Forecasting patient flows with pandemic induced concept drift using explainable machine learning.

Area of Science:

• Healthcare Operations Research
• Public Health Informatics
• Data Science

Background:

• Accurate patient arrival forecasting is crucial for Urgent Care Clinics (UCCs) and Emergency Departments (EDs) resource management and patient care.
• Forecasting patient flow is complex due to multiple influencing factors, further complicated by the COVID-19 pandemic and associated lockdowns.
• Existing models often lack adaptability to rapidly changing public health conditions.

Purpose of the Study:

• To investigate the utility of novel quasi-real-time variables for improving patient flow forecasting models.
• To enhance model adaptability to pandemic-induced disruptions.
• To explore the internal mechanics of forecasting models using eXplainable AI techniques.

Main Methods:

• Employed a Voting ensemble method combining machine learning and statistical techniques.
• Integrated quasi-real-time variables: Google search terms, pedestrian traffic, influenza incidence, and COVID-19 Alert Levels.
• Utilized eXplainable AI tools to analyze model behavior.

Main Results:

• The Voting ensemble model demonstrated superior reliability in forecasting patient arrivals.
• COVID-19 Alert Level, Google search terms, and pedestrian traffic were key features for generalizable forecasts.
• Proxy variables effectively augmented standard autoregressive features, maintaining forecast accuracy.

Conclusions:

• Novel proxy variables significantly enhance the accuracy and adaptability of patient flow forecasting models.
• The proposed features are effective for maintaining forecast accuracy during future public health crises like pandemics.
• eXplainable AI provides deeper insights into model mechanics, aiding in understanding forecasting drivers.