Abstract PR006: A time-to-event model for pancreatic cancer risk prediction across multiple future intervals using longitudinal structured electronic health records
Jifan Gao, Mina Stanikic, Eshna Lahr, Zhiwei Liang, Gargi Porwal, Chen Yuan, Vidya Madineedi, Leigh Culnane, Lauren Brais, Vasilena Gocheva, Jonathan Nowak, Brian Wolpin, Simona Cristea, Michael Rosenthal, William LotterAbstract
Introduction:
To facilitate risk assessment for pancreatic ductal adenocarcinoma (PDAC) in the general population, we developed time-to-event models using longitudinal structured electronic health record (EHR) data to estimate 6–12, 12–24, and 24–36 months PDAC risk.
Methods:
The model was developed using UK Biobank (UKB), a large general-population cohort in the United Kingdom, and was externally validated using an independent general-population cohort from Mass General Brigham/Dana-Farber Cancer Institute (MGB/DFCI), a separate healthcare system in the United States. We used a deep learning framework consisting of a transformer neural network adapted to longitudinal EHR sequences, and implemented censoring-aware discrete-time survival learning to jointly estimate PDAC risk across multiple future intervals. For cases, candidate index dates were selected from outpatient visits within 6–12, 12–24, and 24–36 months intervals before the first documented PDAC diagnosis, with a 6-month exclusion window prior to the first PDAC diagnosis to reduce potential information leakage from the peri-diagnostic period. For controls without PDAC, index dates were selected from outpatient visits using the same requirement. We limited the cohort to patients aged 45–90 years at the index date. Model inputs included ICD diagnosis codes and medication records occurring on or before the index date. Each eligible index date was required to have at least five prior ICD codes. Using UKB data, the model was first pretrained to predict coarse ICD-10 groups, enabling it to learn general medical representations, and was then fine-tuned with a discrete-time survival objective to learn PDAC-specific risk patterns across future intervals. AI tools were used to assist with code organization and text editing.
Results:
After applying eligibility criteria, the UKB data included 463,744 patients, including 1,067 PDAC cases. The external MGB/DFCI data cohort comprised 860,465 patients, including 578 PDAC cases. In the external cohort, the overall C-index was 0.779 ± 0.021, with interval-specific AUROCs of 0.789 ± 0.027 for 6–12 months, 0.786 ± 0.033 for 12–24 months, and 0.747 ± 0.036 for 24–36 months. At 95% specificity, sensitivities were 0.263 ± 0.023, 0.184 ± 0.015, and 0.209 ± 0.019 for the three intervals, respectively. Performance decreased by less than 0.05 in external validation compared to internal cross-validation, suggesting promising generalizability to an independent healthcare system.
Conclusions:
Beyond general risk prediction or single cohort testing, the developed EHR transformer model uses discrete-time survival learning to jointly estimate PDAC interval-specific and cumulative risk, with promising generalization performance validated in an independent cohort. By identifying elevated risk across multiple risk intervals while excluding the peri-diagnostic period, this approach may serve as a first-stage, low-cost clinical stratification tool to identify high-risk individuals before more resource-intensive downstream screening such as pancreas-directed imaging.
Citation Format:
Jifan Gao, Mina Stanikic, Eshna Lahr, Zhiwei Liang, Gargi Porwal, Chen Yuan, Vidya Madineedi, Leigh Culnane, Lauren Brais, Vasilena Gocheva, Jonathan Nowak, Brian Wolpin, Simona Cristea, Michael Rosenthal, William Lotter. A time-to-event model for pancreatic cancer risk prediction across multiple future intervals using longitudinal structured electronic health records [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 PR006.