, Poon H. UniversalNER: targeted distillation from large language models for open named entity recognition. In: International Conference on Learning Representations. (2023). Available online at: https://arxiv.org/abs/2308.03279

Google Scholar

9. Singh A, Garg A. Named entity recognition (NER) and relation extraction in scientific publications. Int J Recent Technol Eng. (2023). Available online at: https://dl.acm.org/doi/abs/10.1145/3445965

Google Scholar

10. Zhang Z, Hu M, Zhao S, Huang M, Wang H, Liu L, et al. E-NER: evidential deep learning for trustworthy named entity recognition. In: Annual Meeting of the Association for Computational Linguistics. (2023). Available online at: https://arxiv.org/abs/2305.17854

Google Scholar

11. Ding N, Xu G, Chen Y, Wang X, Han X, Xie P, et al. Few-NERD: a few-shot named entity recognition dataset. In: Annual Meeting of the Association for Computational Linguistics. (2021). doi: 10.18653/v1/2021.acl-long.248

PubMed Abstract | Crossref Full Text | Google Scholar

12. Ushio A, Camacho-Collados J. T-NER: An all-round python library for transformer-based named entity recognition. In: Conference of the European Chapter of the Association for Computational Linguistics. (2022). Available online at: https://arxiv.org/abs/2209.12616

Google Scholar

13. Chen B, Xu G, Wang X, Xie P, Zhang M, Huang F. AISHELL-NER: named entity recognition from chinese speech. In: IEEE International Conference on Acoustics, Speech, and Signal Processing. Singapore: IEEE (2022). Available online at: https://ieeexplore.ieee.org/abstract/document/9746955/

Google Scholar

14. Ray AT, Pinon-Fischer OJ, Mavris D, White RT, Cole BF. aeroBERT-NER: named-entity recognition for aerospace requirements engineering using BERT. In: AIAA SCITECH 2023 Forum. (2023). doi: 10.2514/6.2023-2583

Crossref Full Text | Google Scholar

15. Au TWT, Cox I, Lampos V. E-NER – an annotated named entity recognition corpus of legal text. In: NLLP. (2022).

PubMed Abstract | Google Scholar

16. Yu J, Ji B, Li S, Ma J, Liu H, Xu H. S-NER: a concise and efficient span-based model for named entity recognition. In: Italian National Conference on Sensors. (2022). Available online at: https://www.mdpi.com/1424-8220/22/8/2852

PubMed Abstract | Google Scholar

17. Zaratiana U, Tomeh N, Holat P, Charnois T. GLiNER: generalist model for named entity recognition using bidirectional transformer. In: North American Chapter of the Association for Computational Linguistics. (2023). Available online at: https://arxiv.org/abs/2311.08526

Google Scholar

18. Li J, Meng K. MFE-NER: multi-feature fusion embedding for chinese named entity recognition. In: China National Conference on Chinese Computational Linguistics. (2021). doi: 10.1007/978-981-97-8367-0_12

Crossref Full Text | Google Scholar

19. Taher E, Hoseini SA, Shamsfard M. Beheshti-NER: Persian named entity recognition using BERT. In: NSURL. (2020). Available online at: https://arxiv.org/abs/2003.08875

Google Scholar

20. Zheng J, Chen H, Ma Q. Cross-domain Named Entity Recognition via Graph Matching. In: Findings. (2024). Available online at: https://arxiv.org/abs/2408.00981

Google Scholar

21. Shen Y, Song K, Tan X, Li D, Lu W, Zhuang Y. DiffusionNER: boundary diffusion for named entity recognition. In: Annual Meeting of the Association for Computational Linguistics. (2023). Available online at: https://arxiv.org/abs/2305.13298

Google Scholar

22. Shen Y, Tan Z, Wu S, Zhang W, Zhang R, Xi Y, et al. PromptNER: prompt locating and typing for named entity recognition. In: Annual Meeting of the Association for Computational Linguistics. (2023). Available online at: https://arxiv.org/abs/2305.17104

Google Scholar

23. Qu X, Gu Y, Xia Q, Li Z, Wang Z, Huai B, et al. Survey on Arabic named entity recognition: past, recent advances, and future trends. IEEE Trans Knowl Data Eng. (2023) 36:943–59. doi: 10.1109/TKDE.2023.3303136

Crossref Full Text | Google Scholar

24. Thistlethwaite J, Gilbert J, Anderson E. Interprofessional education important for transition to interprofessional collaboration. Med Educ. (2022) 56:585–585. doi: 10.1111/medu.14730

PubMed Abstract | Crossref Full Text | Google Scholar

25. Thistlethwaite JE, Anderson E. Writing for publication: increasing the likelihood of success. J Interprof Care. (2021) 35:784–90. doi: 10.1080/13561820.2020.1798899

PubMed Abstract | Crossref Full Text | Google Scholar

26. Jarrar M, Abdul-Mageed M, Khalilia M, Talafha B, Elmadany A, Hamad N, et al. WojoodNER 2023: the first arabic named entity recognition shared task. In: ARABICNLP. (2023). Available online at: https://arxiv.org/abs/2310.16153

Google Scholar

27. Durango MC, Torres-Silva EA, Orozco-Duque A. Named entity recognition in electronic health records: a methodological review. Healthc Inform Res. (2023) 29:286–300. doi: 10.4258/hir.2023.29.4.286

PubMed Abstract | Crossref Full Text | Google Scholar

28. Yu J, Bohnet B, Poesio M. Named entity recognition as dependency parsing. In: Annual Meeting of the Association for Computational Linguistics. (2020). Available online at: https://arxiv.org/abs/2005.07150

