Poor urban air quality is a major public health concern, especially in highly urbanized areas. This study aims to predict health risks associated with air pollution using machine learning techniques based on environmental variables. The dataset used, Urban Air Quality and Health Impact, contains 1,000 rows and 46 columns, including temperature, humidity, wind speed, dew point, ultraviolet (UV) index, and health risk scores from major U.S. cities. As an improvement over previous studies using linear regression and Random Forest (R-squared 0.89; Mean Squared Error/MSE 0.65), this research implements an optimized Extreme Gradient Boosting (XGBoost) model. The model was fine-tuned using Randomized Search on key hyperparameters and evaluated with an 80:20 data split. It achieved an R-squared of 0.9692 and MSE of 0.0122. Dew point and wind speed were identified as the most influential features. Although synthetic, the dataset reflects environmental patterns similar to Indonesian urban areas. This study does not adopt a text mining framework but instead uses a supervised regression approach based on environmental features. Its main novelty lies in the first application of an optimized XGBoost model using complex variables such as feels-like temperature to estimate urban health risk. Limitations include the absence of real-world validation with Indonesian data and the lack of analysis on interactions between variables