Does Artificial Intelligence Improve the Prediction of Acute Pancreatitis Complications?

Acute pancreatitis is a sudden inflammation of the pancreas that can rapidly progress to severe forms. Each year, it affects between 30 and 40 people per 100,000 inhabitants worldwide. Although most cases are mild and resolve spontaneously, nearly 30% of patients develop severe complications, such as persistent organ failure or infected pancreatic necrosis. These situations significantly increase morbidity and mortality, highlighting the importance of early and appropriate management.

Currently, traditional methods for assessing the severity of acute pancreatitis rely on clinical scores such as Ranson or Glasgow-Imrie. However, these tools have limitations. They are based on measurements taken within the first 48 hours, which delays the identification of at-risk patients. Additionally, their accuracy remains modest, with discrimination values around 0.61 for the Ranson score and 0.67 for the Glasgow-Imrie score. Radiologists, on the other hand, evaluate contrast-enhanced CT images, but their visual judgment achieves an average accuracy of 0.63, with a low sensitivity of 42.5%. This means they primarily detect severe cases when signs are obvious but struggle to identify patients whose condition will deteriorate later.

To overcome these limitations, an innovative approach combines artificial intelligence, medical image analysis, and clinical data. This method, called radiomics, extracts quantitative information from images to reveal patterns invisible to the naked eye. It captures features such as intensity, texture, or spatial arrangement of tissues. Another technique, deep learning, automatically identifies complex patterns in images without requiring predefined criteria. These two approaches transform medical imaging into a tool that is not only visual but also quantitative and predictive.

In this study, 284 patients with acute pancreatitis were analyzed. Among them, 140—nearly half—developed severe complications within 30 days of admission. AI models based solely on imaging outperformed traditional methods, with an accuracy of 0.77 for radiomics and 0.76 for deep learning. The integration of laboratory data, such as white blood cell count, C-reactive protein, or lactate dehydrogenase, further improved performance. The model combining radiomics and clinical data achieved an accuracy of 0.80, a notable improvement over other approaches.

The selected biological markers reflect key disease processes. For example, blood urea indicates dehydration and poor renal perfusion, while C-reactive protein signals intense inflammation. Low albumin is associated with an increased risk of organ failure. These markers complement the information derived from images, providing a more comprehensive view of the patient’s condition.

Radiologists, despite their expertise, show moderate disagreement among themselves when evaluating images. Their limited sensitivity reveals that some patients developing severe complications exhibit subtle abnormalities on images that are difficult to detect visually. Artificial intelligence, on the other hand, identifies these early signs by analyzing thousands of quantitative features.

This multimodal approach, combining imaging and clinical data, thus enables more reliable prediction of complications. It could help doctors identify high-risk patients earlier, tailor monitoring, and rapidly initiate appropriate treatments. However, these results remain preliminary and require validation in larger studies before widespread use in clinical practice.

Sources

About This Study

DOI: https://doi.org/10.1007/s00261-026-05588-w

Title: Multimodal AI for early prediction of adverse clinical outcomes in acute pancreatitis

Journal: Abdominal Radiology

Publisher: Springer Science and Business Media LLC

Authors: Ahmet Yasin Karkas; Yavuz B. Taktak; Burak Gultekin; Ziliang Hong; Halil Ertugrul Aktas; Deniz Seyithanoglu; Timurhan Cebeci; Alper Akin; Ece Elustu; Kerem Arisin; Ali Canturk; Mehmet Ilhan; Naci Senkal; Michael B. Wallace; Abraham F. Bezuidenhout; Frank H. Miller; Alpay Medetalibeyoglu; Mehmet Semih Cakir; Gorkem Durak; Ulas Bagci; Sukru Mehmet Erturk