An artificial intelligence platform for predicting postoperative complications in metastatic spinal surgery: development and validation study

Abstract

Abstract Background Metastatic spinal disease often leads to significant morbidity, and accurate prediction of postoperative outcomes can help optimize patient management and resource allocation. The development of such a predictive tool is crucial in clinical decision-making and enhancing patient care. Hence, this study aims to establish and validate an artificial intelligence (AI) platform to predict postoperative complications in patients with metastatic spinal disease undergoing decompressive surgery. Methods The study analyzed 504 surgically treated patients with spinal metastases from four tertiary hospitals. Among them, 379 patients from three hospitals were treated as model derivation cohort and were randomly divided into two cohorts with a 7:3 ratio. The larger cohort (70%) was used to train machine learning-based models, while the smaller cohort (30%) served to internally validate the models. The remaining 125 patients from another tertiary hospital were served as external validation cohort to externally validate the model. The machine learning algorithms employed in this study includes logistic regression (LR), decision tree (DT), random forest (RF), extreme gradient boosting machine (eXGBM), neural network (NN), and k-nearest neighbor (KNN). The evaluation metrics focused on both discrimination and calibration to ensure robust model’s prediction performance. Results The incidence of postoperative complications was found to be 12.14% in the model derivation cohort. Subgroup analysis identified current smoking status (P = 0.001), diabetes (P = 0.025), hypertension (P = 0.013), visceral metastases (P = 0.017), surgical time (P = 0.015), blood transfusion (P = 0.031), and the number of surgical segments (P = 0.021) as significant factors. These variables were thus incorporated as key features in all the predictive models. Among the six machine learning models, the eXGBM model demonstrated the highest prediction accuracy, with an area under the curve (AUC) of 0.924 (95% CI: 0.884–0.965), outperforming the RF model (AUC = 0.892, 95% CI: 0.846–0.938) and the KNN model (AUC = 0.869, 95% CI: 0.818–0.920). The eXGBM model also excelled in accuracy (0.853), precision (0.853), recall (0.853), F1 score (0.853), Brier score (0.101), and log loss (0.347) metrics. External validation of the eXGBM model also demonstrated a favorable prediction performance with an AUC value of 0.880 (95% CI: 0.830–0.929). Feature importance analysis highlighted surgical time, blood transfusion, and the number of surgical segments as the three most critical factors influencing postoperative complications. In addition, the final optimal model has been made available online as a freely accessible AI platform, and the URL was https://v37vx6dtwdyf4y4nn24cib.streamlit.app/ . Conclusions The AI platform can serve as a valuable tool for assessing the risk of postoperative complications among patients with metastatic spinal disease. This predictive tool can assist healthcare professionals in making informed clinical decisions, ultimately improving patient outcomes and optimizing resource use. Further studies are recommended to validate the model’s effectiveness across diverse patient populations and clinical settings, potentially expanding its application in other areas of surgical oncology.

Keywords

• Machine learning
• Metastatic spinal disease
• Postoperative complications
• Artificial intelligence
• Prediction platform