Deep-Transfer-Learning-Based Natural Language Processing of Serial Free-Text Computed Tomography Reports for...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Pancreatic cancer survival prediction remains challenging.
• Radiology reports contain rich, unstructured data.
• Natural Language Processing (NLP) offers potential for extracting prognostic information from text.

Purpose of the Study:

• To evaluate the predictive capability of serial computed tomography (CT) radiology reports for pancreatic cancer survival.
• To develop and validate a deep-transfer-learning-based NLP model for this purpose.

Main Methods:

• Retrospective training and testing of NLP models using serial, free-text CT reports from a Korean tertiary hospital.
• Extraction of survival data for patients diagnosed with pancreatic cancer.
• External validation using data from an independent US tertiary hospital.
• Calculation of concordance index (c-index) and area under the receiver operating characteristic curve (AUROC).

Main Results:

• A ClinicalBERT model trained on serial CT reports achieved a c-index of 0.811 and AUROC of 0.911 for predicting overall survival.
• The model demonstrated generalizability with an AUROC of 0.888 on an external testing set.
• The NLP model showed contextual interpretation capabilities beyond specific phrases.

Conclusions:

• Deep-transfer-learning-based NLP models utilizing serial CT reports can effectively predict pancreatic cancer patient survival.
• The developed model can support clinical decision-making by extracting prognostic information directly from radiology reports.