On Machine Learning-Based Short-Term Adjustment of Epidemiological Projections of COVID-19 in US

KEFAYATI, S.; Huang, H.; Chakraborty, P.; Roberts, F.; Gopalakrishnan, V.; Srinivasan, R.; Pethe, S.; Madan, P.; Deshpande, A.; Liu, X.; Hu, J.; Jackson, G.

Epidemiological models have provided valuable information for the outlook of COVID-19 pandemic and relative impact of different mitigation scenarios. However, more accurate forecasts are often needed at near term for planning and staffing. We present our early results from a systemic analysis of short-term adjustment of epidemiological modeling of COVID 19 pandemic in US during March-April 2020. Our analysis includes the importance of various types of features for short term adjustment of the predictions. In addition, we explore the potential of data augmentation to address the data limitation for an emerging pandemic. Following published literature, we employ data augmentation via clustering of regions and evaluate a number of clustering strategies to identify early patterns from the data. From our early analysis, we used CovidActNow as our underlying epidemiological model and found that the most impactful features for the one-day prediction horizon are population density, workers in commuting flow, number of deaths in the day prior to prediction date, and the autoregressive features of new COVID-19 cases from three previous dates of the prediction. Interestingly, we also found that counties clustered with New York County resulted in best preforming model with maximum of R2= 0.90 and minimum of R2= 0.85 for state-based and COVID-based clustering strategy, respectively.