This week’s list of data news highlights covers October 21, 2020 – November 6, 2020 and includes articles about forecasting flu outbreaks with location data and decreasing offensive language in video game chats.
Researchers at the Massachusetts Institute of Technology have developed an AI algorithm that can analyze coughs recorded on cellphones to distinguish between a cough caused by COVID-19 and a cough caused by other infections. According to the U.S. Centers for Disease Control and Prevention, 40 percent of individuals who have COVID-19 do not know that they are infected with the virus, meaning these individuals can easily spread the virus through communities before officials are able to track it down. To better identify and track the virus, researchers trained their algorithm to identify and extract distinct features of coughs that are inaudible to the human ear. The researchers used 70,000 samples of cough recordings to train the AI system, 2,500 of which were from individuals who tested positive for the virus. When tested, the algorithm accurately identified 98.5 percent of patients who displayed symptoms and had confirmed cases of COVID-19.
Scientists at the Department of Energy’s Lawrence Berkeley National Laboratory have developed a machine learning tool to accelerate the development of cells for specified goals, such as engineering a microbe to produce a certain drug. Traditionally, bioengineering has relied on trial and error, making the process slow and laborious. To speed up this process, the tool predicts how changes in a cell’s DNA or biochemistry will affect its behavior, enabling the tool to make recommendations for the next engineering cycle with probabilistic predictions for how closely different recommendations meet the end goal. When scientists used the tool to guide the bioengineering process for increasing tryptophan production, also known as baker’s yeast, the tool recommended a series of molecular interactions that increased production by 106 percent.
Researchers at the California Institute of Technology have developed a deep learning technique that can easily solve partial differential equations (PDEs) which are a set of mathematical equations that are good at describing physical phenomena, such as heat diffusion. Typically, researchers solve PDEs by defining the problem set using geometry and training neural networks to solve the problem. However, researchers have discovered that defining the inputs and outputs of PDEs using mathematical analysis and integration is better for modeling phenomena such as the motion of air, allowing neural networks to better solve the problem.When tested on the Navier-Stokes equations, a family of PDEs used to describe the motion of any fluid, researchers found that defining PDEs using methods of mathematical analysis enabled neural networks to solve the equations 1,000 times faster and had a 30 percent lower error rate than they would if they had modeled PDEs using geometry.
Researchers from the Yale School of Medicine and Public Health have developed a natural language processing tool that analyzes pathology reports and identifies individuals with HPV-related cancers. Pathology reports are typically written in lengthy free-form text that contains extraneous information for diagnosis, making it time-consuming for doctors to parse through copious amounts of data in order to diagnose a patient in a timely manner. To extract information from pathology reports more efficiently, the researchers trained the tool to identify and extract information from the reports about the physical signs of HPV, such as genital warts, and HPV testing results. The researchers tested the tool on 949 reports and measured the accuracy, recall, and precision of the tool and found that it identified abnormal cells, tissues, and positive HPV tests with an accuracy rate of 91 percent.
Researchers at the Stevens Institute of Technology in New Jersey have developed a predictive analytics algorithm that uses location data to forecast flu outbreaks. Models that do not use location data and rely solely on data about past influenza outbreaks are unable to uncover patterns across space and time, leading to less accurate outbreak predictions. To increase accuracy of influenza outbreak forecasts, the team trained their algorithm using state and regional data from the United States and Japan and also encoded flu infections as interconnected regional clusters. This enabled the algorithm to find patterns in the way influenza infections flow from one region to another, allowing it to predict outbreaks up to 15 weeks in advance. The algorithm also explains why it makes specific predictions and how it thinks outbreaks in different locations are impacting one another.
Blizzard Entertainment, the company behind the popular video game Overwatch, is using machine learning to flag reports about offensive language and behavior. There are over 40 million people worldwide who play Overwatch, but in the past two years developers have seen a decline in the number of players due to reports of negative online interactions. To address these problems, the company has developed an algorithm that identifies and flags offensive language and behavior faster than a human would, enabling administrators to issue appropriate penalties against repeating offenders more quickly. After being implemented in Overwatch and Heroes of the Storm, another Blizzard Entertainment game, disruptive players have spent 50 percent less time in the games than before.
Sentin, a German company that develops software to improve quality controls in production, has developed an AI recognition system that uses neural networks to detect anomalies or errors in production lines, such as dents in machine-produced cups, from photos. Oftentimes human inspectors cannot detect errors because of eye strain or fatigue, but AI systems are able to catch these errors by using cameras to capture images of the products on the line and by using neural networks to compare the captured images to a reference set of images of what the product should look like. When the system detects an error, it produces a signal that informs the production line manager that this product may need to be discarded.
Technology entrepreneur Atima Lui from Massachusetts has partnered with researchers at York University in Toronto to develop Nudemeter, a computer vision tool that analyzes selfies and quiz responses to find users makeup products that match their skin tones. The cosmetics industry has only recently begun to become more accessible for many darker-skinned people with the introduction of more inclusive shade selections, but color matching still relies on guess work. To address this, Lui first trained Nudemeter using skin tone images from volunteers across the world and then worked with York University researchers to ensure the algorithm could match products to users based on users’ true skin tone, regardless of their device or the conditions in which their photo was taken. Currently, Nude Barre, a hosiery company, uses Nudemeter to help shoppers pick the right-colored tights for them.
Researchers from the Massachusetts Institute of Technology have developed a machine learning tool that predicts how a patient with a urinary tract infection (UTI) will respond to initial and secondary antibiotic treatments. To treat UTIs that do not respond to an initial round of antibiotics, doctors often prescribe a second antibiotic treatment, but secondary antibiotics like fluoroquinolone put women at risk for aortic tears, tendon injuries, and staph infections. To reduce the likelihood that a woman will need a secondary antibiotic, researchers trained the tool to predict the probability that an initial antibiotic could treat a patient’s UTI based on information from 10,000 electronic health records. Using these probabilities, the tool then makes a treatment recommendation that selects the treatment that is least likely to result in treatment failure. When tested, the tool allowed clinicians to reduce the use of second-line antibiotics by 67 percent.
French AI software company Therapixel has developed MammoScreen, an AI tool that identifies different abnormalities, such as lesions or lumps, in the breast using digital mammogram scans. Even though mammograms detect breast cancer at an earlier stage and improve prognosis and reduce mortality, radiologists still miss 30 to 40 percent of breast cancers during screenings. MammoScreen works by identifying suspicious regions on the breast and uses four different positions on the mammogram to score the likelihood the region is malignant. When radiologists applied the tool to 120 mammograms, it achieved an accuracy rate of 80 percent, which is 3 percent more than the accuracy rate radiologists achieved when reading mammograms without the tool’s assistance.
Image: Florian Olivo