Background and Objective Postoperative pulmonary complications (PPCs) are common after elective colorectal surgery and are associated with prolonged hospitalization, increased morbidity, and delayed recovery. Existing PPC prediction tools were mainly developed in thoracic or mixed surgical populations and rely predominantly on clinical variables; preoperative chest CT information remains underused, and a population-specific tool for elderly colorectal cancer patients is still lacking. This study aims to develop and validate a multimodal deep learning model that integrates preoperative chest CT-derived features with multidimensional clinical variables to predict PPCs in patients aged ≥65 years undergoing elective radical colorectal cancer surgery, and to quantify its incremental value and clinical net benefit.
Methods This is a single-center retrospective cohort study. Eligible patients will be assigned to a development window (August 2020-August 2024), split chronologically at a 7∶1∶2 ratio into training, validation, and test sets with five-fold cross-validation, and a locked temporal external validation window (September 2024-September 2026), which will be used only after the final model is fixed. The primary outcome is 30-day composite PPCs defined according to the EPCO criteria; the Melbourne Group Scale will serve as a parallel sensitivity definition. Preoperative clinical variables and chest CT images obtained within four weeks before surgery will be collected. A convolutional neural network imaging branch-or 2D transfer learning based on RadImageNet if the sample size is insufficient-will extract CT deep features, which will be fused with clinical variables at the feature (intermediate) level. Six classifiers, including logistic regression, support vector machine, random forest, XGBoost, LightGBM, and multilayer perceptron, will be tuned within a unified search space. Discrimination (AUC with DeLong test and Bootstrap 95% CI), calibration (calibration curves, Brier score), incremental value (NRI, IDI), and clinical utility (decision curve analysis) will be evaluated against the ARISCAT and GSU-Pulmonary scores, a clinical-only model, and an imaging-only model. Grad-CAM and SHAP will be applied to interpret the model.
Expected Outcome The study is expected to provide an interpretable and deployable preoperative risk stratification tool for elderly patients undergoing colorectal surgery, clarify the incremental contribution of preoperative chest CT information, and define the threshold range in which the model yields net clinical benefit.
More