Cancer patients undergoing active treatment often face adverse clinical events, including unexpected re-hospitalizations and mortality. Using real-world EHR and genomic data, we leveraged machine learning to attempt to predict such adverse patient outcomes. After applying a variety of machine learning algorithms (including logistic regression, random forest, neural networks, and multi-task learning), we found that the model with the highest accuracy and reliability was multi-task learning using the neural network as a basis, with the area under the receiver operating characteristic curve (AUC) value being 0.79. For future directions, our models can be applied across different types of cancer and can be further optimized to account for sources of bias, such as overfitting and sample size. Given some promising preliminary results, our predictive models can potentially be used to help clinicians make decisions about treatment plans and timely intervention for cancer patients.