Abstract
OBJECTIVE: Gastrointestinal dysfunction following esophageal cancer surgery represents a prevalent postoperative complication. This study aims to develop a time-dynamic machine learning model to predict the duration of postoperative gastrointestinal dysfunction (POGID) by integrating preoperative and perioperative continuous blood data with dynamic ultrasound characteristics, thereby facilitating early clinical intervention. METHODS: A retrospective cohort of 826 patients who underwent radical esophagectomy between 2017 and 2024 was enrolled and stratified into a training set (70%), a validation set (15%), and a test set (15%). Predictive variables encompassed baseline demographic and clinical data, blood routine parameters at five distinct time points, and ultrasound features at three time points. Four machine learning models were constructed: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), and Random Forest (RF). Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R(2)), and Mean Absolute Percentage Error (MAPE). Feature importance was assessed via SHapley Additive exPlanations (SHAP) analysis. RESULTS: The LSTM model demonstrated superior predictive performance on the test set, achieving an MAE of 1.23 ± 0.31 days, an RMSE of 1.56 ± 0.42 days, an R(2) of 0.78 ± 0.06, and an MAPE of 12.3% ± 3.1%, significantly outperforming the RF model (all P < 0.001). The top five influential predictors were postoperative day 1 white blood cell count, preoperative day 1 antral cross-sectional area, postoperative day 3 platelet count, Tumor Node Metastasis (TNM) stage, and postoperative day 2 intestinal peristalsis frequency. Subgroup analyses confirmed the model's robust predictive capability, with R(2) values ranging from 0.72 to 0.83. CONCLUSION: The time-dynamic LSTM model, which integrates continuous blood data, ultrasound features, and baseline characteristics, accurately predicts POGID duration and identifies actionable intervention targets. This model can be integrated into clinical decision support systems to optimize perioperative management and enhance postoperative recovery.