The increasing frequency and intensity of drought events-periods of significant decrease in water availability-are among the most alarming impacts of climate change. Monitoring and detecting these events is essential to mitigate their impact on our society. However, traditional drought indices often fail to accurately detect such impacts as they mostly focus on single precursors. In this study, we leverage machine learning algorithms to define a novel data-driven, impact-based drought index reproducing as target the Vegetation Health Index, a satellite signal that directly assesses the vegetation status. We first apply novel dimensionality reduction methods that allow for interpretable spatial aggregation of features related to precipitation, temperature, snow, and lakes. Then, we select the most informative and non-redundant features through filter feature selection. Finally, linear supervised learning methods are considered, given the small number of samples and the aim of preserving interpretability. The experimental setting focuses on ten sub-basins of the Po River basin, but the aim is to design a machine learning-based workflow applicable on a large scale.