Mikhail Potievskiy1*, Pavel Sokolov1, Vladimir Trifanov1, Valerii Starinskii1, Denis Larionov1, Ruslan Moshurov2, Peter Shegai1 and Andrei Kaprin1
1National Medical Research Center of Radiology, Ministry of Health (Russia)
2Voronezh Oncological Center
potievskiymikhail [at] gmail.com
Abstract
The treatment of liver and pancreatic tumors remained one of the most challenging areas in abdominal surgery. We developed approaches to model the risk of surgical complications in patients undergoing surgery for hepatopancreatobiliary tumors. Scientifically valuable data were often fragmented and stored in local databases rather than integrated into medical information systems. To address this limitation, we implemented the project “Unified Digital System for the Collection and Storage of Scientific Data” at the National Medical Research Center of Radiology in 2020, which enabled structured data aggregation and deployment of predictive models.
We included 150 patients and 22 clinical, laboratory, and perioperative parameters from a single-center retrospective study. We developed machine learning models to predict biochemical leakage and grade B/C pancreatic fistulas. We compared logistic regression, Random Forest, and CatBoost algorithms using ROC AUC and selected the best-performing model. We evaluated risk factors using Shapley values and relative importance metrics.
CatBoost demonstrated the highest predictive accuracy (ROC AUC 74–86%). Tumor vascular invasion, age, and body mass index were identified as the main preoperative and intraoperative predictors of postoperative fistulas (ROC AUC 70%). For grade B/C fistulas, the predictors were similar. Postoperative data from days 3–5 showed that blood amylase, drain amylase, and leukocyte count were significant predictors of biochemical leakage and clinically relevant fistulas (ROC AUC 86% and 75%, respectively).
The application of evolutionary biology concepts played an important role in modeling treatment outcomes. Artificial intelligence enabled the analysis of tumor microevolution, mechanisms of therapeutic resistance, and population-level differences in disease progression and treatment response. The integration of artificial intelligence into digital healthcare systems facilitated the combination of clinical, biological, and population-based data, supporting the development of personalized treatment strategies and improving outcomes in oncological care.
Keywords: Cancer predictive analytics, AI, database