Mild cognitive impairment (MCI) is regarded as an early stage of Alzheimer’s disease (AD), and early diagnosis is essential for timely intervention in disease progression. However, the features extracted by existing methods often exhibit poor performance, and many diagnostic approaches fail to adequately integrate clinical information. Structural Magnetic Resonance Imaging (sMRI) is widely used in clinical settings due to its high signal-to-noise ratio, robustness against artifacts, and ability to provide stable morphological features. In this study, we proposed a framework based on sMRI image data and clinical information for the diagnosis of MCI, particularly utilizing gradient extraction to enhance the representation of brain structures. Firstly, we constructed a morphological brain network by extracting intensity and texture features from sMRI images and computed the global gradients of the brain network. These global gradients were used as node features representing brain regions, enabling the learning of individual embeddings. Next, by concatenating cognitive scores with the learned embeddings, we constructed node features for each individual in the population graph. We then calculated the correlations between risk factors (e.g., age, gender, and APOE gene) to define edge features in the population graph. Finally, we applied population graph learning to classify MCI. The results demonstrated that leveraging global gradients significantly improves diagnostic accuracy for MCI, offering promising insights for future clinical applications. The code is available at https://github.com/AstonshisL/PopGradientGNN .