A new AI model trained on 165,391 emergency department records from Shuang-Ho Hospital achieves an AUROC of 0.88, effectively identifying patients unlikely to need CT imaging and outperforming traditional methods and biomedical language models.

At the core is a clinical language engineering pipeline that normalizes language, standardizes terminology, and learns semantic representations from noisy narrative data to extract meaningful diagnostic signals.

Supported by Taiwan’s National Science and Technology Council, the work appears in Engineering Applications of Artificial Intelligence (2024), recognized for its impact on AI-driven healthcare innovation.

Led by Professor Yung-Chun Chang of Taipei Medical University and Dr. Ting-Yun Huang of Shuang-Ho Hospital, the study analyzes real-world, multilingual, and unstructured emergency department notes.

The project highlights interdisciplinary collaboration between AI engineers and frontline clinicians and demonstrates the potential of large language models to optimize healthcare systems in multilingual, unstructured data settings.

Designed for real-world deployment, the model runs on standard hospital computing infrastructure and can be integrated with electronic health records to provide real-time decision support.

Unlike traditional models relying on structured data, this system uses unstructured clinical narratives—chief complaints, symptom descriptions, medical history, and pain severity—as primary inputs to capture subtle patterns.

The original article is titled Enhancing the Effectiveness of Emergency Department Computed Tomography Scans Using Pre-Trained Language Models, with a profile of Yung-Chun Chang.

TMU researchers developed an AI-driven prediction model to help emergency departments decide when CT scans are necessary, aiming to speed decisions, reduce unnecessary imaging, and optimize resources.