Severe maternal morbidity (SMM) is a critical challenge in obstetric care, where early identification of high-risk patients is essential for preventing life-threatening outcomes. In this study, we present a novel framework, named Graph-based eXplainable Prediction of Severe Maternal Morbidity (SMMGraphX), that constructs a heterogeneous graph from electronic health records to improve the early prediction of SMM. Unlike conventional approaches that treat clinical data as flat tables, our framework explicitly models the rich relational structure connecting patients, diagnosis codes, and laboratory results. This heterogeneous graph captures the multifaceted nature of clinical interactions, preserving both node and relation-type information essential for understanding complex care pathways. The graph-based model learns to prioritise critical information embedded in this network, capturing which types of clinical relationships and entities contribute most to risk prediction. Importantly, the framework offers interpretability through analysis of learned importance weights, allowing post hoc identification of influential clinical factors such as abnormal laboratory values and key diagnostic patterns. This provides clinicians with transparent insights into the drivers of individual risk predictions, supporting more informed and trustworthy decision-making. Evaluated on the MIMIC-IV dataset, the framework achieves an accuracy of 91.78%, precision of 95.40%, recall of 84.50%, and F1-score of 89.53%, outperforming classical machine learning models and state-of-the-art homogeneous graph neural networks. This work offers a robust and transparent solution for risk stratification in maternal health, leveraging heterogeneous graph construction and interpretable learning to inform clinical decision-making and improve outcomes for at-risk mothers.