This week’s list of data news highlights covers July 27-August 2, 2019, and includes articles about an AI system that can detect a deadly crop disease and a system that can spot deepfake images.
1. Predicting a Deadly Kidney Condition
Researchers from DeepMind have developed an AI system to predict if a patient will develop acute kidney injury, a potentially fatal condition that affects 20 percent of patients in U.S. hospitals. The researchers used the electronic health records of 700,000 adults that received treatment at U.S. Department of Veteran Affairs clinics to develop the system’s model. The system analyzes thousands of data points, such as a patient’s age, blood pressure, and movement from one hospital bed to another, to make predictions two days before a potential injury with 56 percent accuracy, which is an improvement upon existing statistical models.
Researchers from North Carolina State University have developed a smartphone-based system that can detect late blight, a common crop disease that causes more than $6 billion in annual losses worldwide. Plant health influences what kinds of chemicals they produce and the researchers’ system uses a strip of sensors that plug into a smartphone to assess plant emissions. The system detects late blight in tomato leaves with greater than 95 percent accuracy.
3. Spotting Links Between Artwork
Researchers from the Czech Technical University have shown how computer vision systems can help art historians identify connections between different artists. The researchers used OpenPose, an open-source program to detect human poses, to analyze 37,000 images of artwork based on the similarity of human poses in the artwork. The algorithms discovered similarities between artworks, such as a sitting Virgin Mary, that would be difficult to match using standard image retrieval methods because other elements of the artwork are substantially different.
4. Detecting Irregular Heartbeats
Researchers from the Mayo Clinic have developed an AI system that can diagnose atrial fibrillation, an irregular heartbeat, in electrocardiographs (ECGs) with 90 percent accuracy. The researchers trained their system on 450,000 ECGs, teaching it to identify small changes in the electrical activity of the heart. The system could help improve the effectiveness and efficiency of ECG screening.
5. Spotting Image Manipulations at the Pixel Level
Researchers from JD.com, a Chinese online retailer, the University of California, Santa Barbara, and the University of California, Riverside, have developed an AI system that can detect deepfake images—images realistically altered by deep learning software—with up to 95 percent accuracy. The researchers trained the system on thousands of deepfake and authentic images, teaching the system to spot manipulations at the single-pixel level. The researchers found that boundaries of objects in deepfake images are often smoother than in natural images.
Researchers from Tsinghua University in China have developed a hybrid AI chip, which they have shown can enable autonomous bicycles. The chip combines traditional computer science designs and architectures that attempt to mimic functions of the brain. This combination enables the chip, which runs neural networks, to help a bicycle balance, steer itself, detect objects, and respond to voice commands.
7. Detecting Esophageal Cancer
Researchers from the Indian Institute of Technology have developed an AI system that can act as a pre-screening tool for esophageal cancer. The researchers developed the system using data of 3,000 people, including clinical data, data on tobacco use, cancer history in the family, and data concerning difficulty swallowing. The system was up to 99 percent accurate and had an almost zero percent false-negative rate for predicting if an individual has esophageal cancer.
Researchers from the University of Colorado and Duke University have developed an AI system that can accurately classify certain human emotions in images. The researchers trained and tested their system on a combined 2,185 videos that displayed 27 distinct emotions. The system classifies emotions such as horror and sexual desire accurately but struggles in classifying confusion, awe, and surprise.
9. Creating a Wearable for Plants
Researchers from the University of Nebraska and Iowa State University have designed a fitness tracker-like bracelet that can measure corn and thick-stemmed crops’ sap flow, which indicates how much water a plant is using and conserving. The bracelet uses a micro-heater to apply small amounts of heat to a plant’s stalk while sensors above and below the heater measure how fast the sap is carrying heat away. This measurement indicates the speed of the sap flow which can help researchers better understand how different strains of crops respond to drought conditions and lead to the development of more drought-resistant crops.
10. Automatically Annotating Radiology Reports for Cancer Outcomes
Researchers from the Dana-Farber Cancer Institute in Boston have developed an AI tool that can automatically label radiology reports by the cancer outcome. Data concerning how tumors respond to treatments are typically only available as unstructured text in reports, making it difficult for researchers to use the data in computational analysis. The researchers trained their tool using reports concerning over 1,000 patients, finding it could accurately identify treatment outcomes, such as disease-free survival. In addition, the tool can annotate 30,000 reports with the correct treatment outcome in ten minutes. The same task would take a human reviewer six months.
Image: Dwight Sipler