Predicting incidence density of COVID-19 rebound using tree-based machine learning algorithms

Introduction: A series of strategies adopted by the Chinese government can indeed control the COVID-19 epidemic, but they can also cause negative impact on people's mental health and economic incomes. How to balance the relationship between epidemic prevention and social development is an urgent topic for current research. Methods: : We included 122 rebound events involved 96 cities caused by Delta variant from May 21, 2021 to February 23, 2022 and corresponding 32 social environmental factors. Principal Component Analysis and K-Means were used for dimensionality reduction. Conventional logistic regression model, Random Forest model, and extreme Gradient Boosting model were used to model the factors for incidence density. Results: : A total of 96 cities were clustered into six categories. Cities with the number of cases or incidence density above the median are concentrated in cluster 1 and cluster 6. We selected “older”, “urbanratio”, “unemploy”, “serve”, and “air” as the optimal features, and constructed three concise models. The three models showed good discriminatory powers with AUCs of 0.666, 0.795, and 0.747. Conclusion: Based on available public data, high prediction accuracy of the incidence density of COVID‐19 rebound can be achieved by machine learning methods. Developed level of cities may confer the rebound of COVID-19.