(4). That’s when the real issue became clear. The problem wasn’t just the model, it was the data.
Many publicly available datasets (like Kaggle) are collected in controlled environments. They are clean, balanced, and often lack the variability that define real clinical settings.