Deep Learning–Based Classification of Lumbar Spine Disorders Using Biomechanical Features: Model Development, Evaluation, and Clinical Implications

Abstract

This project is designed for Automated Identification of HERNIA, NORMAL, and SPONDYLOLISTHESIS conditions, since it is indeed cumbersome to accurately classify subtle and overlapping biomechanical patterns within the lumbar spine data. Secondly, a DNN model using three deep layers, activation using ReLU, L2 regularization, batch normalization, and dropout values of 0.5 and 0.3 were constructed to mitigate the overfitting effect. The proposed idea was suitable for small & complex medical datasets, whereas Early Stopping and adaptive learning-rate scheduling were applied to stabilize training and improve generalization. The proposed model performed well in classification, obtaining a test accuracy of 80.65% at the end. The highest detection accuracy corresponded to spondylolisthesis, which is due to its more specific biomechanical characteristics, and Normal revealed moderate results. Hernia was the most challenging class, as the feature patterns are rather specific, but the model still distinguished between it meaningfully. In conclusion, the results validate the proposed methodology, achieving state-of-the-art performance in lumbar spine condition classification and laying a strong foundation for future clinical decision-support systems.