Multimodal temporal deep learning for postoperative complication risk stratification in colorectal surgery: a transformer-based analysis of inflammatory–metabolic dynamics

Postoperative complications after colorectal surgery are time-dependent and heterogeneous, complicating recovery and nursing resource planning. Traditional static metrics fail to capture dynamic physiological evolution. This study developed an interpretable multimodal temporal deep-learning model to characterize postoperative inflammatory–metabolic stress dynamics for predicting nursing-related complication risk. A retrospective multicenter cohort study used MIMIC-IV and eICU for model development and internal validation, and ACS NSQIP for a harmonized external transportability assessment based on available structured perioperative and outcome variables. Adult colorectal surgery patients with postoperative vital signs and laboratory data were included in the temporal development cohorts. Vital signs, inflammatory markers, metabolic indicators, and nursing proxy variables were integrated into multimodal temporal inputs. A Transformer-based model was established, with SHAP for interpretability and unsupervised clustering for trajectory identification. Performance was evaluated using discrimination, precision-recall, and calibration metrics. The primary analysis was defined as offline sequence classification over the 0–72 h postoperative window; leakage-aware sensitivity analyses used prespecified 6-, 12-, 24-, 48-, and 72-h landmarks and an online-style causal attention mask. Among 3,184 included patients, 22.4% developed complications. The multimodal Transformer model achieved an AUC of 0.84 in internal validation and 0.79 in the harmonized ACS NSQIP external assessment. The model showed higher discrimination than the static logistic-regression and random-forest baselines included in the current analysis; however, architecture-specific superiority cannot be inferred without comparisons against stronger temporal models. SHAP analysis identified the 12–36 h postoperative window as a key interval for risk separation. Unsupervised clustering revealed three inflammatory-metabolic stress trajectories with complication gradients (10.9%, 24.6%, 41.3%; p for trend < 0.001). Nursing proxy variables contributed to prediction but were interpreted as severity- and care-context markers rather than direct measures of nursing quality or causal intervention effects. This retrospective multicenter analysis suggests that multimodal temporal modeling can characterize early postoperative inflammatory-metabolic stress patterns and stratify complication risk in colorectal surgery. The findings support further prospective evaluation of dynamic monitoring strategies rather than immediate clinical deployment. This study retrospectively analyzed de-identified public databases without any health care intervention on human participants. Therefore, trial registration was not required.