We explore challenges in applying machine learning to neuroscience and psychiatry, focusing on explainability, data noise, confounders, and generalizability to enhance clinical applications.