Related Experiment Video
Updated: Jul 8, 2026
Testing Targeted Therapies in Cancer using Structural DNA Alteration Analysis and Patient-Derived Xenografts
Published on: July 25, 2020
When survival models fail: An interpretable anomaly-detection approach for high-risk phenotypes in resected solid
Claudio Ricci1, Laura Alberici2, Vincenzo D'Ambra1
1Department of Internal Medicine and Surgery (DIMEC), Alma Mater Studiorum, University of Bologna, Bologna, Italy.
Background:
Solid pseudopapillary tumor of the pancreas is a rare neoplasm with an excellent prognosis after resection. Conventional survival models are unreliable because of the low number of cancer-specific deaths. This study aimed to identify clinicopathologic patterns associated with the "high-risk phenotype" through an interpretable anomaly-detection framework.
Methods:
Patients undergoing pancreatic resection for solid pseudopapillary tumors were identified from the Surveillance, Epidemiology, and End Results registry (2000-2021). Cancer-specific death was modeled as an ultrarare event using 2 unsupervised algorithms: Isolation Forest and Local Outlier Factor. Group differences in anomaly scores were assessed using the Mann-Whitney U test with Cliff's δ as effect size. Feature relevance was explored using a Random Forest surrogate model with SHapley Additive exPlanations interpretation. The final calculator was externally validated on a cohort reconstructed through a systematic review.
Results:
In the Surveillance, Epidemiology, and End Results cohort (n = 387), patients who died from solid pseudopapillary tumors showed significantly higher anomaly scores (Isolation Forest = -0.07 ± 0.06 vs 0.02 ± 0.05; P = .007; δ = 0.783). The most influential contributors to the death-like anomaly pattern were lymph node ratio (0.204 ± 0.083), hospital type (0.168 ± 0.037), male sex (0.130 ± 0.028), atypical resection (0.120 ± 0.028), and stage IV disease (0.096 ± 0.047). The model achieved an in-sample area under the receiver operating characteristic curve of 0.892, a cross-validation area under the receiver operating characteristic curve of 0.733 ± 0.351, and an external-validation area under the receiver operating characteristic curve of 0.975.
Conclusion:
An interpretable anomaly-detection approach can identify hidden high-risk phenotypes within apparently indolent diseases. In resected pancreatic solid pseudopapillary tumors, elevated lymph node ratio, treatment in nonmetropolitan settings, and male sex define the "high-risk phenotype."
