Pancreatic cancer risk prediction using sequential models (www.cell.com)

Researchers trained sequential deep-learning models on large, real-world VA electronic health records to predict future pancreatic ductal adenocarcinoma (PDAC) risk using longitudinal diagnostic (ICD) and medication trajectories. The development set included 210,848 patients (18,126 with cancer) and the held-out evaluation set 987,693 patients (1,300 with cancer). Cancer cases had substantially denser histories (median ICD events 195 vs 91; prescriptions 73 vs 50), and a combined model that jointly used diagnosis and medication sequences—with temporal exclusion windows to avoid leakage—improved prediction over non-sequential baselines. Significance for AI/ML and clinical practice: the study demonstrates that sequential modeling of EHR time series (implemented in PyTorch, with interpretability via Captum and public code) can identify a high-risk PDAC bracket from routine care data, enabling targeted surveillance and hypothesis generation about drug associations. Top predictive diagnostic features included hypertension, diabetes, lipidemias and acute pancreatitis; medication signals included lisinopril, statins, diuretics and metformin. The work highlights technical implications—benefit of temporal sequence models and multimodal EHR inputs—and cautions: demographic performance disparities, confounding in medication associations, and time-to-diagnosis limitations require prospective validation before clinical deployment.