Abstract A095: Quantifying the burden of undiagnosed, untreated pancreatic cancer facilitated through automated computational review of abdominal imaging
Chase A. Shipp, Siva P. Madala, Adrianna Kapusta, Christine Molmenti, Daniel KingAbstract
Background:
Pancreatic masses on radiologic reports are often overlooked, leaving a significant population undiagnosed and untreated for pancreatic cancer. We trained a natural language processing (NLP) model to identify suspected pancreatic malignancies and quantified and characterized a patient population with pancreatic masses who are not captured by registries or undergo appropriate oncologic follow up.
Methods:
The NLP model reviewed all abdominal CT and MRI imaging reports in the Northwell health system between April 27 and July 21, 2025 and flagged pancreatic masses. Chart review was conducted to identify patients presenting with newly radiographically suspected disease. Biopsy, treatment, sociodemographic, and clinical data were abstracted from the medical record.
Results:
The NLP model identified 359 reports with pancreatic masses, among which 96 (27%) had an initial presentation of a pancreatic mass. Of these, 38 (40%) did not receive a biopsy. Of those who did, most had adenocarcinoma (PDAC) or poorly differentiated carcinoma (79%). Among those with PDAC or carcinoma, 67% received treatment with surgery or chemotherapy. Median time from radiology report to biopsy and radiology report to treatment was 4 days (IQR 2 - 9) and 32 days (IQR 27 - 46), respectively. Among the 38 patients with radiographically-suspected cancer who did not receive a biopsy, reasons for lack of biopsy were as follows: lost to follow-up (29%), mass assumed to be benign (26%), patient preference (16%), mass assumed to be different primary (13%), underlying disease or frailty (11%), biopsy difficulty (3%), and disease progression or death (3%). Among the 15 patients with diagnosed PDAC or carcinoma who did not receive treatment, the reasons for lack of treatment were as follows: patient preference (47%), underlying disease or frailty (27%), disease progression or death (13%), and lost to follow-up (13%).
Conclusions:
The NLP model uncovered patients with radiographically suspected pancreatic cancer, a heretofore poorly characterized population not captured in registries, and found that the majority (53 of 96) never received treatment for PDAC. The reasons for lack of biopsy or treatment should be further studied to decrease time to treatment in pancreatic cancer. In particular, future studies should explore whether the advent of targeted RAS inhibitors will impact the proportion of patients declining biopsy or treatment due to patient preference for avoiding further workup or therapy. Ultimately, NLP may be a promising approach to study epidemiology and referral patterns of suspected or recently diagnosed pancreatic cancer or other malignancies with incipient radiologic findings.
Citation Format:
Chase A. Shipp, Siva P. Madala, Adrianna Kapusta, Christine Molmenti, Daniel King. Quantifying the burden of undiagnosed, untreated pancreatic cancer facilitated through automated computational review of abdominal imaging [abstract]. In: Proceedings of the AACR Conference on Pancreatic Cancer: New Frontiers in Biology and Therapeutic Development; 2026 Sep 25-28; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(18_Suppl_2):Abstract nr A095.