A deep-learning–based time-series analysis study capable of predicting acute exacerbations of Chronic Obstructive Pulmonary Disease (COPD) in advance has been conducted at Jeonbuk National University (JBNU) and is attracting attention. While existing machine learning models have shown limitations in capturing long-term time-series dependencies and in handling severe class imbalance, the proposed predictive framework improves these issues and increases potential for clinical application.

The research team led by Professor Jae-Hyeok Cho of the Department of Software Engineering, College of Engineering at JBNU announced that they presented the paper "Time-Series–Based Prediction of COPD Exacerbation Using Multivariate Clinical Data" at the 2026 ICOBAR-SMART Joint Conference held at BINUS University in Indonesia and received the Best Paper Award.

The study involved Professor Jae-Hyeok Cho of the JBNU Applied AI Laboratory, Dr. Hye-Shin Kim, PhD candidate Seoheon Yoo, and master's student Jaehong Kim. The conference was an international academic event jointly organized by the Korean Institute of Information Technology (KIIT) and BINUS University, serving as a forum to share the latest research results in information technology.

The research team extracted a cohort of COPD patients from the MIMIC-IV intensive care clinical database established by MIT in the United States, and developed a Transformer-architecture time-series prediction model based on eight multivariate vital signs such as heart rate, blood pressure, oxygen saturation, respiratory rate, and body temperature.

In particular, they applied a patching mechanism that divides continuous physiological signals into fixed-length segments to effectively learn local contextual information, and introduced a Balanced Random Undersampling technique to address data imbalance, in which exacerbation events accounted for only about 1.1% of the dataset.

In experiments, the model achieved an AUROC of 0.7502 and a recall of 86.27% for predicting acute exacerbation one hour in advance. This indicates the model could detect approximately 86% of actual exacerbation events in advance, suggesting potential use as a clinical decision-support system to enable proactive medical interventions such as intensified bronchodilator administration, antibiotic prescription when infection is suspected, and adjustment of oxygen therapy.

Professor Jae-Hyeok Cho stated, 'This study is significant in that it has laid the groundwork for a clinical early warning system capable of predicting acute exacerbations of ICU COPD patients in advance to enable timely intervention,' and added, 'We plan to expand the work with multicenter external validation and long-term prediction over 6–12 hour intervals, and to enhance real-world applicability through prospective clinical evaluation.'

Meanwhile, this research was carried out with support from the BK21 Four program.