A Machine Learning Approach for Predicting Failed First-Attempt Radial Artery Puncture Patients with Heart Failure

Abstract

Objective: To identify risk factors for first-attempt failure of radial arterial puncture in heart failure patients and to develop and compare predictive models using logistic regression with advanced feature engineering and ensemble learning approaches.

Method: A retrospective study was conducted involving 789 heart failure patients who underwent radial arterial puncture. Patients were divided into a training set (80%) and a test set (20%) using a stratified hold-out method. A logistic regression framework incorporating feature engineering (interaction terms, transformations, composite scores, PCA) and a dual-stage variable selection strategy (LASSO followed by stepwise selection with p<0.1 threshold) was employed. Four ensemble models were developed in parallel. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and calibration metrics through a multi-repeat validation framework.

Results: The final logistic regression model identified eleven variables, with edema degree (β = -0.0959, OR = 0.9085, p = 0.0160) and the interaction between ejection fraction and log-transformed BNP (β = 0.1507, OR = 1.1627, p = 0.0371) reaching statistical significance. The model demonstrated fair discriminative ability with an average test AUC of 0.693 (±0.033), high specificity (95.59% ± 2.55%), but lower sensitivity (31.88% ± 13.98%). The ensemble learning models showed weaker discriminative performance but exhibited potential overfitting on the training set. The optimal probability cutoff for the logistic model was 0.658.

Conclusion: Both modeling approaches developed effective prediction tools. The logistic regression model provided clinically interpretable risk factors, while the ensemble learning models achieved higher discriminatory power at the cost of interpretability. These models can assist in pre-procedural identification of high-risk patients, allowing for tailored strategies to improve first-attempt success rates and reduce patient discomfort.