EUS-Based Artificial Intelligence to Predict Outcomes in Pancreatic Cancer

An Observational Study on the Application of Artificial Intelligence Model to Predict Diagnosis, Prognosis, and Molecular Alterations in Pancreatic Cancer

The EUS-AI-R study is an observational, single-center, two-phase (retrospective-prospective) study designed to develop and validate artificial intelligence (AI) models for predicting chemotherapy response and oncological outcomes in patients with pancreatic ductal adenocarcinoma (PDAC).

Patients who underwent endoscopic ultrasound (EUS) with tissue acquisition (EUS-FNA/FNB) for suspected pancreatic lesions at IRCCS San Raffaele Hospital between January 1st, 2019 and January 2026 will be retrospectively included. All patients have a histologically confirmed diagnosis of PDAC and a minimum follow-up of six months. These data derive from an IRB-approved institutional study (BIOPANCREAS; NCT06552078).

Retrospective multimodal data, including EUS imaging (B-mode, elastography, contrast-enhanced EUS), clinical and laboratory variables, CT/MRI imaging, digital pathology, and molecular data when available, will be used to develop and internally validate multiple AI models.

In the prospective phase, the best-performing AI model will be applied to an independent cohort of patients undergoing EUS at the same institution to evaluate feasibility, calibration, and real-world performance.

No additional procedures beyond standard clinical practice will be performed.

PDAC remains one of the leading causes of cancer-related mortality, largely due to late diagnosis and limited predictive tools for treatment response. EUS represents the most sensitive modality for detecting pancreatic lesions and allows tissue acquisition for histological confirmation.

Recent advances in AI, including machine learning (ML) and deep learning (DL), have demonstrated strong potential in improving diagnostic accuracy, prognostic stratification, and prediction of treatment response in oncology.

The EUS-AI-R study aims to integrate multimodal data, including EUS imaging, clinical variables, radiological imaging, digital histopathology, and molecular data, into AI-based predictive models capable of estimating chemotherapy response and survival outcomes in PDAC patients.

The study consists of two phases:

• Retrospective phase: development, training, and internal validation of AI models using approximately 500 patients from an institutional database.
• Prospective phase: application of the selected model to an independent cohort (200 patients) to assess feasibility, calibration, and real-world performance.

Multiple modality-specific models (EUS-based, clinical-based, radiology-based, pathology-based) and a multimodal integrated model will be developed and compared. Model performance will be evaluated using AUC-ROC, sensitivity, specificity, calibration, and concordance index for survival outcomes.

The study is observational and does not modify standard clinical practice.

Inclusion Criteria:

• Pathologically confirmed diagnosis (final pathology report) of pancreatic cancer obtained through endoscopic ultrasound-guided tissue sampling (EUS-FNA or EUS-FNB)
• Age ≥ 18 years at the time of diagnosis
• Minimum follow-up duration of 6 months after diagnosis
• Absence of other concomitant neoplastic diseases
• Age>= 18 years
• Capacity to understand and make informed decisions
• Written informed consent provided by the patient

Exclusion Criteria:

• All patients who underwent endoscopic ultrasound with tissue sampling at the Pancreato-Biliary Endoscopy and Endoscopic Ultrasound Unit of IRCCS San Raffaele Hospital but do not meet the inclusion criteria will be excluded from the final analysis cohorts.
• Age<18 years
• Inability to understand and make informed decisions
• Refusal to participate in the study