Abstract
INTRODUCTION: Lumbar prolapsed disc (LPD) is a leading cause of low back pain, contributing significantly to global disability and healthcare burden. This study aimed to develop machine learning models to predict the risk of LPD by analysing gene expression profiles for early detection. METHODS: Transcriptomic data from peripheral blood samples were obtained from the Gene Expression Omnibus (GEO) database, with dataset GSE150408 used for training and GSE124272 for testing. The training dataset included 17 patients with sciatica resulting from LPD, all of whom had magnetic resonance imaging confirmation of single-level LPD at either the L4/5 or L5/S1 levels. Data from 17 healthy volunteers were used as controls. Recursive feature elimination (RFE) was employed to identify the most relevant gene signatures among 23 pain-related genes. Machine learning models, including support vector machine (SVM), random forest, k-nearest neighbours (KNN), logistic regression, and Extreme Gradient Boosting (XGBoost), were trained and evaluated. Model performance was assessed using accuracy, area under the curve (AUC), F1 score, and Matthews correlation coefficient (MCC). RESULTS: Eight key gene signatures were identified as significant predictors of LPD, with MMP9 exhibiting the highest importance score. Most of these genes were differentially expressed between patients with LPD and healthy controls (p < 0.05). Among the models, random forest demonstrated the highest accuracy (0.80, 95% CI 0.73-0.85) and MCC (0.64, 95% CI 0.53-0.76), followed by KNN, XGBoost, and SVM. Overall, the random forest model exhibited the most robust performance in predicting the risk of LPD. CONCLUSION: The results of our study suggest that machine learning models based on pain-related gene signatures may identify patients at high risk of developing LPD with reasonably high accuracy. These prediction models could perhaps be integrated into clinical diagnostic tools to enhance early diagnosis and prevention.