Pancreatic cancer risk predicted from disease trajectories using deep learning

Abstract

Pancreatic cancer is an aggressive disease that typically presents late with poor patient outcomes. There is a pronounced medical need for early detection of pancreatic cancer, which can be facilitated by identifying high-risk populations. Here we apply artificial intelligence (AI) methods to a large corpus of more than 6 million patient records spanning 40 years with 24,000 pancreatic cancer cases in the Danish National Patient Registry. In contrast to existing methods that do not use temporal information, we explicitly train machine learning models on the time sequence of diseases in patient clinical histories. In addition, the models predict the risk of cancer occurrence in time intervals of 3 to 60 months duration after risk assessment. For cancer occurrence within 12 months, the performance of the best model trained on full trajectories (AUROC=0.91) substantially exceeds that of a model without time information (AUROC=0.81). For the best model, lower performance (AUROC=0.86) results when disease events within a 3 month window before cancer diagnosis are excluded from training, reflecting the decreasing information value of earlier disease events. These results raise the state-of-the-art level of performance of cancer risk prediction on real-world data sets and provide support for the design of real-world population-wide clinical screening trials, in which high risk patients are assigned to serial imaging and measurement of blood-based markers to facilitate earlier cancer detection. AI on real-world clinical records has the potential to shift focus from treatment of late- to early-stage cancer, benefiting patients by improving lifespan and quality of life.

Developing and validating a pancreatic cancer risk model for the general population using multi-institutional electronic health records from a federated network

“…Other researchers have used EHR data to develop PDAC risk prediction models for the general population [3,5,7,8,22]. Data set sizes range from 1,792 PDAC cases/1.8M controls [8] to 24,000 PDAC cases/6.2M controls [22].…”

“…While some studies work with data obtained from multiple organizations [7,8,22], none work with a federated data network that harmonizes and standardizes the data, none provides a clear path to clinical deployment, and none supports the seamless deployment of the model to new HCOs as they join the federated network. Some previous studies evaluate the ability of their models to identify high-risk individuals either until or shortly before the date of PDAC diagnosis [7,8,22], when clinical benefit is improbable. To focus on time frames in which detection of early stage disease and potential cure are most likely, we evaluate the ability of our models to identify high-risk patients at least six months before diagnosis.…”

Early Detection of Pancreatic Cancer

“…Only a handful of research studies have used ML to build predictive models with EHR data in this field. 11 , 15 , 16 These studies have demonstrated that by leveraging AI/ML and EHRs, subpopulations at high risk for PDAC can be identified 1 to 2 years before diagnosis. Such efforts also highlight specific challenges and opportunities for improving the secondary use of EHR data with AI and innovative data science solutions.…”

Structured deep embedding model to generate composite clinical indices from electronic health records for early detection of pancreatic cancer

“…The AUROC from cross-application of the model on an external dataset, however, decreased to 0.78, which addresses the limitation of model generalizability, likely due to different coding practices across different health systems. 15 The Med-BERT, a contextualized embedding model pre-trained on a structured EHR dataset of 28,490,650 patients, has shown some promise for establishing a generalizable AI model for medical/clinical applications. Med-BERT enables utilization of small local training datasets for realistic disease prediction tasks.…”