A novel approach for selecting combination clinical markers of pathology applied to a large retrospective cohort of surgically resected pancreatic cysts.

Objective: Our objective was to develop an approach for selecting combinatorial markers of pathology from diverse clinical data types. We demonstrate this approach on the problem of pancreatic cyst classification.

Materials And Methods: We analyzed 1026 patients with surgically resected pancreatic cysts, comprising 584 intraductal papillary mucinous neoplasms, 332 serous cystadenomas, 78 mucinous cystic neoplasms, and 42 solid-pseudopapillary neoplasms. To derive optimal markers for cyst classification from the preoperative clinical and radiological data, we developed a statistical approach for combining any number of categorical, dichotomous, or continuous-valued clinical parameters into individual predictors of pathology. The approach is unbiased and statistically rigorous. Millions of feature combinations were tested using 10-fold cross-validation, and the most informative features were validated in an independent cohort of 130 patients with surgically resected pancreatic cysts.

Results: We identified combinatorial clinical markers that classified serous cystadenomas with 95% sensitivity and 83% specificity; solid-pseudopapillary neoplasms with 89% sensitivity and 86% specificity; mucinous cystic neoplasms with 91% sensitivity and 83% specificity; and intraductal papillary mucinous neoplasms with 94% sensitivity and 90% specificity. No individual features were as accurate as the combination markers. We further validated these combinatorial markers on an independent cohort of 130 pancreatic cysts, and achieved high and well-balanced accuracies. Overall sensitivity and specificity for identifying patients requiring surgical resection was 84% and 81%, respectively.

Conclusions: Our approach identified combinatorial markers for pancreatic cyst classification that had improved performance relative to the individual features they comprise. In principle, this approach can be applied to any clinical dataset comprising dichotomous, categorical, and continuous-valued parameters.