Early Anomaly Detection in Time Series: A Hierarchical Approach for Predicting Critical Health Episodes