Objective: This study aims to address inequalities in mild cognitive impairment (MCI) risk among Chinese middle-aged and older adults by developing an integrated learning framework to predict MCI risk and identify key contributing factors.
Methods: Using CHARLS data of 4626 participants, we developed a convolutional neural network-bidirectional long short-term memory-attention (CNN-BiLSTM-Attention) model to capture the temporal and spatial features of MCI progression. SHAP (Shapley Additive Explanations) analysis quantified feature importance and enhanced interpretability, while mediation analysis explored causal pathways, particularly focusing on the role of education. Model performance was compared with eight other frameworks, including LSTM-based models, using Receiver Operating Characteristic (ROC) curves and classification metrics.
Results: The CNN-BiLSTM-Attention model demonstrated relatively promising predictive performance (AUC: 0.7317), with moderately high sensitivity (0.6902) and a high negative predictive value (NPV) of 0.9414. Education emerged as the most critical predictor, followed by Instrumental Activities of Daily Living (IADL) and gender. Mediation analysis revealed that education influenced MCI risk indirectly through health insurance, social interaction, physical activity, and depression.
Conclusion: We present an interpretable, data-driven framework for predicting MCI risk while uncovering key inequality factors, particularly the pivotal role of education. The model's robust performance and interpretability highlight its potential to inform public health strategies and interventions aimed at addressing inequalities in dementia risk.
Keywords: CNN-BiLSTM-Attention; Mediation analysis; SHAP analysis; inequality; integrated learning; mild cognitive impairment.
© 2025 Bi et al.