DSPONVNet: a multimodal deep learning model integrating intraoperative monitoring and clinical features for predicting postoperative nausea and vomiting risk

Postoperative nausea and vomiting (PONV) is a frequent and distressing complication that affects patient comfort and recovery following surgery. Conventional risk prediction models rely heavily on static clinical features, often overlooking real-time physiological signals. This study aimed to develop a robust prediction model that integrates intraoperative monitoring data with structured clinical variables to enhance the accuracy of PONV risk assessment. We proposed DSPONVNet, a multimodal deep learning model incorporating a multilayer perceptron (MLP) for static features, a long short-term memory (LSTM) network for dynamic intraoperative monitoring data, and a self-attention mechanism for feature fusion. A total of 53,250 patients who underwent general anesthesia were retrospectively included. The model was trained and evaluated using stratified data partitioning and compared with five baseline models. DSPONVNet achieved superior performance, with a ROC-AUC of 0.9376 and F1 score of 0.8701, outperforming all baseline models. SHAP analysis revealed that both traditional risk factors (e.g., female sex, prior PONV) and intraoperative heart rate fluctuation significantly contributed to risk prediction, highlighting the value of integrating dynamic physiological data. DSPONVNet demonstrates enhanced predictive capability and interpretability by incorporating intraoperative monitoring data, enabling more accurate and individualized PONV risk assessment. These findings support the use of real-time data fusion in clinical decision support systems for perioperative care optimization.