数据驱动的前庭功能磁共振成像在PPPD及BPPV后残余头晕中的应用：基于人工智能表型分析与纵向建模的叙述性方法学综述。

文献信息

中文摘要

背景： 眩晕和慢性头晕的神经影像学研究已发现前庭、视觉、躯体感觉及认知-情绪网络存在异常，但传统的组水平分析在捕捉个体异质性和纵向变化方面存在局限。本叙述性方法学综述整合现有证据，并提出一个结合功能磁共振成像（fMRI）、人工智能（AI）和先进统计建模的框架，锚定于两个优先研究场景：PPPD表型分析和BPPV复位后残余头晕的纵向特征描述。方法：AI，特别是无监督学习，可能有助于从高维fMRI特征中识别候选网络亚型，而混合效应模型、非线性模型和潜在轨迹模型可以评估其纵向稳定性、临床相关性和时间演变。结果：这些方法共同构成了一个互补的模式发现-纵向验证框架。结论：然而，当前证据仍受限于小样本、单中心设计、技术异质性和外部验证稀缺。因此，候选PPPD亚组和BPPV复位后提出的恢复轨迹应被视为产生假设，而非已确立的fMRI表型。该框架因此主要旨在作为研究方法来用于机制研究、候选生物标志物发现和前瞻性验证，而非临床可部署的决策工具。未来的进展将需要更大的多中心队列、协调的采集和预处理、严格控制混杂因素和多重检验、独立验证，以及证明其超越现有临床评估的增量价值。

英文摘要

Neuroimaging studies of vertigo and chronic dizziness have identified abnormalities across vestibular, visual, somatosensory, and cognitive-emotional networks, but conventional group-level analyses are limited in capturing individual heterogeneity and longitudinal change. This narrative methodological review integrates current evidence and proposes a framework combining functional magnetic resonance imaging (fMRI), artificial intelligence (AI), and advanced statistical modeling, anchored in two priority research scenarios: PPPD phenotyping and longitudinal characterization of residual dizziness after BPPV repositioning. AI, particularly unsupervised learning, may help identify candidate network subtypes from high-dimensional fMRI features, whereas mixed-effects, nonlinear, and latent trajectory models can evaluate their longitudinal stability, clinical relevance, and temporal evolution. Together, these approaches form a complementary pattern-discovery-longitudinal-validation framework. Current evidence, however, remains limited by small samples, single-center designs, technical heterogeneity, and scarce external validation. Accordingly, candidate PPPD subgroups and proposed recovery trajectories after BPPV repositioning should be regarded as hypothesis-generating rather than established fMRI phenotypes. The framework is therefore intended primarily as a research methodology for mechanistic investigation, candidate biomarker discovery, and prospective validation rather than as a clinically deployable decision tool. Future progress will require larger multicenter cohorts, harmonized acquisition and preprocessing, rigorous control of confounding and multiple testing, independent validation, and demonstration of incremental value beyond established clinical assessment.