Forecasting extreme heat weeks in advance is crucial for public health and agriculture, but traditional weather models often struggle beyond a few days. In this study, we developed a machine learning (ML) method that enhances the ECMWF weather model for predicting summer temperatures in the U.S., especially during extreme heat. Our approach combines ECMWF forecasts with other climate drivers, soil moisture, ocean temperatures, and large-scale patterns like El Niño and the Madden-Julian Oscillation. Using data from 2003–2023, we trained an ML model (XGBoost) to predict weekly average temperatures and the likelihood of extreme heat events up to 3 weeks ahead. Our hybrid method consistently outperformed both the ECMWF baseline and standard ML models, especially in weeks 2 and 3. Performance was strongest in Texas and the Southeast. Soil moisture was a key predictor in southern states, while ocean patterns gained importance with longer lead times. This approach improves early warning of extreme heat and could support climate-resilient planning. However, regular updates will be needed as the climate evolves.