Forecasting Patient Flows with Pandemic Induced Concept Drift using Explainable Machine Learning

Real-world disruptions require adaptive AI models: The study demonstrates that traditional forecasting methods fail during major disruptions like pandemics. By incorporating real-time variables such as Google search terms, pedestrian traffic, and COVID-19 Alert Levels, the researchers improved the model's ability to adapt to "concept drift"—the phenomenon where patterns change unexpectedly—showing that AI systems need flexibility to remain accurate during crises.

Explainability revealed which factors actually matter: Using explainable AI tools, the researchers identified that COVID-19 Alert Levels, Google search terms, and pedestrian traffic were the most effective predictors of patient flows. This transparency is crucial for healthcare decision-makers who need to understand why a model makes predictions before trusting it with resource allocation decisions.

Proxy variables can substitute for direct data: The study found that indirect indicators like Google search behavior and foot traffic patterns effectively predicted patient arrivals, even without direct access to comprehensive health data. This has important implications for privacy and data accessibility—healthcare systems may not always need invasive data collection if they can use publicly available proxy measures responsibly.

Ensemble methods outperformed single approaches: A voting ensemble combining machine learning and statistical techniques proved most reliable, suggesting that responsible AI implementation benefits from combining multiple methodologies rather than relying on any single algorithm, reducing the risk of systematic failures.

Healthcare resource planning can be more equitable: Accurate patient flow forecasting enables better staffing and resource distribution, which directly impacts care quality and wait times—particularly important for vulnerable populations who depend on urgent care services.