Investigating the utilization of LIME-enhanced AI and generated data for evaluating the risks associated with radial artery puncture procedures.

Radial artery puncture, a routine arterial cannulation procedure for perioperative and critical care settings, is limited by high first-attempt failure and hematoma risk under experience-based practice. We aim to develop artificial intelligence (AI) - powered precise puncture guidance to reduce procedure-related complications and improve patient experience. In this study, Synthetic Data Vault (SDV) algorithm is used to generate patient data as training set data according to the real data provided by Biostudies. Ten machine learning models were built via Python, and performance was evaluated using accuracy, precision, recall and area under the receiver operating characteristic curve (AUC). In the real-world test cohort, there were 185 positive cases of the first-attempt radial artery puncture success, corresponding to a positive proportion of 72.27%. For the postoperative hematoma outcome, 40 positive cases were identified, with a positive proportion of 15.63%. Local Interpretable Model-agnostic Explanations (LIME) integrated with decision tree classifier identified the top 3 predictors of puncture-related adverse events, enabling individualized risk stratification for clinical decision-making. For first-attempt puncture success prediction in the test set, all models except logistic regression, Gaussian Naive Bayes (GNB) and k-Nearest Neighbors (KNN) achieved accuracy > 0.750; all except KNN and Light Gradient Boosting Machine (LGBM) had precision > 0.830. LGBM (0.951) and CatBoost (0.881) yielded the highest recall, while CatBoost (0.763) and eXtreme Gradient Boosting (XGB) (0.750) had the highest AUC. For post-puncture hematoma prediction, all models except KNN had accuracy > 0.840; the top three precision values were from Multilayer Perceptron Classifier (MLPC, 0.857), Linear Support Vector Classification (LinearSVC, 0.667) and Adaptive Boosting (ADAB, 0.556). GNB achieved the highest recall (0.450), and all models except decision tree and KNN had AUC > 0.750. XGB achieved the optimal overall performance for first-attempt radial artery puncture success prediction in the tested single-center cohort, while MLPC yielded the highest precision for post-puncture hematoma prediction among the 10 evaluated algorithms.