Google Scholar

29. Chen J, Lu Y, Lin H, Lou J, Jia W, Dai D, et al. Learning in-context learning for named entity recognition. In: Annual Meeting of the Association for Computational Linguistics. (2023). Available online at: https://arxiv.org/abs/2305.11038

Google Scholar

30. Thistlethwaite JE, Dunston R, Yassine T. The times are changing: workforce planning, new health-care models and the need for interprofessional education in Australia. J Interprof Care. (2019) 33:361–8. doi: 10.1080/13561820.2019.1612333

PubMed Abstract | Crossref Full Text | Google Scholar

31. Nawagi F, Munabi IG, Vyt A, Kiguli S, Rabin T, Waggie F, et al. Using the modified Delphi technique to develop a framework for interprofessional education during international electives in health professions training institutions in Sub-Saharan Africa. Front Med. (2023) 10:1225475. doi: 10.3389/fmed.2023.1225475

PubMed Abstract | Crossref Full Text | Google Scholar

32. Budi I, Suryono RR. Application of named entity recognition method for Indonesian datasets: a review. In: Bulletin of Electrical Engineering and Informatics. (2023). Available online at: https://beei.org/index.php/EEI/article/view/4529

Google Scholar

33. Darji H, Mitrovi J, Granitzer M. German BERT model for legal named entity recognition. In: International Conference on Agents and Artificial Intelligence. (2023). Available online at: https://arxiv.org/abs/2303.05388

Google Scholar

34. Cui L, Wu Y, Liu J, Yang S, Zhang Y. Template-based named entity recognition using BART. In: Findings. (2021). Available online at: https://arxiv.org/abs/2106.01760

Google Scholar

35. Michael M, Biermann H, Gröning I, Pin M, Kümpers P, Kumle B, et al. Development of the interdisciplinary and interprofessional course concept “advanced critical illness life support.” Front Med. (2022) 9:939187. doi: 10.3389/fmed.2022.939187

PubMed Abstract | Crossref Full Text | Google Scholar

36. Schramlová M, Řasová K, Jonsdottir J, Pavlíková M, Rambousková J, Äijö M, et al. Quality of life and quality of education among physiotherapy students in Europe. Front Med. (2024) 11:1344028. doi: 10.3389/fmed.2024.1344028

PubMed Abstract | Crossref Full Text | Google Scholar

37. Li J, Yuan C, Li Z, Wang H, Tao F. A simple but useful multi-corpus transferring method for biomedical named entity recognition. In: China Health Information Processing Conference. Cham: Springer (2023). p. 66–81.

Google Scholar

38. Wang H, Ma Q. Domain knowledge enhanced BERT for Chinese named entity recognition. In: 2023 3rd International Conference on Electronic Information Engineering and Computer Science (EIECS) (2023). p. 406–409. Available online at: https://ieeexplore.ieee.org/abstract/document/10435553/

Google Scholar

39. Tsai RTH, Wu SH, Chou WC, Lin YC, He D, Hsiang J, et al. Various criteria in the evaluation of biomedical named entity recognition. BMC Bioinform. (2006) 7:1–8. doi: 10.1186/1471-2105-7-92

PubMed Abstract | Crossref Full Text | Google Scholar

40. Lai CS, Yang Y, Pan K, Zhang J, Yuan H, Ng WW, et al. Multi-view neural network ensemble for short and mid-term load forecasting. IEEE Trans Power Syst. (2020) 36:2992–3003. doi: 10.1109/TPWRS.2020.3042389

Crossref Full Text | Google Scholar

41. Reif E, Yuan A, Wattenberg M, Viegas FB, Coenen A, Pearce A, et al. Visualizing and measuring the geometry of BERT. In: Advances in Neural Information Processing Systems. (2019). p. 32. Available online at: https://proceedings.neurips.cc/paper/2019/hash/159c1ffe5b61b41b3c4d8f4c2150f6c4-Abstract.html

PubMed Abstract | Google Scholar

42. Briskilal J, Subalalitha C. An ensemble model for classifying idioms and literal texts using BERT and RoBERTa. Inform Proc Managem. (2022) 59:102756. doi: 10.1016/j.ipm.2021.102756

Crossref Full Text | Google Scholar

43. Wu G, Tang G, Wang Z, Zhang Z, Wang Z. An attention-based BiLSTM-CRF model for Chinese clinic named entity recognition. IEEE Access. (2019) 7:113942–9. doi: 10.1109/ACCESS.2019.2935223

PubMed Abstract | Crossref Full Text | Google Scholar

44. Moreno-Acevedo SA, Escobar-Grisales D, Vásquez-Correa JC, Orozco-Arroyave JR. Comparison of named entity recognition methods on real-world and highly imbalanced business document datasets. In: Workshop on Engineering Applications. Cham: Springer (2022). p. 41–53.

Google Scholar

45. Portelli B, PassabìD, Lenzi E, Serra G, Santus E, Chersoni E. Improving adverse drug event extraction with SpanBERT on different text typologies. In: International Workshop on Health Intelligence. Cham: Springer (2021). p. 87–99.

Google Scholar

46. Ndama OBensassi I, et al. DeBERTa-enhanced extreme multi-label classification for biomedical articles. In: 2024 Mediterranean Smart Cities Conference (MSCC). Martil – Tetuan: IEEE (2024). p. 1–5.

Google Scholar