机器学习在孤立性快速眼动睡眠行为障碍识别与临床转化预测中的应用进展

Advances in Machine Learning for the Identification and Phenoconversion Prediction of Isolated Rapid Eye Movement Sleep Behavior Disorder

摘要: 孤立性快速眼动睡眠行为障碍(isolated Rapid eye movement sleep Behavior Disorder, iRBD)是α-突触核蛋白病的重要前驱期临床表型，可进一步转化为帕金森病、路易体痴呆或多系统萎缩，但个体的转化时间与疾病终点存在明显异质性。机器学习能够整合视频多导睡眠监测、脑电、心电、神经影像、临床量表、生物标志物及数字化行为数据，为iRBD筛查、辅助识别和个体化表型转化预测提供新的技术路径。本文采用叙述性综述方法，以临床任务为主线，归纳自动识别、固定时间窗转化分类、转化时间预测及疾病特异性转化预测的代表性研究，并比较其研究对象、输入模态、预测终点、验证方式和评价指标。现有证据显示，自动识别模型已获得较好的内部判别性能，而转化时间及疾病特异性结局预测仍主要处于小样本探索和内部验证阶段。样本量与转化事件不足、标签及终点定义不统一、患者级数据泄漏、外部验证缺乏和概率校准不足，是限制模型可信度和临床应用的主要问题。未来应建立标准化多中心纵向队列，采用生存分析、竞争风险和动态预测框架，并评价模型的校准度、临床净获益、成本、公平性及工作流程影响，推动算法由内部性能比较转向可解释、可验证且能够支持临床管理的风险工具。

Abstract: Isolated rapid eye movement sleep behavior disorder (iRBD) is an important prodromal clinical phenotype of α-synucleinopathies. Individuals with iRBD may subsequently develop Parkinson’s disease, dementia with Lewy bodies, or multiple system atrophy, although the timing and clinical endpoint of phenoconversion vary substantially among individuals. Machine learning can integrate video-polysomnography, electroencephalography, electrocardiography, neuroimaging, clinical scales, biomarkers, and digital behavioral data, providing new approaches to iRBD screening, computer-aided identification, and individualized phenoconversion prediction. In this narrative review, we use clinical tasks as the organizing framework to summarize representative studies of automated identification, fixed-horizon phenoconversion classification, time-to-phenoconversion prediction, and disease-specific phenoconversion prediction. We also compare their study populations, input modalities, prediction endpoints, validation strategies, and performance metrics. Current evidence indicates that automated identification models have achieved good discriminatory performance in internal validation. By contrast, the prediction of phenoconversion timing and disease-specific outcomes remains largely limited to exploratory studies with small samples and internal validation. Insufficient sample sizes and phenoconversion events, inconsistent definitions of labels and endpoints, patient-level data leakage, limited external validation, and inadequate probability calibration are the main barriers to model reliability and clinical application. Future studies should establish standardized, multicenter longitudinal cohorts; adopt survival analysis, competing-risk, and dynamic prediction frameworks; and assess model calibration, clinical net benefit, cost, fairness, and effects on clinical workflows. These efforts will help shift the field from comparisons of internal model performance toward interpretable and externally validated risk tools that can support clinical management.