Machine-learning-based prediction of functional recovery in deep-pain-negative dogs after decompressive thoracolumbar hemilaminectomy for acute intervertebral disc extrusion.

Abstract

Objective: To develop and compare machine-learning algorithms to predict recovery of ambulation after decompressive surgery for acute intervertebral disc extrusion (IVDE).

Study design: Multicenter retrospective cohort study.

Sample population: Deep-pain-negative dogs with acute IVDE (n = 162).

Methods: Clinical variables were preprocessed for machine learning and split into independent training and test sets in an 80:20 ratio. Each model was trained and internally validated on the full test set (Test_full) and the XGBoost algorithm validated on the same test set with preoperative variables withheld (Test_wh).

Results: Recovery of ambulation was recorded in 86/162 dogs (53.1%) in this sample population after decompressive surgery. The XGBoost algorithm achieved the best performance with an area under the receiver operating characteristic curve (AUC) of .9502 (95% CI: .8919-.9901), an accuracy of .8906 (95% CI: .8125-.9531), a sensitivity of .8750, and a specificity of .9063 on Test_full. XGBoost performance on Test_wh was decreased, with an AUC of .8271 (95% CI: .7186-.9209), an accuracy of .7187 (95% CI: .6093-.8281), a sensitivity of .5625, and a specificity of .8750.

Conclusion: Machine-learning algorithms may predict outcomes accurately in deep-pain-negative dogs with IVDE after decompressive surgery. The XGBoost algorithm performed best on tabular data from this veterinary population undergoing spinal surgery.

Clinical significance: Machine-learning algorithms outperform current methods of prognostication. Pending external validation, machine-learning algorithms may be useful as assistive tools for surgical decision making.