Objective To preliminarily construct a machine learning model, aimed at predicting the risk of mild cognitive impairment (MCI), and to provide a reference for healthcare professionals in early and rapid screening of MCI.Methods From July to September 2024, a convenience sampling method was employed to select 294 elderly individuals from two neighborhoods in Nanjing.Facial features were extracted using OpenFace 3.0 while subjects viewed happy, neutral, and sad videos.Significant facial features were categorized into seven combinations:happy, neutral, sad, happy+neutral, happy+sad, neutral+sad, and happy+neutral+sad.The feature combinations were used as input variables, and the presence of MCI was the outcome variable.The dataset was split into training and testing sets in a 7∶3 ratio to construct the XGBoost machine learning model.The model′s discriminative perfor-mance was evaluated using accuracy, precision, recall, F1 score, and area under the curve (AUC-ROC) values, with SHAP analysis conducted on the best-performing facial feature combination model.Results Comparison of facial features revealed significant differences in AU04_AUI, AU06_AUI, AU10_AUP, and AU12_AUP among the MCI group while watching happy videos compared to the non-MCI group.The MCI group also exhibited significant differences in nine facial features when watching neutral videos and eight features when watching sad videos.All XGBoost models constructed from facial feature combinations showed AUC values greater than 0.6, with the sad video model achieving the highest AUC of 0.71.SHAP analysis of the sad video model indicated that the top three predictive factors were AU04_AUI, AU20_AUP, and AU07_AUI.Conclusion A preliminary XGBoost machine learning model based on facial features has been constructed to assist in the early identification of MCI risk, providing a refe-rence for early warning and intervention strategies for MCI.