Hierarchical Spatio-Temporal Transformer for Coherent Emergency Department Forecasting

Researchers introduced HierSTT, a hierarchical spatio-temporal Transformer for emergency department forecasting across hospital, regional, and national levels.

The model combines a Temporal Fusion Transformer for national patterns with spatio-temporal encoder-decoder modules for lower levels, plus a coherence loss to reduce inconsistencies.

On a Portuguese dataset covering 81 hospitals in 5 regional health administrations, HierSTT cut average WAPE by 32% versus the best non-hierarchical deep learning baseline.

It also outperformed classical forecast reconciliation methods, suggesting more reliable demand planning for staffing, beds, and capacity.