2021/06/08 by Harshavardhan Kamarthi, Kamarthi, Harshavardhan, Alexander Rodríguez +3 · 1 citation
Computer Science · Mathematics · Medicine · #Anomaly Detection Techniques and Applications #Artificial Intelligence (cs.AI) #COVID-19 diagnosis using AI #COVID-19 epidemiological studies #Computers and Society (cs.CY) #Data-Driven Disease Surveillance #FOS: Computer and information sciences #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2106.04420
openalex publication_date 2021/06/08 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
In real-time forecasting in public health, data collection is a non-trivial\nand demanding task. Often after initially released, it undergoes several\nrevisions later (maybe due to human or technical constraints) - as a result, it\nmay take weeks until the data reaches to a stable value. This so-called\n'backfill' phenomenon and its effect on model performance has been barely\nstudied in the prior literature. In this paper, we introduce the multi-variate\nbackfill problem using COVID-19 as the motivating example. We construct a\ndetailed dataset composed of relevant signals over the past year of the\npandemic. We then systematically characterize several patterns in backfill\ndynamics and leverage our observations for formulating a novel problem and\nneural framework Back2Future that aims to refines a given model's predictions\nin real-time. Our extensive experiments demonstrate that our method refines the\nperformance of top models for COVID-19 forecasting, in contrast to non-trivial\nbaselines, yielding 18% improvement over baselines, enabling us obtain a new\nSOTA performance. In addition, we show that our model improves model evaluation\ntoo; hence policy-makers can better understand the true accuracy of forecasting\nmodels in real-time.\n