可解释机器学习用于识别中国阻塞性睡眠呼吸暂停患者的轻度认知障碍：一项多中心横断面研究

文献信息

中文摘要

背景： 轻度认知障碍（MCI）在阻塞性睡眠呼吸暂停（OSA）中具有临床相关性。我们开发并外部验证了一个可解释的机器学习模型，用于识别同时存在的临床定义MCI，以支持优先安排正式认知评估而非用于诊断。
方法： 这项多中心横断面诊断预测研究纳入了来自长春的400例符合条件的OSA患者，分为训练集（n=280）和用于算法选择的内部留出集（n=120），另加一个独立的深圳外部验证样本（n=192）。MCI分类整合了报告的认知下降、客观认知证据、基本保留的日常独立性、痴呆排除以及有医生参与的评估团队共识。在27个候选预测变量中，冗余审查后保留24个，LASSO-Boruta共识保留9个。特征选择在重复交叉验证前于完整训练集上一次性完成，未嵌套于重采样折内。八种算法使用重复10折交叉验证进行调优，仅在训练折内使用SMOTE，并以PR-AUC作为主要指标。RF阈值（0.3228）由训练集折外概率得出，并在验证前锁定。
结果： MCI在93/280（33.2%）、38/120（31.7%）和57/192（29.7%）名参与者中被分类。九个RF预测变量为咖啡摄入、体力活动、AHI、MSaO2、PHQ-9、GAD-7、ESS、CPSS-14和PSQI总分。在用于算法选择的内部留出集中，RF达到ROC-AUC 0.953、PR-AUC 0.936和Brier评分0.060；这些以算法选择为条件的估计并非无偏的泛化估计。在独立外部验证中，ROC-AUC为0.957、PR-AUC为0.809、Brier评分为0.051。校准截距-0.616和斜率0.701表明尽管区分度强，但存在系统性概率高估。
结论： RF模型在深圳表现出强区分度，但外部校准不良和系统性概率高估意味着高AUC并不能使原始概率可迁移；在临床使用前需要进行本地校准评估，必要时重新校准。MCI结局无法仅凭问卷/量表评分复现，补充表S10是回顾性操作规范而非原始CRF。性能证据几乎完全来自男性OSA患者；在女性中的有效性尚未确立，该模型不应在女性中用于临床评估。它是一种研究性筛查优先排序工具，而非诊断、因果、预后或临床影响工具。

英文摘要

BACKGROUND: Mild cognitive impairment (MCI) is clinically relevant in obstructive sleep apnea (OSA). We developed and externally validated an interpretable machine-learning model for concurrent clinically defined MCI to support prioritization for formal cognitive assessment rather than diagnosis.
METHODS: This multicenter cross-sectional diagnostic-prediction study included 400 eligible patients with OSA from Changchun, divided into a training set (n = 280) and an internal algorithm-selection hold-out (n = 120), plus an independent Shenzhen external-validation sample (n = 192). MCI classification integrated reported cognitive decline, objective cognitive evidence, largely preserved everyday independence, dementia exclusion, and assessment-team consensus with physician participation. Of 27 candidate predictors, redundancy review left 24 and LASSO-Boruta consensus retained nine. Feature selection was performed once on the full training set before repeated cross-validation and was not nested within resampling folds. Eight algorithms were tuned using repeated 10-fold cross-validation with training-fold-only SMOTE and PR-AUC as the primary metric. The RF threshold (0.3228) was derived from training out-of-fold probabilities and locked before validation.
RESULTS: MCI was classified in 93/280 (33.2%), 38/120 (31.7%), and 57/192 (29.7%) participants. The nine RF predictors were coffee consumption, physical activity, AHI, MSaO2, PHQ-9, GAD-7, ESS, CPSS-14, and PSQI total scores. In the internal hold-out used for algorithm selection, RF achieved ROC-AUC 0.953, PR-AUC 0.936, and Brier score 0.060; these selection-conditioned estimates are not unbiased generalization estimates. In independent external validation, ROC-AUC was 0.957, PR-AUC 0.809, and Brier score 0.051. Calibration intercept - 0.616 and slope 0.701 indicated systematic probability overestimation despite strong discrimination.
CONCLUSIONS: The RF model discriminated strongly in Shenzhen, but external miscalibration and systematic probability overestimation mean that high AUC does not make raw probabilities transportable; local calibration assessment and, if needed, recalibration are required before clinical use. The MCI outcome cannot be reproduced from questionnaire/scale scores alone, and Supplementary Table S10 is a retrospective operational specification rather than the original CRF. Performance evidence comes almost entirely from male OSA patients; validity in women is unestablished, and the model should not be used for clinical assessment in females. It is a research screening-prioritization tool, not a diagnostic, causal, prognostic, or clinical-impact instrument.