Forecasting Local Surges in COVID-19 Hospitalizations through Adaptive Decision Tree Classifiers

Area of Science:

• Public Health Surveillance
• Epidemiology
• Machine Learning in Healthcare

Background:

• The COVID-19 pandemic strained healthcare systems, necessitating accurate prediction of hospital capacity risks.
• Existing risk metrics, like CDC's Community Levels, had limitations in real-time updates and direct outcome prediction.
• Lack of interpretability in some models hindered decision-making for mitigating hospital surges.

Purpose of the Study:

• To evaluate adaptive decision tree classifiers for predicting surges in COVID-19 hospitalizations.
• To compare the performance of decision tree classifiers against logistic regression, neural networks, and CDC's Community Levels.
• To assess the interpretability and real-time update capabilities of decision tree models for public health decision-making.

Main Methods:

• Real-time decision tree classifiers were developed to predict COVID-19 hospitalization surges from July 2020 to November 2022.
• Performance was evaluated using metrics such as area under the receiver-operating characteristic curve (auROC) and area under the precision-recall curve (auPRC).
• Comparisons were made with logistic regression, neural network models, and the CDC's Community Levels.

Main Results:

• Decision tree classifiers achieved an auROC >80% for most weeks, outperforming CDC's Community Levels in predicting high hospital occupancy.
• The auPRC, sensitivity, and specificity varied between 20%-100% over time, aligning with pandemic waves.
• Decision tree models showed comparable performance to logistic regression and neural networks, offering superior interpretability.

Conclusions:

• Routinely collected hospital surveillance data can be used to adaptively update decision tree classifiers for predicting hospitalization surges.
• Decision tree classifiers provide interpretable rules for understanding and responding to epidemic changes.
• The sensitivity and specificity of these classifiers can fluctuate significantly across different pandemic waves.