Depression is a common mental disorder that affects millions of people worldwide. Psychological assessments remain the most commonly used diagnostic tools. However, this reliance highlights the opportunity to explore alternative approaches based on the use of machine learning models. This study explores a multimodal graph-based machine learning approach that combines electroencephalography (EEG), voice signals, demographic information, and psychological test results to detect depression. Two groups of graphs were generated using different combinations of features. Feature selection was subsequently performed, yielding distinct subsets of relevant features derived from each group of graphs. The graph2vec model was then employed to generate embeddings for each graph group and each subset of relevant features. Seven machine learning algorithms were trained using the embeddings as feature vectors. The results demonstrate competitive performance compared to those reported in the literature, achieving F1-scores above 0.85 while relying on less complex methods and using only a small subset of the extracted features. The methodology employed and the results obtained are promising, highlighting the potential of graph-based approaches for performing multimodal classification tasks. However, there are limitations mainly related to associated with computational resources that should be analyzed in greater detail.