Using Machine Learning to Estimate Unobserved COVID-19 Infections in North America.

BACKGROUND: The detection of coronavirus disease 2019 (COVID-19) cases remains a huge challenge. As of April 22, 2020, the COVID-19 pandemic continues to take its toll, with >2.6 million confirmed infections and >183,000 deaths. Dire projections are surfacing almost every day, and policymakers worldwide are using projections for critical decisions. Given this background, we modeled unobserved infections to examine the extent to which we might be grossly underestimating COVID-19 infections in North America. METHODS: We developed a machine-learning model to uncover hidden patterns based on reported cases and to predict potential infections. First, our model relied on dimensionality reduction to identify parameters that were key to uncovering hidden patterns. Next, our predictive analysis used an unbiased hierarchical Bayesian estimator approach to infer past infections from current fatalities. RESULTS: Our analysis indicates that, when we assumed a 13-day lag time from infection to death, the United States, as of April 22, 2020, likely had at least 1.3 million undetected infections. With a longer lag time-for example, 23 days-there could have been at least 1.7 million undetected infections. Given these assumptions, the number of undetected infections in Canada could have ranged from 60,000 to 80,000. Duarte's elegant unbiased estimator approach suggested that, as of April 22, 2020, the United States had up to >1.6 million undetected infections and Canada had at least 60,000 to 86,000 undetected infections. However, the Johns Hopkins University Center for Systems Science and Engineering data feed on April 22, 2020, reported only 840,476 and 41,650 confirmed cases for the United States and Canada, respectively. CONCLUSIONS: We have identified 2 key findings: (1) as of April 22, 2020, the United States may have had 1.5 to 2.029 times the number of reported infections and Canada may have had 1.44 to 2.06 times the number of reported infections and (2) even if we assume that the fatality and growth rates in the unobservable population (undetected infections) are similar to those in the observable population (confirmed infections), the number of undetected infections may be within ranges similar to those described above. In summary, 2 different approaches indicated similar ranges of undetected infections in North America. LEVEL OF EVIDENCE: Prognostic Level V. See Instructions for Authors for a complete description of levels of evidence.

Risk of a second wave of Covid-19 infections: using artificial intelligence to investigate stringency of physical distancing policies in North America.

PURPOSE: Accurately forecasting the occurrence of future covid-19-related cases across relaxed (Sweden) and stringent (USA and Canada) policy contexts has a renewed sense of urgency. Moreover, there is a need for a multidimensional county-level approach to monitor the second wave of covid-19 in the USA. METHOD: We use an artificial intelligence framework based on timeline of policy interventions that triangulated results based on the three approaches-Bayesian susceptible-infected-recovered (SIR), Kalman filter, and machine learning. RESULTS: Our findings suggest three important insights. First, the effective growth rate of covid-19 infections dropped in response to the approximate dates of key policy interventions. We find that the change points for spreading rates approximately coincide with the timelines of policy interventions across respective countries. Second, forecasted trend until mid-June in the USA was downward trending, stable, and linear. Sweden is likely to be heading in the other direction. That is, Sweden's forecasted trend until mid-June appears to be non-linear and upward trending. Canada appears to fall somewhere in the middle-the trend for the same period is flat. Third, a Kalman filter based robustness check indicates that by mid-June the USA will likely have close to two million virus cases, while Sweden will likely have over 44,000 covid-19 cases. CONCLUSION: We show that drop in effective growth rate of covid-19 infections was sharper in the case of stringent policies (USA and Canada) but was more gradual in the case of relaxed policy (Sweden). Our study exhorts policy makers to take these results into account as they consider the implications of relaxing lockdown measures.

Deep learning COVID-19 detection bias: accuracy through artificial intelligence.

BACKGROUND: Detection of COVID-19 cases' accuracy is posing a conundrum for scientists, physicians, and policy-makers. As of April 23, 2020, 2.7 million cases have been confirmed, over 190,000 people are dead, and about 750,000 people are reported recovered. Yet, there is no publicly available data on tests that could be missing infections. Complicating matters and furthering anxiety are specific instances of false-negative tests. METHODS: We developed a deep learning model to improve accuracy of reported cases and to precisely predict the disease from chest X-ray scans. Our model relied on convolutional neural networks (CNNs) to detect structural abnormalities and disease categorization that were keys to uncovering hidden patterns. To do so, a transfer learning approach was deployed to perform detections from the chest anterior-posterior radiographs of patients. We used publicly available datasets to achieve this. RESULTS: Our results offer very high accuracy (96.3%) and loss (0.151 binary cross-entropy) using the public dataset consisting of patients from different countries worldwide. As the confusion matrix indicates, our model is able to accurately identify true negatives (74) and true positives (32); this deep learning model identified three cases of false-positive and one false-negative finding from the healthy patient scans. CONCLUSIONS: Our COVID-19 detection model minimizes manual interaction dependent on radiologists as it automates identification of structural abnormalities in patient's CXRs, and our deep learning model is likely to detect true positives and true negatives and weed out false positive and false negatives with > 96.3% accuracy.