Machine Learning Offers Exciting Potential for Predicting Postprocedural Outcomes: A Framework for Developing Random Forest Models in IR

Abstract

Purpose: To demonstrate that random forest models trained on a large national sample can accurately predict relevant outcomes and may ultimately contribute to future clinical decision support tools in IR.

Materials and methods: Patient data from years 2012-2014 of the National Inpatient Sample were used to develop random forest machine learning models to predict iatrogenic pneumothorax after computed tomography-guided transthoracic biopsy (TTB), in-hospital mortality after transjugular intrahepatic portosystemic shunt (TIPS), and length of stay > 3 days after uterine artery embolization (UAE). Model performance was evaluated with area under the receiver operating characteristic curve (AUROC) and maximum F1 score. The threshold for AUROC significance was set at 0.75.

Results: AUROC was 0.913 for the TTB model, 0.788 for the TIPS model, and 0.879 for the UAE model. Maximum F1 score was 0.532 for the TTB model, 0.357 for the TIPS model, and 0.700 for the UAE model. The TTB model had the highest AUROC, while the UAE model had the highest F1 score. All models met the criteria for AUROC significance.

Conclusions: This study demonstrates that machine learning models may suitably predict a variety of different clinically relevant outcomes, including procedure-specific complications, mortality, and length of stay. Performance of these models will improve as more high-quality IR data become available.