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

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Pancreatic cancer presents late, leading to poor patient outcomes.
• Early detection is crucial for improving survival and quality of life.
• Current diagnostic methods have limitations in identifying at-risk individuals proactively.

Purpose of the Study:

• To develop and validate artificial intelligence (AI) models for predicting pancreatic cancer risk.
• To assess the performance of AI models using large-scale clinical datasets from Denmark and the US.
• To evaluate the feasibility of implementing AI-driven surveillance programs for early pancreatic cancer detection.

Main Methods:

• Machine learning models were trained on sequential disease codes from electronic health records.
• Data from the Danish National Patient Registry (DNPR) and US Veterans Affairs (US-VA) databases were utilized.
• Model performance was evaluated using the area under the receiver operating characteristic (AUROC) curve for predicting cancer occurrence within various time windows.

Main Results:

• The best model trained on Danish data achieved an AUROC of 0.88 for predicting cancer within 36 months.
• Excluding recent disease events (3 months prior) reduced performance to AUROC 0.83.
• The model identified a relative risk of 59 for the highest-risk group (over 50 years old).
• Cross-application to US data showed lower AUROC (0.71), necessitating model retraining for improved performance (AUROC 0.78).

Conclusions:

• AI models demonstrate significant potential for early pancreatic cancer risk prediction.
• These models can inform the design of targeted surveillance programs for high-risk populations.
• Early detection through AI can potentially enhance patient lifespan and quality of life.