Pancreatic ductal adenocarcinoma (PDAC) has a 5-year survival rate below 12%, largely because most cases are diagnosed at late stages when the disease has already spread beyond the pancreas. Patients diagnosed at early stages, when the tumor is still surgically resectable, have dramatically better outcomes.
The challenge is that PDAC typically causes no specific symptoms until it is advanced. Early-stage disease may produce subtle, non-specific signs that appear in the clinical record months or years before diagnosis, including changes in laboratory values, new-onset diabetes, weight loss, and biliary symptoms.
Electronic Health Records (EHR) capture longitudinal clinical data including diagnoses, medications, laboratory results, and procedure records across time. Machine learning applied to EHR data offers the opportunity to detect subtle pre-diagnostic patterns that might signal early pancreatic cancer before symptoms become overt.
Individual institutions lack sufficient patients with rare early-stage PDAC to train robust predictive models. EHR data sharing networks such as TriNetX, PCORnet, and i2b2 aggregate de-identified patient data from multiple hospitals, dramatically increasing the available sample sizes for research.
A critical enabler for multi-site research is the use of Common Data Models (CDMs) such as OMOP (Observational Medical Outcomes Partnership). CDMs standardize how clinical concepts are coded across different hospital systems, so that a diagnosis of diabetes from one hospital's EHR maps to the same code as the same diagnosis from another hospital.
Without CDM harmonization, researchers would need to write separate data extraction code for each institution's proprietary format. CDM adoption allows AI models trained on data from one network to be more readily applied at new sites, accelerating the path from research findings to clinical deployment.
Supervised learning models for PDAC risk prediction are typically trained to distinguish patients who will develop pancreatic cancer within a specified time window from matched controls who do not. This case-control design using historical EHR data allows model training without prospective data collection.
Natural Language Processing (NLP) is essential for extracting clinically relevant information from free-text components of the EHR such as physician notes, radiology reports, and pathology records. Structured EHR data captures diagnoses and labs, but much of the clinically relevant information exists only in unstructured text.
Common machine learning approaches include gradient boosting methods like XGBoost, which handle the mixed data types and sparse feature spaces typical of EHR data, as well as recurrent neural networks that can model the temporal sequence of clinical events leading up to a potential cancer diagnosis.
Federated learning enables AI models to be trained across multiple hospitals without sharing raw patient data. Each participating site trains the model locally on its own patients, and only the resulting model parameters (not patient records) are shared with a central coordinating server.
The central server aggregates parameter updates from all participating sites into an improved global model, which is then distributed back to each site for the next training round. This iterative process continues until the global model converges, effectively learning from all patients across all sites without any individual hospital ever exposing its patient data.
Federated learning addresses a fundamental tension in AI-driven healthcare: the need for large diverse training datasets to build robust models versus strict patient privacy regulations such as HIPAA in the United States and GDPR in Europe. Federated approaches make multi-site collaboration both technically feasible and legally compliant.
Data quality is a persistent challenge for EHR-based AI. Missing data, inconsistent coding practices, and documentation bias mean that EHR data reflects what clinicians recorded rather than the complete clinical picture. Models trained on data from one health system may fail at another where documentation practices differ.
The AI chasm refers to the gap between research validation and real-world clinical deployment. A model that achieves excellent AUC in a retrospective study may underperform in prospective deployment due to distribution shift, physician workflow disruption, or alert fatigue from too many false positive notifications.
Incentive alignment is also a barrier. Detecting PDAC earlier requires action (imaging, follow-up, potential intervention) that generates cost. Healthcare systems need aligned financial and clinical incentives that reward early detection to motivate widespread adoption of AI screening programs.
Despite these challenges, EHR-based AI for PDAC risk stratification holds significant promise. Even modest improvements in detecting cases months earlier could translate to meaningful survival benefits given the steep relationship between stage at diagnosis and survival outcomes in pancreatic cancer.