A landmark-based dynamic prediction model for anastomotic leakage after rectal cancer surgery: integrating perioperative and postoperative trajectories

Figures

• 

  Figure 1. Flowchart of the two-stage landmark-based dynamic prediction model for anastomotic leakage (AL). (1) Cohort selection: From 853 screened patients,general exclusions yielded the complete eligible surgical cohort (N = 816). This complete cohort was specifically utilized for a sensitivity analysis to verify the baseline stability of Model 1 immediately after surgery (Supplementary Table 1). Subsequently, landmark exclusion of early AL diagnosed ≤ POD3 formally established the POD3 landmark primary cohort (N = 815). (2) Feature streams: Sixteen consensus baseline predictors were identified from 24 candidates using a majority voting strategy across four feature selection algorithms (LASSO, RF-MDA, SVM-RFE, and Stepwise AIC). Concurrently, postoperative longitudinal NLR trajectories were stratified via PAM clustering, alongside the tracking of dynamic markers (albumin <30 g/L, corrected calcium). (3) Modelling: Model 1 estimated immediate postoperative AL risk utilizing baseline predictors. Model 2 dynamically updated this risk on postoperative day 3 (POD3) by integrating early postoperative tracking variables. (4) Validation and translation: Model performance was rigorously assessed…

• 

  Figure 2. Forest plots of multivariable logistic regression analyses for the prediction models. (A) Model 1 (baseline risk assessment): Adjusted odds ratios (aORs) for the 16 consensus preoperative and intraoperative predictors. (B) Model 2 (sequential risk refinement): Updated risk profile on postoperative day 3 (POD3) integrating longitudinal trajectory and dynamic biological markers. Centers and error bars represent aORs with 95% confidence intervals (CIs). The vertical dashed line indicates an OR of 1.0. P < 0.05 indicates statistical significance.

• 

  Figure 3. Discriminative performance and clinical utility of the machine learning algorithms and the two-stage prediction models. (A) Receiver operating characteristic (ROC) curves comparing the discriminative ability of the four individual machine learning algorithms and the consensus baseline model (Model 1). (B) ROC curves demonstrating the superior discrimination of the enhanced postoperative model (Model 2) compared to Model 1 (DeLong test, P < 0.001). (C) Decision curve analysis (DCA) evaluating the clinical net benefit of the initial algorithms and Model 1. (D) DCA illustrating the increased clinical net benefit of Model 2 over Model 1 across a wide range of threshold probabilities.

• 

  Figure 4. Consensus of variable selection across four distinct algorithms visualized by an UpSet plot. This UpSet plot displays the intersection sizes (top bar chart) and set sizes (left horizontal bar chart) of variables selected by four different feature selection techniques: Stepwise AIC, Support Vector Machine-Recursive Feature Elimination (SVM-RFE), Random Forest-Mean Decrease Accuracy (RFMDA), and Least Absolute Shrinkage and Selection Operator (LASSO). (Top Bar Chart) Shows the number of variables unique to specific combinations of selection methods. The leftmost bar confirms that 7 key variables (including Cluster NLR, Age, etc.) were consistently selected by all four algorithms. (Left Horizontal Bar Chart) Shows the total number of variables selected by each individual algorithm. (Matrix) The dark circles and lines connect the specific combinations of algorithms forming each intersection.