An ensemble learning framework based on comprehensive gray matter features for identification of mild cognitive impairment in leukoaraiosis

White matter hyperintensities (WMH), also known as leukoaraiosis (LA), is strongly associated with cognitive impairment and lead to an increased risk of dementia. The purpose of this study is to develop a model to effectively and objectively identify WMH patients with cognitive impairment (WMH-MCI). Firstly, the comprehensive multiple cortical morphological measurements were extracted from magnetic resonance imaging (MRI) to enrich the disease characterization information. Then, based on the general eXtreme Gradient Boosting classifier (XGBoost), we designed a data-level fusion resampling method (Fusion + XGBoost) and an algorithm-level focal loss improved XGBoost model (FL-XGBoost), respectively, to solve the imbalance learning problem of classifying WMH-MCI (minority class of 27 samples) and the WMH population without cognitive impairment (WMH-nCI, majority class of 70 samples). Moreover, an ensemble framework based on weighted soft-voting was developed to combine the two models to further improve the overall classification performance and stability of the model. Compared with the baseline XGBoost model trained on the original imbalance dataset (Bacc: 78.20%), both the Fusion + XGBoost model (Bacc: 80.53%) and the FL-XGBoost model (Bacc: 81.25%) could improve the identification accuracy of WMH-MCI, the improvements were 2.33% and 3.05%, respectively. The overall model accuracy with weighted ensemble learning achieved 84.80%, with high sensitivity (85.50%) and specificity (84.14%) at the same time, which was better than that of the single model and significantly improved than the baseline model. The developed model could accurately identify the cognitive impairment in the WMH population, which could assist early clinical diagnosis and timely decision-making.