AUG 11, 2021 4:00 PM PDT

Machine Learning Offers New Way to Detect COVID-19 Infection

WRITTEN BY: Ryan Vingum

Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests have emerged as the standard for testing and confirming COVID-19 infections, with nearly 80 million performed in the US thus far since the start of the pandemic. Additionally, methods such as chest x-rays can complement these tests for a more robust picture of infection. However, according to new research published in PLOS ONE, RT-PCR testing and other mechanisms for confirming infection can be ineffective or even unavailable in certain locations, especially in poorer countries, making accurate and timely reporting of COVID-19 infection difficult.

According to the study published in PLOS ONE, researchers are looking at a way to improve machine learning’s ability to predict COVID-19 infections without a full set of diagnostic data, called the denoising fully connected network (DFCN). The authors of the study concluded that “in the context of the current pandemic and the limitations of the RT-PCR test, this study introduces a model that is accurate and applicable for use in many diverse settings around the world.”

The DFCN is designed to work well in the absence of complete data inputs, which may be beneficial given the inconsistency in which some clinical information is not available or able to be collected about COVID-19. Specifically, researchers note that “the model is trained with randomly masked inputs which, combined with this unique regularization method, instills a robustness to missing input data.”

The study first gathered information about COVID-19 patients, including chest x-rays and other laboratory information to train the DFCN on how to recognize and predict COVID-19 infections. Chest x-rays specifically, were chosen because of the ways COVID-19 can present similarly to pneumonia.

Then, to compare the DFCN's abilities with other types of machine learning used to make predictions in clinical contexts, researchers used data from an ablation study, including imaging scans and laboratory information. Their findings highlighted the superiority of the DFCNs abilities, suggesting their model may be more effective for predicting COVID-19 infections from limited data. 

In their analysis of the DFCN compared to other machine learning networks, the researcher team highlighted that denoising autoencoders are a common approach used in these machine learning applications in clinical contexts. The problem, however, is that autencoders in most machine learning tools, such as the fully convolutional network (FCN), create a bottleneck “efficiency” constraint. The DFCN overcomes this limitation.

More research, including more data sets, could help prove and validate the DFCNs capabilities for predicting clinical information in the absence of complete data.

Sources: PLOS ONE; Centers for Disease Control and Prevention 

About the Author
  • Science writer and editor, with a focus on simplifying complex information about health, medicine, technology, and clinical drug development for a general audience.
You May Also Like
AUG 16, 2021
Infographics
How Can We Make Sure Artificial Intelligence Won't Destroy Humanity?
AUG 16, 2021
How Can We Make Sure Artificial Intelligence Won't Destroy Humanity?
Artifical intelligence (AI) is becoming increasingly pervasive in daily life. While still far from the robitic humanoids ...
AUG 12, 2021
Technology
New Process for Making Pure Synthetic RNA
AUG 12, 2021
New Process for Making Pure Synthetic RNA
Researchers may have found a way to create synthetic RNA that is more “pure” and cost-effective than previou ...
AUG 13, 2021
Earth & The Environment
As Above, So Below: Restoring Wetlands and Mapping Watersheds
AUG 13, 2021
As Above, So Below: Restoring Wetlands and Mapping Watersheds
Water is essential for all life on earth, and much of it is stored in wetlands. Biodiversity is high in wetlands, often ...
SEP 01, 2021
Chemistry & Physics
New Study Suggests One-third of Binary Stars Have Engulfed Their Planets
SEP 01, 2021
New Study Suggests One-third of Binary Stars Have Engulfed Their Planets
While our solar system is relatively calm and inert, some sun-like stars out there will literally eat the planets in the ...
SEP 28, 2021
Earth & The Environment
Using UAVs to Study the Environment
SEP 28, 2021
Using UAVs to Study the Environment
Unmanned aerial vehicles (UAVs), or drones, are a relatively new technology for the general public at an affordable cost ...
OCT 05, 2021
Chemistry & Physics
Researchers Propose Pathway to Plastics Free of Carbon Emissions
OCT 05, 2021
Researchers Propose Pathway to Plastics Free of Carbon Emissions
Plastics may, in general, be a product of fossil fuels like petroleum, but that has not stopped an international team of ...
Loading Comments...