AI That Reads the ER: TMU Model Helps Doctors Make Smarter CT Scan Decisions

In busy emergency departments, every minute and every medical resource matters. An AI-driven prediction model developed by Taipei Medical University may help physicians determine more efficiently whether patients need CT scans, supporting faster decision-making, reducing unnecessary imaging, and improving hospital resource use.

The model learns from real-world emergency department records, including mixed languages notes, abbreviations, and fragmented symptom descriptions. By turning unstructured medical language into practical decision support, the system demonstrates how artificial intelligence can assist frontline clinicians in complex and fast-paced care settings.

The study was led by Professor Yung-Chun Chang from the Graduate Institute of Data Science at Taipei Medical University (TMU), in collaboration with Dr. Ting-Yun Huang from the Emergency Department of Shuang-Ho Hospital.

Addressing Emergency Department Overcrowding

Emergency department overcrowding remains a critical global healthcare challenge affecting treatment efficiency, patient flow, and medical safety. Computed tomography (CT) imaging is essential for diagnosing serious and potentially life-threatening conditions, but previous studies have estimated that 20–40% of CT scans may be unnecessary. Overuse of CT imaging can increase healthcare costs, place pressure on hospital resources, and expose patients to avoidable radiation.

To address this issue, the research team developed an AI-based predictive system designed to support early clinical decision-making and optimize medical resource allocation.

Turning Clinical Language into AI Decision Support

A key innovation of this study lies in its ability to tackle the complexities of real-world emergency departments. Clinical records in Taiwanese hospitals often include mixed Chinese–English terminology, domain-specific abbreviations, and institution-specific expressions. In emergency settings, these notes are often brief, fragmented, and highly unstructured due to time constraints and urgent clinical workflows.

To overcome these challenges, the research team developed a comprehensive clinical language engineering pipeline capable of processing multilingual and highly heterogeneous medical text. The system performs automated language normalization, terminology standardization, correction of inconsistent clinical expressions, and semantic representation learning from noisy narrative data

This enables the model to extract meaningful diagnostic information from incomplete or irregular documentation and improves its robustness in real-world clinical settings.

The proposed framework integrates large-scale clinical text processing, multimodal feature fusion, prompt-based feature representation, and data augmentation using large language models. Unlike conventional prediction methods that rely mainly on structured numerical data or laboratory results, the system leverages unstructured clinical narrative text—such as chief complaints, symptom descriptions, medical history, and pain severity—as the primary source of predictive information. This design allows the model to capture subtle clinical patterns and contextual relationships often overlooked by traditional machine learning approaches.

Using 165,391 emergency department records from Shuang-Ho Hospital, the model demonstrated strong predictive performance, achieving an AUROC of 0.88. It showed high reliability in identifying patients who were unlikely to require CT imaging and outperformed traditional machine learning methods as well as existing biomedical language models.

The system was also developed with practical clinical implementation in mind. After training, the model can operate on standard hospital computing infrastructure and has the potential to be integrated with electronic health record systems, enabling real-time clinical decision support in diverse healthcare environments.

Advancing AI-Enabled Healthcare Innovation

This research highlights the importance of interdisciplinary collaboration between AI engineers and frontline clinicians. It also demonstrates the transformative potential of large language models in healthcare system optimization.

By addressing the complexities of multilingual and unstructured clinical data, the study provides a foundation for future AI-driven medical decision support systems. The research was supported by Taiwan’s National Science and Technology Council and published in Engineering Applications of Artificial Intelligence, an international journal with a 2024 impact factor of 8.0 and a ranking in the top 2.5% of the Engineering, Multidisciplinary category. These achievements reflect TMU’s growing impact in AI-enabled healthcare innovation and its commitment to advancing precision medicine and intelligent healthcare.

---

Look for More Information

• Original Article: Enhancing the Effectiveness of Emergency Department Computed Tomography Scans Using Pre-Trained Language Models
• Author Profile: Yung-Chun Chang, Professor, Graduate Institute of Data Science, College of Management