With the abundance of data at our fingertips, data-driven solutions are being created to solve our modern problems. For the past couple of years, the Covid-19 pandemic has been a major global issue, and as a result, machine learning methods have been used in response. For this post, I will summarize the 2020 article Applications and challenges of AI-based algorithms in the COVID-19 pandemic by Danai Khemasuwan and Henri G Colt. In this literature review, the researchers explored the strengths and limitations of how machine learning algorithms are being used to respond to the Covid-19 pandemic. This is relevant not only in regard to Covid-19 but also in other instances where algorithms and public health intersect.

One use case that the researchers explored is outbreak detection. Outbreak detection is necessary to identify clusters of infection to prevent further spread. Using natural language processing, researchers analyzed seasonal Twitter data for signs of Covid-19 or flu symptoms. Using supervised and semi-supervised techniques, tweets were analyzed to understand the prevalence of Covid-19 symptoms. However, there are explicit biases in terms of the data. Since only a subset of the general population only uses Twitter, analysis from the model can only generalize to Twitter users. These may be limited to certain locations and age groups.

Forecasting of Covid-19 cases, hospitalizations, and death were also used throughout the pandemic. An example was in Brazil, where forecasts of cases for the next three and six days were published. Daily case data was used to create a model that was based on ridge regression, random forest, and support vector regression techniques. Models forecasting cases for the next three days had errors of less than 5%, and forecasts for six days ahead had errors of less than 7%. Short-term forecasting allowed Brazillian officials to have insight into public health policies. However, since the models only used Brazillian health data, they may not be easily generalizable to other countries and populations.

Image classification was used to detect Covid-19 in chest x-rays with a convolutional neural network. Because there was not a lot of chest x-ray data early on in the pandemic, pretrained CNN models in related contexts were applied. By using transfer learning, a majority of the model training had already been done, and the limited Covid-19 chest x-ray data was used to fine-tune it. The CNN models were able to reach a sensitivity rate of 98%. This shows how researchers were able to make do with limited data and still achieve results.

Overall, many different machine-learning solutions have been used throughout the Covid-19 pandemic. This post only provides a small sampling. It is evident that machine-learning techniques have a great positive impact on public health. At the same time, there are limitations. Privacy, data accessibility, and population biases must all be considered. Furthermore, models may not be easily generalizable. Since infection data varies based on location, models that perform well in one location may not apply to others. Overall, it is important to understand the impacts of machine-learning techniques as it relates to public health.