AI 中文总结
研究针对神经质量模型在模拟推理中存在的问题，引入NMM-SBI审计框架，先评估模型对观测数据的覆盖，再训练后验估计器评估信息损失，通过多轨迹联合后验检查一致性，应用于两个数据集，区分失败来源，防止误读，提供可扩展约束。
AI 中文摘要
神经质量模型使用低维动力学参数描述群体水平的神经活动，并通过基于模拟的推理在显式似然难以处理时实现后验估计。但模拟数据上良好的后验恢复并不能保证模型覆盖真实观测、摘要统计保留目标信息或多个参数可独立解释。我们引入了NMM-SBI审计框架，它首先评估候选模型配置是否覆盖观测数据，然后为多个层次的参数坐标训练单独的后验估计器以评估摘要诱导的信息损失，使用多轨迹联合后验检查解释的一致性并报告分级证据。我们将该框架应用于两个真实数据集，结果区分了四种失败来源，防止将模拟器内的强恢复误读为有效的生理解释，为可支持结论提供了具有明确边界的可扩展约束。
英文摘要
Neural mass models describe population activity with low-dimensional dynamics, but simulation-based posterior recovery does not ensure that a model fits real observations or that inferred parameters support physiological interpretation. We introduce NMM-SBI Audit, a hierarchical framework that evaluates whether a model configuration covers observed data, assesses recoverability across multilevel parameter coordinates and summary representations, and examines joint parameter compensation and cross-track consistency. In experiments with known ground truth, the framework controlled empirical error rates and detected prespecified failures. Applied to real data, a single-source Epileptor model failed to cover core seizure statistics of SOZ-local iEEG, rendering simulation-recoverable targets unsuitable for patient-specific mechanistic interpretation. In contrast, a CMC-inspired auditory network model showed no systematic representation-level mismatch and supported conditional recovery of selected superficial-layer and inhibitory gains, while revealing parameter compensation, summary information loss, and instability of the active structure. These results show that observation fit, target recoverability, and joint interpretability provide distinct levels of evidence. NMM-SBI Audit offers a scalable approach to limiting unsupported mechanistic claims in simulation-based inference of neural dynamics.
