Predictive accuracy of machine learning models for conservative treatment failure in thoracolumbar burst fractures

Abstract

Background

The management of patients with thoracolumbar burst fractures remains a topic of debate, with conservative treatment being successful in most cases but not all. This study aimed to assess the utility of machine learning models (MLMs) in predicting the need for surgery in patients with these fractures who do not respond to conservative management.

Methods

A retrospective analysis of 357 patients with traumatic thoracolumbar burst fractures treated conservatively between January 2017 and October 2023 was conducted. Various potential risk factors for treatment failure were evaluated, including age, gender, BMI, smoking, diabetes, vertebral body compression rate, anterior height compression, Cobb angle, interpedicular distance, canal compromise, and pain intensity. Three MLMs—random forest (RF), support vector machine (SVM), and k-nearest neighborhood (k-NN)—were used to predict treatment failure, with the RF model also identifying factors associated with treatment failure.

Results

Among the patients studied, most (85.2%) completed conservative treatment, while 14.8% required surgery during follow-up. Smoking (OR: 2.01; 95% CI: 1.54–2.86; p = 0.011) and interpedicular distance (OR: 2.31; 95% CI: 1.22–2.73; p = 0.003) were found to be independent risk factors for treatment failure. The MLMs demonstrated good performance, with SVM achieving the highest accuracy (0.931), followed by RF (0.911) and k-NN (0.896). SVM also exhibited superior sensitivity and specificity compared to the other models, with AUC values of 0.897, 0.854, and 0.815 for SVM, RF, and k-NN, respectively.

Conclusion

This study underscores the effectiveness of MLMs in predicting conservative treatment failure in patients with thoracolumbar burst fractures. These models offer valuable prognostic insights that can aid in optimizing patient management and clinical outcomes in this specific patient population.