基于fNIRS衍生脑功能连接特征预测卒中后吞咽困难严重程度的列线图模型的开发与验证

文献信息

中文摘要

目的： 本研究旨在开发并验证一种基于静息态功能近红外光谱（fNIRS）数据的列线图模型，用于预测卒中后吞咽困难（PSD）的严重程度，为PSD的个体化评估和干预提供依据。
方法： 这项多中心回顾性研究连续纳入了2023年3月至2026年4月期间来自两家医院的178例PSD患者。患者被分为轻度（标准化吞咽评估[SSA]评分<26）和重度（SSA≥26）PSD组。收集静息态fNIRS数据以计算18个吞咽相关脑区之间的功能连接强度，生成513个候选特征。在训练集中，首先使用带10折交叉验证的最小绝对收缩和选择算子（LASSO）回归进行初步特征筛选，随后进行Bootstrap重采样（B=200）以进行稳定性选择，确定最终预测因子。基于所选特征构建多变量逻辑回归模型，并据此建立可视化列线图。全面评估模型性能：通过受试者工作特征曲线下面积（AUC）评估区分能力，通过校准曲线评估校准度，通过决策曲线分析（DCA）评估临床实用性。
结果： 最终将三个稳定的功能连接特征纳入模型：LIPG_RIPG（额下回半球间连接）、RDLPFC_RVAC（右侧背外侧前额叶皮层与右侧视觉联合皮层之间的连接）和LFEF_LFEF（左侧额叶眼区区域内连接）。该模型表现出优异的区分性能，训练集AUC为0.928（95% CI：0.878-0.978），内部验证集AUC为0.857（95% CI：0.727-0.986），外部验证集AUC为0.881（95% CI：0.769-0.994）。校准曲线显示预测风险与实际观察之间高度一致，DCA证实该模型在广泛的阈值概率范围（0.05-0.9）内产生正的净临床获益。敏感性分析进一步验证了三个fNIRS特征的预测价值独立于常规临床变量，且不受PSD严重程度截断值选择的影响。
结论： 基于fNIRS功能连接的列线图模型能够有效预测PSD严重程度，具有良好的区分度、校准度和临床转化潜力。它为临床医生提供了一种无创、易用的定量工具，用于早期识别高风险PSD患者，并支持基于证据制定个性化康复策略，促进PSD管理从主观量表评估向客观神经功能精准评估的转变。

英文摘要

OBJECTIVE: This study aimed to develop and validate a nomogram model based on resting-state functional near-infrared spectroscopy (fNIRS) data to predict the severity of post-stroke dysphagia (PSD), providing a basis for individualized assessment and intervention for PSD.
METHODS: This multicenter retrospective study consecutively enrolled 178 PSD patients from two hospitals between March 2023 and April 2026. Patients were classified into mild (Standardized Swallowing Assessment [SSA] score < 26) and severe (SSA ≥ 26) PSD groups. Resting-state fNIRS data were collected to calculate functional connectivity strength among 18 swallowing-related brain regions, generating 513 candidate features. In the training set, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was first used for preliminary feature screening, followed by Bootstrap resampling (B = 200) for stability selection to identify final predictors. A multivariate logistic regression model was constructed based on the selected features, and a visualized nomogram was established accordingly. Model performance was comprehensively evaluated: discriminative ability via the area under the receiver operating characteristic curve (AUC), calibration via calibration curves, and clinical utility via decision curve analysis (DCA).
RESULTS: Three stable functional connectivity features were finally included in the model: LIPG_RIPG (interhemispheric connectivity of the inferior prefrontal gyrus), RDLPFC_RVAC (connectivity between right dorsolateral prefrontal cortex and right visual association cortex), and LFEF_LFEF (intraregional connectivity of left frontal eye field). The model demonstrated excellent discriminative performance, with AUCs of 0.928 (95% CI: 0.878-0.978) in the training set, 0.857 (95% CI: 0.727-0.986) in the internal validation set, and 0.881 (95% CI: 0.769-0.994) in the external validation set. Calibration curves showed high consistency between predicted risk and actual observation, and DCA confirmed the model yielded positive net clinical benefit across a wide threshold probability range (0.05-0.9). Sensitivity analyses further verified that the predictive value of the three fNIRS features was independent of conventional clinical variables and not affected by the choice of PSD severity cutoff.
CONCLUSION: The nomogram model based on fNIRS functional connectivity can effectively predict PSD severity, with sound discrimination, calibration, and clinical translational potential. It provides clinicians with a non-invasive, easy-to-use quantitative tool for early identification of high-risk PSD patients and supports evidence-based formulation of personalized rehabilitation strategies, promoting the transition of PSD management from subjective scale assessment to objective neurofunctional precision evaluation.