Yonsei University

A new deep learning model-based tool developed in South Korea can be used to both diagnose coronary artery disease and predict major adverse cardiac events in emergency cases. HOW IT WORKS Developed by a team of researchers from Yonsei University Severance Hospital, Keimyung University Dongsan Hospital, and medical imaging AI company Phantomics, the AI tool automatically assesses coronary CT angiography (CCTA) scans and classifies stenosis as normal, non-occlusive, or occlusive. The model also utilises the YOLO architecture, which simultaneously locates and classifies objects, to quickly process images. FINDINGS In a study, the AI model was tested using CCTA data from 408 patients who presented with acute chest pain at three emergency departments from 2018 to 2022. Findings, which were published in the Radiology: Artificial Intelligence journal of the Radiology Society of North America, noted that the deep learning-driven analysis of the degree of stenosis was a better predictor of major adverse cardiac events (MACEs) than common clinical risk factors such as hyperlipidemia or the cardiac enzyme level troponin-T. Moreover, pairing the AI-driven analysis with common risk factors improved the MACE prediction by 14% points to 90%. WHY IT MATTERS CT angiography, which is used to assess artery stenosis for CAD prognosis, usually takes a long time to process results, with analyses varying depending on the reader, according to Severance Hospital. The AI tool developed by the Korean research team not only detects CAD but also predicts MACE risks in patients who present to the emergency department. "This study suggests the possibility that deep learning models can be applied to predict patient prognosis beyond simply determining the presence or absence of CAD in emergency rooms, where rapid diagnosis and treatment decisions are important," said Dr Jin Hur, professor at Severance Hospital's Department of Radiology. "The AI technology can be applied beyond simple diagnostic assistance to become a clinical decision support tool," he added. MARKET SNAPSHOT Latest research projects across Asia-Pacific have also utilised AI to improve CAD diagnosis. Singaporean startup Health BETA is developing a solution that considers genetic and lifestyle factors in providing an enhanced polygenic risk score for CAD. Meanwhile, three major heart hospitals in Singapore – the National Heart Centre Singapore, the National University Hospital and Tan Tock Seng Hospital – are set to pilot a new machine learning-driven system for rapid CAD prediction. In Australia, publicly listed medical device companies Echo IQ and Artrya have recently obtained 510(k) clearance from the United States Food and Drug Administration for their respective AI-powered software for diagnosing CAD. Echo IQ's product is specifically indicated for detecting severe aortic stenosis, while Artrya's AI-powered software, Salix, delivers a 10-minute point-of-care assessment of CCTA scans.

A consortium led by South Korea's top universities, Yonsei University and Seoul National University, will embark on a national project promoting the early detection of autism spectrum disorder in the country. Backed by the National Center for Mental Health and the Ministry of Health and Welfare, the project received 9.2 billion won ($6.6 million) in government funding. According to a media release, the project aims to develop AI-powered digital and medical devices for screening autism, which are targeted for release by 2028. The devices will be run by algorithms based on data from a cohort of 1,200 infants and toddlers below 48 months. Additionally, applications for an innovative medical device designation from the Ministry of Food and Drug Safety will be pursued. WHY IT MATTERS A recent article noted an overall increase in cases of ASD among Korean children and adolescents from 2011 to 2021. It has been estimated that the prevalence of DSM-5-defined ASD is at 2.20% . This trend may necessitate calibrating strategies to enable timelier and more effective prevention and care of children with autism. However, current approaches in ASD diagnosis, according to the Yonsei-SNU-led consortium, are largely dependent on specialists' findings and subjective observations of guardians, which often render them inconsistent and short of objectivity. "The AI-based [ASD] screening assistive medical device to be developed through this project will be an opportunity to shift the paradigm of early autism diagnosis and treatment," explained Chun Geun-ah, professor at Yonsei University Severance Hospital and principal project investigator. "The consortium will establish a foundation to overcome the limitations of diagnosing [ASD] and lead to better treatment outcomes," Kim Boong-nyeon, SNU Hospital (SNUH) professor and co-principal investigator, also said. The consortium involves several major hospitals, such as Yonsei's Gangnam Severance Hospital, Ewha Womans University Hospital in Seoul and Mokdong, and Seoul St. Mary’s Hospital, and digital health companies, including HurayPositive and Adotcure. THE LARGER TREND In 2023, SNUH opened a living laboratory to collect data that will be used to develop algorithm-driven models for detecting and personalising the treatment of ASD. The lab, also supported by the NCMH, serves as a bedrock for new digital therapeutics and the discovery of autism biomarkers. AI has seen increasing applications in ASD screening and treatment. EarliTec Diagnostics in the United States is a major developer in this space. It received regulatory approval for its autism decision support tool in 2022. Meanwhile, in Japan, new research utilised eye-tracking technology to demonstrate how predictable movement stimuli can be potentially used as a behavioural marker for autism .

A new study from South Korea has demonstrated the use of AI models to possibly diagnose and stratify attention deficit hyperactivity disorder by analysing eye images. FINDINGS Researchers from Yonsei University Health System have used four machine learning models and the AutoMorph deep learning pipeline to analyse approximately 1,108 retinal fundus photographs from over 600 children with ADHD and children with typical development. Based on findings published in npj Digital Medicine, all four models showed high accuracy, reaching up to 96.9%. These AI models also showed high performance of up to 87.3% in predicting the degree of impairment in visual selective attention, an ability that ADHD patients are usually challenged with. The study also identified representative symptoms of ADHD by deriving key retinal features via a Shapley Additive Explanations analysis. These include increased vascular density, decreased arterial vessel width, and changes in the optic disc structure. WHY IT MATTERS ADHD is a neurodevelopmental disorder that is difficult to diagnose quickly, says the Yonsei University researchers. Challenges in diagnosis include patients' high subjectivity, the variability of symptoms among individuals, and overlapping symptoms with other existing conditions. Considering the established role of dopamine in retinal function and ADHD symptoms, the researchers looked at retinal images as a potential ADHD screening biomarker in their study. Eventually, they proved the potential of retinal images as an ADHD biomarker and rapid screening tool. "Fundus examinations are very simple, taking less than five minutes. It seems these can be used as a rapid test to monitor the effectiveness of ADHD treatments," said Keun-ah Cheon, professor of Paediatrics at Severance Hospital and co-research lead. THE LARGER TREND Most digital innovations that came out over the past years have only complemented ADHD management, like a mobile application developed at Flinders University in Australia and a digital therapeutics-based program by Singapore's Institute of Mental Health and local startup Neeuro. Meanwhile, eye and eye image-based screening technologies, augmented by AI, have also been applied to diagnose or predict other neurodevelopmental conditions, particularly autism spectrum disorder. Recently, a researcher from Waseda University in Japan demonstrated the use of eye-tracking technology to test children's responses to predictable movement stimuli, which was found to have potential as a behavioural marker for early autism diagnosis.