在基于大语言模型的听力损失计算模拟中探索耳鸣样中枢过度活跃

文献信息

中文摘要

背景： 耳鸣通常与听力损失后适应不良的神经可塑性相关。尽管大语言模型（LLM）在语言处理和表征方面与人脑具有相似性，但其用于建模耳鸣相关神经机制的潜力仍然有限。本研究旨在开发一个基于LLM的框架，以研究耳鸣相关的适应不良神经可塑性。方法：将一个LLM划分为八个可独立训练的区域。通过在常识推理任务训练期间施加词元特异性的高频衰减，构建了与外周听力损失相关的传入神经阻滞改变探索性计算类比。通过区域特异性可塑性消融（一次选择性地禁用一个区域的可塑性），量化了各区域对耳鸣样过度活动的贡献。使用中心化核对齐评估区域间的表征关系。结果：该计算性传入神经阻滞类比在很大程度上保持了任务准确率，但表现出活动增加。区域1是唯一对过度活动具有净抑制性贡献的区域，而区域2表现出最大的正向贡献；区域6和区域8也有正向贡献。区域1和区域2在所有条件下均保持表征上的独特性。传入神经阻滞倾向于降低区域1与区域2的相似性，而除区域6和区域8外，大多数区域的可塑性消融倾向于增加该相似性。结论：所提出的基于LLM的框架捕捉到了与当前听力损失相关耳鸣理论一致的网络层面功能分离，识别出不同的抑制性、转换性和高阶调制性作用。这些发现凸显了LLM在研究耳鸣背后适应不良神经改变方面的潜力。该框架可能为构建患者特异性数字孪生奠定基础。

英文摘要

Tinnitus is commonly associated with maladaptive neural plasticity following hearing loss. Although large language models (LLMs) share similarities with the human brain in language processing and representation, their potential for modeling tinnitus-related neural mechanisms remains limited. This study aimed to develop an LLM-based framework to investigate tinnitus-related maladaptive neural plasticity. An LLM was partitioned into eight independently trainable regions. An exploratory computational analogy of deafferentation-related changes associated with peripheral hearing loss was developed by applying token-specific high-frequency attenuation during training on a commonsense reasoning task. Region-specific contributions to tinnitus-like excess activity were quantified using region-wise plasticity ablation, selectively disabling plasticity in one region at a time. Representational relationships among regions were evaluated using centered kernel alignment. The computational deafferentation analog largely preserved task accuracy but showed increased activity. Region 1 was the only region with a net suppressive contribution to excess activity, whereas region 2 exhibited the largest positive contribution; regions 6 and 8 also contributed positively. Regions 1 and 2 remained representationally distinct across all conditions. Deafferentation tended to reduce region 1-region 2 similarity, whereas plasticity ablation in most regions, except for regions 6 and 8, tended to increase it. The proposed LLM-based framework captured network-level functional dissociations consistent with current theories of hearing loss-associated tinnitus, identifying distinct suppressive, transformation, and higher-order modulatory roles. These findings highlight the potential of LLMs to investigate maladaptive neural changes underlying tinnitus. This framework may provide the foundation for constructing patient-specific digital twins.