Machine Learning Algorithm to Predict Change in the Decision-Making for Thoracolumbar Fractures Without Neurological Deficit After MRI: A Multicenter Study.

Abstract

Study DesignA multicenter study.ObjectiveTo develop a machine learning algorithm to predict when magnetic resonance imaging (MRI) may change the thoracolumbar AO Spine injury severity score (TLAOSIS) treatment recommendation for thoracolumbar fractures (TLFs) without neurological deficits.MethodsThree trauma centers recruited 619 neurologically intact TLFs (AO Spine A-fractures) who underwent computed tomography (CT) and MRI. CT findings indicating posterior ligamentous complex (PLC) injury were defined as facet malalignment, horizontal laminar fracture, spinous process fracture, and interspinous widening ≥4 mm. A single positive CT finding indicated an M1 modifier. The primary outcome was any change in the TLAOSIS treatment recommendation among conservative (≤3), grey zone (4-5), and surgical (>5) groups after MRI. The derivation and validation sets utilized 80% and 20% of the samples, respectively. A classification and regression tree (CART) was developed using the M1 modifier, AO fracture subtype (A1-A4), and spine level. Model discrimination was quantified using the area under the receiver operating curve (AUC).ResultsMRI altered TLAOSIS recommendations in 82 (13.2%) cases. The CART used the M1 modifier, A subtype, and spine level (importance = 0.914, 0.055, and 0.031, respectively). The model achieved an AUC of 0.93, sensitivity of 87.5%, specificity of 96.3%, and mean accuracy of 92.9% (±12.0%) in cross-validation in predicting TLAOSIS recommendation change.ConclusionThe CART model accurately predicted changes in the TLAOSIS recommendation after MRI. This algorithm provides cost-effective indications for MRI in neurologically intact AO A-type fractures, ensuring accurate PLC assessment while minimizing unnecessary imaging.