Machine learning using gastric curvature angle and iliac crest level predicts nasoenteric tube success in critical care

Successful placement of nasoenteric tubes in critically ill patients remains challenging. Anatomical parameters such as gastropyloric angle and tube-to-iliac-crest distance may influence success, but their combined predictive value has not been systematically evaluated using machine learning approaches. This study aimed to investigate the independent predictive roles of the gastropyloric angle and the vertical distance from the lowest point of the nasojejunal tube to the highest point of the iliac crest for successful nasojejunal tube placement, and to develop and compare machine learning models for individualized risk prediction. A retrospective study was conducted on 287 critically ill patients who underwent nasojejunal tube placement. Eight candidate predictors were analyzed, including the gastropyloric angle and the vertical distance from the lowest point of the nasojejunal tube to the highest point of the iliac crest (distance_tube_to_iliac_crest). Five models—Logistic Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—were developed and compared. Model performance was evaluated using accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). Feature importance was derived from the RF and LR models. The overall success rate was 80.13% (230/287). LR achieved the highest AUC (0.7464), and KNN achieved the highest accuracy (0.8276) and F1 score (0.8980). No single model outperformed others across all metrics, and all AUC values remained below 0.75, indicating only moderate discriminative ability. SVM showed perfect recall (1.00) but low AUC (0.5851), suggesting overfitting. Feature importance analysis revealed that distance_tube_to_iliac_crest was the most important predictor (0.2424), followed by the gastropyloric angle (0.2244) and insertion depth (0.1430). These three features collectively accounted for more than 60% of the predictive information. Five-fold cross-validation confirmed the stability of the LR model (mean AUC 0.7348 ± 0.0668). The gastropyloric angle and the vertical distance from the lowest tube point to the iliac crest are independent anatomical predictors of successful nasojejunal tube placement, supporting the “turning resistance hypothesis”. LR was selected for the web-based risk calculator primarily due to its interpretability, although its AUC of 0.7464 indicates only moderate predictive performance. KNN showed superior classification accuracy and F1 score, suggesting it may be preferred when the clinical goal is to maximize correct classification. External validation is warranted before clinical deployment. Based on the model, a three-tier risk stratification strategy was proposed to guide the selection of blind versus image-guided placement, aiming to improve first-attempt success and reduce complications.