386 An Externally Validated Machine Learning Ensemble Model Accurately Predicts Important Neurosurgical Outcomes

Abstract

Abstract INTRODUCTION Machine learning (ML) uses sophisticated computer algorithms to “learn” patterns in complex, heterogeneous datasets, and then to apply those patterns to never-before-seen data. In this proof-of-concept study, we use a novel approach of algorithm selection and combination to build and externally validate a ML-based model that predicts extended hospitalization (>7 days) in patients undergoing craniotomy for brain tumor, an important outcome for providers and patients. METHODS A training dataset of 41,222 patients was created from the National Inpatient Sample (NIS). 26 ML algorithms were trained on 26 preoperative variables to predict extended hospitalization, and the most predictive algorithms selected and combined to create an ensemble model. The area under the curve (AUC) of the receiver operating characteristic curve was calculated for the trained ensemble model to demonstrate the ensemble's discriminative ability. The negative predictive value, positive predictive value, sensitivity, and specificity of the ensemble were calculated at the optimal F1 score. Finally, the ensemble was externally validated using a dataset of 4592 patients generated from the National Surgical Quality Improvement Program (NSQIP) database. RESULTS >An ensemble model comprising two tree-based algorithms and a regularized logistic regression best predicted the outcome of interest. On internal validation, the ensemble had an AUC of 0.8024, and on external validation, an AUC of 0.7666. CONCLUSION A ML-based ensemble model predicts extended hospitalization for patients undergoing craniotomy for brain tumor with good discrimination on both internal and external validation. This proof of concept study is the first to describe guided algorithm selection and ensemble model creation to predict neurosurgical outcomes. This broadly applicable technique optimizes predictive accuracy, allowing researchers to confidently build the most predictive models for their data, potentially revolutionizing outcomes research.

Ähnliche Arbeiten