Hypoxia, defined by low dissolved oxygen (DO) levels, is an important ecological challenge in semiarid river ecosystems, which are particularly vulnerable due to agricultural and urban pressures. This study introduces an explainable machine learning framework combined with survival analysis (SA) to predict hypoxia risk in the Karkheh River Basin, Iran. The model integrates SHAP (SHapley Additive exPlanations)-driven interpretability to enhance transparency, allowing stakeholders to identify key environmental predictors. Eleven months of water quality monitoring data collected from 8 stations across the basin were used to train a random forest model, with temperature identified as the dominant predictor for hypoxia onset, with a threshold at 18 degrees C. Additional key factors such as turbidity, total suspended solids, and nutrient concentrations also significantly influenced DO levels. The model achieved a predictive accuracy of R2 = 0.84, demonstrating high reliability in forecasting hypoxia risks. The SA component further quantified the timing and duration of hypoxia, revealing that temperature is a crucial factor in hypoxia risk. This early-warning framework is cost-effective and scalable, offering actionable insights for water resource management in regions with limited monitoring infrastructure. The study contributes to Sustainable Development Goals 6 (Clean Water and Sanitation) and 13 (Climate Action) by providing a tool for sustainable management of river ecosystems in data-scarce regions. The interpretability provided by SHAP allows for clear communication of model predictions, facilitating decision-making and stakeholder engagement.