Pancreatic cancer has one of the lowest survival rates of any cancer, with most patients surviving less than a year after diagnosis. Accurately predicting how long a patient is likely to survive — and what factors most influence that — can help doctors choose the right treatment, counsel patients, and prioritize clinical trial enrollment.
Traditional survival analysis methods rely heavily on staging and a few clinical variables. Machine learning models can incorporate a much wider range of data including imaging features, genomics, lab results, and treatment history, potentially offering more precise and individualized predictions.
This paper reviewed the published literature on machine learning methods applied to pancreatic cancer survival prediction. The survey covered studies using a range of algorithms including decision trees, random forests, support vector machines, neural networks, and deep learning models.
Studies were evaluated based on the data types used — clinical, imaging, genomic, or combinations — and the performance metrics reported. The authors assessed how different approaches compared on tasks like predicting one-year survival, overall survival time, and resection outcomes.
A key observation across the literature was that ensemble methods and deep learning models tended to outperform simpler approaches, especially when richer multi-modal data was available. However, small dataset sizes remained a persistent limitation across many studies.
Across studies reviewed, the most commonly used data types were clinical variables such as tumor stage, surgical margins, and CA 19-9 levels. Imaging-based features from CT scans and genomic data were used less frequently but tended to add predictive value when combined with clinical data.
Models that combined multiple data types — sometimes called multi-modal models — consistently showed better performance than single-data-type models. This suggests that survival in pancreatic cancer is driven by a complex interaction of biological, anatomical, and clinical factors.
Despite progress, a major limitation identified in the review was the lack of external validation. Most models were developed and tested on the same dataset, raising questions about how well they would generalize to new patient populations.
One recurring challenge is data scarcity. Pancreatic cancer is relatively uncommon, and datasets with complete multi-modal information are rare. This limits the complexity of models that can be reliably trained and makes it difficult to achieve statistical power for subgroup analyses.
Interpretability is another concern. Many high-performing models such as deep neural networks are difficult to explain to clinicians. For survival prediction tools to be adopted in practice, they need to provide not just a prediction but also some explanation of which factors are driving it.
The survey concludes that while machine learning has made meaningful progress in pancreatic cancer survival prediction, the field needs larger multicenter datasets, standardized benchmarks for model comparison, and more rigorous external validation before clinical deployment is feasible.
Future research should focus on explainable AI methods that can communicate predictions to clinicians in interpretable terms, and on prospective studies that test whether AI-assisted survival prediction actually improves clinical decision-making and patient outcomes.