Accurate prediction of intensive care unit (ICU) length of stay (LOS) is crucial for optimizing resource allocation and improving patient outcomes. We propose ContrastLOS, a novel graph-based deep learning model that integrates graph transformer networks with contrastive pre-training to enhance ICU LOS prediction. By constructing dynamic patient similarity graphs, ContrastLOS captures intricate inter-patient relationships and temporal dependencies. Through contrastive learning on unlabeled data, the model develops robust patient embeddings suitable for both classification and regression tasks. Experiments on three comprehensive ICU datasets—MIMIC-III, MIMIC-IV v3.0, and eICU—show that ContrastLOS consistently outperforms state-of-the-art baselines. For classification (predicting whether LOS exceeds 3 or 7 days), it achieves AUROC scores of up to 82.1%, improving performance by up to 5.2% compared to competitive methods. In regression tasks, ContrastLOS attains the lowest RMSE (2.41 days) and MAE (1.48 days) on the multi-center eICU dataset. Notably, it maintains an AUROC of 76.8% with only 10% labeled data, highlighting its effectiveness in low-resource settings. These findings suggest that ContrastLOS can serve as a robust clinical decision support tool for critical care management.