Pancreatic surgery, particularly pancreaticoduodenectomy (Whipple procedure) and distal pancreatectomy, carries substantial risk of massive intraoperative blood loss (IBL). Predicting which patients are at high risk before surgery could allow surgeons to better prepare with blood products, surgical teams, and modified operative strategies.
This study used a retrospective cohort of 175 patients who underwent surgery for pancreatic ductal adenocarcinoma (PDAC) to build a predictive model for massive bleeding. The dataset included 34 preoperative variables covering patient demographics, laboratory values, comorbidities, and planned surgical procedure.
The approach centers on decision tree modeling, a machine learning method that creates interpretable if-then rules from patient data. Decision trees are particularly valuable in surgical contexts because clinicians can easily follow the logic and apply it prospectively without specialized software.
A striking finding was that 50.3% of patients in the cohort experienced massive intraoperative blood loss (IBL), defined using established surgical thresholds. This high prevalence confirms that massive bleeding is not a rare complication in pancreatic surgery but rather a common event that surgeons must be prepared to manage.
The cohort underwent either pancreaticoduodenectomy (PD) or distal pancreatectomy (DP), two fundamentally different operations with different anatomical challenges and bleeding risks. This distinction was captured in the model as a key variable, with surgical procedure type serving as the first split in the decision tree.
The high event rate (50.3%) is favorable from a statistical modeling standpoint, as it ensures both classes are well-represented in training data and reduces the risk of class imbalance distorting model performance. It also reinforces the clinical importance of a reliable pre-surgical risk assessment tool.
The decision tree algorithm identified surgical procedure type (PD vs. DP) as the most informative split at the root node - meaning this single variable best separated patients into high and low bleeding risk groups at the first level of the tree. This finding aligns with clinical intuition, as the Whipple procedure (PD) is anatomically more complex and typically involves more vascular dissection than distal pancreatectomy.
For patients undergoing pancreaticoduodenectomy, the tree's second split involved diabetes mellitus, suggesting that diabetic patients undergoing PD carry different bleeding risk than non-diabetic patients. Diabetes may reflect differences in tissue fragility, vascular anatomy, or healing capacity.
The tree was validated using a separate testing dataset, and performance metrics included sensitivity (ability to correctly identify true high-bleeding cases) and specificity (ability to correctly identify true low-bleeding cases). Model performance was also compared to standard logistic regression to assess whether the tree added value over simpler approaches.
Complementary logistic regression analysis identified ALT (alanine aminotransferase) - a liver enzyme - and performance of distal pancreatectomy (DP) as statistically significant predictors of massive bleeding. Elevated ALT may indicate hepatic involvement or portal hypertension, which can increase surgical complexity and bleeding risk.
The fact that DP appeared as a protective factor in logistic regression (relative to PD) is consistent with the decision tree finding that surgical procedure is the primary discriminator. However, logistic regression provides a continuous probability estimate, whereas the decision tree provides discrete risk categories that are more directly actionable in clinical settings.
The combination of decision tree and logistic regression analyses provides complementary perspectives: the tree offers interpretable if-then rules for clinical use, while regression provides odds ratios and confidence intervals for understanding the magnitude of individual predictors' contributions to bleeding risk.
The decision tree achieved 98.5% sensitivity on the training dataset and 100% sensitivity on the testing dataset, meaning it correctly identified virtually all patients who would go on to experience massive bleeding. High sensitivity is the critical performance metric in this context, since missing a true high-risk patient (a false negative) could result in inadequate preparation and preventable complications.
These very high sensitivity values should be interpreted cautiously, as they may reflect some degree of overfitting or the relatively small sample size of the testing set. External validation in independent pancreatic surgery cohorts from different institutions would be necessary to confirm that these sensitivity estimates generalize.
The specificity values (correctly identifying low-risk patients) were not reported as being as extreme, reflecting the typical precision-recall tradeoff in predictive models. In surgical planning, it may be acceptable to over-predict bleeding risk (preparing blood products unnecessarily) but not to under-predict it (being caught unprepared).
A validated preoperative decision tree for massive bleeding could be integrated into surgical planning workflows at pancreatic cancer centers. High-risk patients could receive preoperative interventions such as nutritional optimization, correction of coagulopathy, cross-matched blood product preparation, and assignment to the most experienced surgical teams.
The use of only preoperative variables (no intraoperative data) makes this tool practically applicable: risk assessment can occur days or weeks before surgery during the routine evaluation period. This is a key advantage over models that incorporate intraoperative factors discovered only during the procedure itself.
Limitations include the single-center retrospective design, the relatively small cohort of 175 patients, and the need for external validation. Additionally, surgical technique and institutional experience vary considerably across pancreatic surgery centers, potentially limiting the generalizability of a model trained at one institution to others with different case volumes and outcomes.