Pan Ting, Shen Jiang, Xu Man
College of Management and Economy, Tianjin University, Tianjin, China.
Business School, Nankai University, Tianjin, China.
Front Neurosci. 2025 May 20;19:1606038. doi: 10.3389/fnins.2025.1606038. eCollection 2025.
Neurosurgical intelligent question-answering (Q&A) systems offers a novel paradigm to enhance perceptual intelligence-simulating human-like cognitive processing for contextual understanding and emotion interaction. While retrieval-based models lack perceptual adaptability to rare clinical scenarios, and generative LLMs, despite fluency, fail to ground outputs in domain-specific neurosurgical knowledge or doctor expertise. Hybrid frameworks struggle to emulate clinician perceptual workflows (e.g., contextual prioritization, empathy modulation). These present challenges for further improving the semantic understanding, memory integration, and trustworthiness of intelligent Q&A systems in neurosurgery.
To address these challenges, we propose a Multi-Task Knowledge Graph-Augmented Answer Generation model (MT-KGAG), designed to enhance perceptual fidelity. It uses a hybrid attention mechanism to introduce neurosurgical knowledge graph and doctor features in the answer generation model to prioritize clinically salient information akin to human perceptual workflows. Simultaneously, the model employs a multi-task learning framework, jointly optimizing answer generation, candidate answer ranking, and doctor recommendation tasks aligning machine outputs with clinician decision-making patterns while embedding safeguards against hallucination or inappropriate emotional mimicry. Experiments utilize real-world data from a Chinese online health platform, validated through perceptual coherence metrics and ethical robustness assessments.
The MT-KGAG model outperformed all baselines. It achieved an Embedding Average of 0.9439, DISTINCT-2 of 0.2681, and a medical entity density of 0.2471. Medical experts rated patient safety at 4.02/5 and health outcomes at 3.89/5. Additionally, it attained MRR scores of 0.6155 for candidate answer ranking and 0.6169 for doctor recommendation, confirming its multi-task synergy.
MT-KGAG pioneers perception-aware AI in neurosurgery, where LLMs transcend text generation to simulate clinician-like contextual reasoning and ethical judgment. By fusing LLM's generative adaptability with domain-specific knowledge graphs, the model navigates complex trade-offs between empathetic interaction and perceptual safety-delivering responses that are both contextually nuanced and ethically constrained. This work highlights the transformative potential of perceptual intelligence in medical AI, enabling systems to "interpret" patient needs, "recall" specialized knowledge, and "prioritize" clinical relevance while mitigating risks of anthropomorphic overreach.
神经外科智能问答系统提供了一种新的范例，以增强感知智能——模拟类人认知处理以实现情境理解和情感交互。基于检索的模型缺乏对罕见临床场景的感知适应性，而生成式语言模型（LLMs）尽管流畅，但未能将输出基于特定领域的神经外科知识或医生专业知识。混合框架难以模拟临床医生的感知工作流程（例如，情境优先级排序、共情调节）。这些都给进一步提高神经外科智能问答系统的语义理解、记忆整合和可信度带来了挑战。
为应对这些挑战，我们提出了一种多任务知识图谱增强答案生成模型（MT-KGAG），旨在提高感知保真度。它使用混合注意力机制在答案生成模型中引入神经外科知识图谱和医生特征，以优先处理类似于人类感知工作流程的临床显著信息。同时，该模型采用多任务学习框架，联合优化答案生成、候选答案排序和医生推荐任务，使机器输出与临床医生的决策模式保持一致，同时嵌入防止幻觉或不适当情感模仿的保障措施。实验利用来自中国在线健康平台的真实世界数据，通过感知连贯度指标和伦理稳健性评估进行验证。
MT-KGAG模型优于所有基线模型。它的嵌入平均值为0.9439，DISTINCT-2值为0.2681，医学实体密度为0.2471。医学专家对患者安全性的评分是4.02/5，对健康结果的评分是3.89/5。此外，它在候选答案排序方面的MRR分数为0.6155，在医生推荐方面的MRR分数为0.6169，证实了其多任务协同效应。
MT-KGAG开创了神经外科领域感知感知人工智能的先河，其中语言模型超越了文本生成，能够模拟类似临床医生的情境推理和伦理判断。通过将语言模型的生成适应性与特定领域的知识图谱相结合，该模型在共情交互和感知安全之间进行复杂的权衡，提供既具有情境细微差别又受到伦理约束的回应。这项工作突出了感知智能在医学人工智能中的变革潜力，使系统能够“解读”患者需求、“回忆”专业知识并“优先考虑”临床相关性，同时降低拟人化过度的风险。
