Predicting Critical Care Ultrasound Penetration Using Ensemble Least Absolute Shrinkage And Selection Operator In High-Dimensional Data

Critical care ultrasound (CCUS) has emerged as an essential bedside diagnostic modality. Nevertheless, its penetration among practitioners, defined as multi-procedural competency, remains poorly understood and unevenly distributed across geographic, specialty, and training contexts. Existing literature is dominated by descriptive surveys and clinical findings that employ standard logistic regression, with no prior application of machine learning to predict individual-level CCUS penetration using high-dimensional data. This study aimed to develop an interpretable machine learning model to predict Critical Care Ultrasound Penetration (CCUSP) and identify key determinants of adoption using high-dimensional clinical survey data. A binary CCUSP target variable was constructed from the Critical Care Ultrasound Practice (CUSP) Survey of 2,748 practitioners across 39 countries, classifying practitioners as High CCUSP (proficient in two or more of four core ultrasound-guided procedures) or Low CCUSP. A Tuned Ensemble LASSO model was developed within a nested cross-validation framework, incorporating SMOTE for class imbalance and SHapley Additive exPlanations (SHAP) for model interpretability. The Tuned Ensemble LASSO, optimized with a Youden threshold of 0.402, achieved an F1-score of 0.839, recall of 85.9\%, accuracy of 76.6\%, and AUC-ROC of 0.724 on the held-out test set. SHAP analysis identified Adult versus Paediatric Practitioner (mean $|$SHAP$|$ = 0.392), Advanced POCUS Certification (mean $|$SHAP$|$ = 0.363), and Practice in a High-Income Country (mean $|$SHAP$|$ = 0.256) as the three most influential predictors. This study demonstrates that ensemble machine learning with SHAP-based interpretability can effectively predict CCUSP from high-dimensional survey data, providing ranked, quantified evidence on the determinants of ultrasound adoption.