An interpretable AI multimodal framework for predicting anesthesia-related hemodynamic instability: radiomics and deep feature extraction from preoperative contrast-enhanced CT combined with intraoperative clinical vital signs

We developed and externally validated an interpretable multimodal deep learning framework for predicting Hemodynamic instability (HDI) in patients undergoing elective abdominal and pelvic cancer surgery under general anesthesia, integrating (1) radiomic features and deep imaging representations from preoperative contrast-enhanced computed tomography (CT), and (2) peri-induction vital-sign summaries. This retrospective multicenter cohort study enrolled 456 patients across three centers (training/validation: n = 312, Centers 1–2; external test: n = 144, Center 3). HDI was defined as sustained MAP < 65 mmHg for > 5 min or a > 20% MAP reduction requiring vasopressor intervention. Four feature modalities were evaluated: (1) a clinical-only benchmark (L2-regularized logistic regression on seven preoperative variables), (2) radiomic-only CT features, (3) deep imaging features from four pretrained backbone architectures (ConvNeXt-Base, Swin-V2-B, EfficientNetV2-L, ViT-L/16), and (4) a full multimodal combination of CT features and peri-induction vital-sign statistics. Imaging modalities were evaluated across four tabular deep learning (DL) classifiers (TabNet, FT-Transformer, SAINT, DANet) and four feature-selection strategies (LASSO, Boruta, mRMR, SHAP-RFE), yielding 16 configurations per imaging modality. HDI occurred in 138 patients (30.3%), with comparable prevalence across centers ( χ 2 = 0.04, p = 0.98). The clinical-only L2-LR benchmark achieved an external test AUC-ROC of 0.728 (95% CI 0.667–0.787). Radiomic-only models improved performance to AUC 0.786 (0.726–0.842; ΔAUC + 0.058, p = 0.031). Deep feature-only models with the ViT-L/16 backbone achieved AUC 0.805 (0.746–0.859; ΔAUC + 0.077, p = 0.008). The primary multimodal model (SAINT + ViT-L/16 + mRMR) achieved the highest external test AUC of 0.836 (0.784–0.884) (ΔAUC + 0.108 vs clinical-only L2-LR; Westfall–Young p < 0.001), with sensitivity 0.726, specificity 0.798, Brier score 0.174, and MCC 0.511. A perioperative multimodal framework integrating preoperative CT radiomic and deep imaging features with peri-induction vital-sign data significantly outperforms clinical-only risk stratification for HDI prediction.