Placido · Nature medicine 2023 · Retrospective prognostic cohort and deep learning validation study · n=9000000

A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories.

Cited 338 times in the scientific literature.

Level 3 - non-randomized controlled study

Retrospective cohort study developing and validating a prognostic algorithm across clinical registries.

PubMed 37156936 · doi:10.1038/s41591-023-02332-5 · record verified 2026-08-28

What was done

Researchers developed and validated a deep learning model (CancerRiskNet) to predict pancreatic cancer risk using sequential disease codes from electronic health records. The model was trained and evaluated on clinical data from 6 million patients (24,000 pancreatic cancer cases) in the Danish National Patient Registry (DNPR) and tested/retrained on 3 million patients (3,900 cases) from the US Veterans Affairs (US-VA) health system across incremental prediction time windows.

What was found

For predicting pancreatic cancer within 36 months, the best DNPR model achieved an AUROC of 0.88, which dropped to 0.83 when excluding clinical events within 3 months prior to diagnosis. Among the 1,000 highest-risk individuals over age 50, the estimated relative risk was 59. Direct application of the Danish model to the US-VA cohort yielded an AUROC of 0.71; retraining the model on US-VA data improved performance to an AUROC of 0.78 (0.76 when excluding the final 3 months of data).

Why it matters

Early detection of pancreatic cancer is critical for improving survival, but broad population screening is impractical due to low overall incidence. Using existing medical coding trajectories allows automated risk stratification to identify high-risk subsets of patients who may benefit from targeted surveillance programs.

Limits

The algorithm relies strictly on structured diagnostic codes, which may contain misclassifications or missing entries. The model demonstrated reduced generalizability when transferred directly between different healthcare systems (AUROC dropped from 0.88 to 0.71), necessitating cohort-specific retraining. Clinical utility and cost-effectiveness in prospective screening workflows remain unproven.

Cited by