Heatwaves are becoming more frequent and intense due to climate change, posing extreme threats to human health, ecosystems, and infrastructure. However, most existing forecasting systems remain limited in arid regions where ground observations are sparse. To enhance heat-wave forecasting, we developed an attention-based Graph Neural Network (GNN) trained on historical weather data (1990-2022) incorporating several atmospheric parameters. The proposed model leverages graph structures to capture spatial dependencies among weather stations, while the attention mechanism enhances the relative importance of meteorological predictors associated with heatwave persistence and stable atmospheric and reduced ventilation conditions. The model was tested using out-of-sample data from 2023 to 2024. Our analysis reveals that during 1990-2024, there were 9995 heatwave days across 48 UAE stations, with an upward trend since the mid-1990s. The southern and southwest desert regions experience the highest heatwave frequency (>220 events), whereas coastal areas exhibit fewer heatwave days but greater humidity persistence, intensifying thermal discomfort. The attention-based GNN achieved high predictive accuracy (0.96) and effectively captured spatial-temporal patterns of extreme heat events, with robust generalisation and a two-day lead time in predicting heatwave onset. Sensitivity analysis highlighted the relative importance of temperature persistence (a diagnostic indicator of heatwave intensity) and low wind speed conditions associated with atmospheric stagnation in predicting heatwave occurrence. This study demonstrates the applicability of attention-based GNNs for short-term heatwave forecasting in data-sparse arid regions, using the United Arab Emirates (UAE) as a case study. These findings highlight the promise of graph-based machine learning frameworks for improving climate-hazard forecasting and provide a foundation for operational early-warning systems and heat-risk adaptation planning in vulnerable environments.