HierSTT Reports 32% Lower Forecast Error Across Portugal’s Emergency Departments

Researchers have proposed a Transformer architecture designed to forecast emergency department demand at three connected levels: individual hospitals, regional health administrations and the national healthcare system. In an arXiv preprint, the team reports that the model, called HierSTT, reduced average Weighted Absolute Percentage Error by 32% relative to its strongest non-hierarchical deep learning baseline.

The result comes from experiments on a Portuguese dataset covering 81 hospitals and five regional health administrations. It should be read as an author-reported research result rather than an independently validated product claim: the paper is a preprint, and the available source does not establish that the system has been deployed in clinical operations.

What the researchers built

Many forecasting systems train separate models for different parts of an organization. In healthcare, that can leave planners with hospital forecasts that do not add up to a region’s predicted total, or regional estimates that conflict with the national projection.

HierSTT instead generates forecasts across the hierarchy within one end-to-end system. According to the paper’s abstract, a Temporal Fusion Transformer captures national demand patterns. Spatio-temporal Transformer encoder-decoder modules then model regional and hospital demand while conditioning their predictions on forecasts from the level above.

This makes the work relevant to machine learning teams building forecasting tools for organizations with nested structures. The same technical problem can arise wherever local units roll up into larger operational groups, although the paper’s reported evidence is limited to emergency department data.

How the model encourages consistency

The architecture uses a coherence-aware loss during training. This loss penalizes disagreement between levels, such as when the combined hospital forecasts do not align with the relevant regional estimate.

That approach differs from post-processing methods that first generate independent forecasts and then reconcile them. HierSTT attempts to learn temporal patterns, relationships among locations and cross-level consistency together.

The authors describe the resulting forecasts as “near-coherent,” not perfectly coherent. That distinction matters: the method is designed to reduce aggregation discrepancies, but the abstract does not claim that every output is guaranteed to satisfy exact summation constraints.

The reported evidence at a glance

What the 32% figure means

The headline figure is a relative improvement in average WAPE, a metric that compares total absolute forecast error with the scale of observed demand. It is not a claim that the model predicts demand with 32% absolute error, nor does it mean that every hospital improved by the same amount.

The abstract says HierSTT achieved the reduction against the best non-hierarchical deep learning baseline included in the experiments. It also says the model outperformed all classical hierarchical reconciliation methods evaluated by the researchers. Assessing the strength of those comparisons requires examining the full paper’s forecast horizons, train-test splits, baseline tuning and results at each level of the hierarchy.

Why the approach matters to AI developers

A single accurate aggregate forecast is not enough when decisions are distributed across multiple organizational levels. Hospital teams may need local estimates for staffing and bed planning, while regional and national authorities work with aggregated demand. Conflicting outputs can make it difficult to decide which forecast should guide resource allocation.

For developers, HierSTT offers one design for treating hierarchical consistency as a training objective rather than a final correction step. Its combination of a Temporal Fusion Transformer, spatio-temporal modules and a coherence penalty could also inform research on other nested forecasting tasks. The current paper, however, does not establish performance outside the Portuguese emergency care setting.

What remains unclear

The available evidence does not show whether HierSTT has been tested prospectively in live hospital workflows or whether its forecasts improve operational outcomes such as staffing efficiency or patient waiting times. Forecast accuracy alone cannot establish those effects.

Generalization is another open question. Portugal’s administrative structure, data collection practices and demand patterns may differ from those in other health systems. The source summary also does not provide deployment costs, inference latency, missing-data behavior or the amount of historical data needed for reliable training.

The arXiv announcement identifies the paper as a version-two replacement. As with any revised preprint, readers evaluating the reported benchmark should check the paper’s version history and confirm which tables, experimental settings and code revision correspond to the 32% result.

What readers can verify

Technical teams can inspect the full preprint for per-level error results, ablation studies and the precise formulation of the coherence-aware loss. They should also review the linked repository for data-access conditions, preprocessing instructions, reproducibility files and the commit associated with the reported experiments. Until independent replication or prospective testing is available, HierSTT is best treated as a research framework with promising benchmark results, not a validated hospital planning product.

Source: arXiv, “Hierarchical Spatio-Temporal Transformer for Coherent Emergency Department Forecasting,” https://arxiv.org/abs/2607.27106; associated project repository, https://github.com/FilipaLino/HierSTT.