USING ENSEMBLE LEARNING AND FEATURE SELECTION IN THE DIAGNOSIS OF LOW BACK PAIN

Aim:A wide variety of research is currently being conducted on how artificial intelligence can assist clinical decision-making and improve clinician judgments. The goal of this research was to develop a computer-aided diagnostic (CAD) approach that can aid healthcare professionals in identifying lumbosacral pathologies. Material and Methods:The study included 633 abnormal and 442 normal lateral lumbosacral radiographs, and the You Only Look Once algorithm was used to automate the cropping task. This study used pre-trained VGG-16, ResNet-101, and MobileNetV2 models for transfer learning. Feature extraction was performed from the intermediate layer of VGG-16, resulting in 512 features. Then, a variance threshold was applied, resulting in 221 selected features with a variance threshold of 0.01. Then, support vector classifier, logistic regression, random forest classifier, and k-nearest neighbours machine learning models were trained using both sets of 512 extracted features and 221 selected features separately. Results:The results from the ensemble learning model with the stacking classifier using features selected using a threshold value 0.01 from features extracted were: accuracy 93.0% (best); sensitivity, 91.8%; specificity, 94.1%; precision, 92.9%; F1 score, 92.3% (best); area under the receiver operating characteristic curve, 0.97 (one of the best); and Cohen's kappa, 0.86 (best). Conclusion:The ensemble learning model with a stacking classifier using features selected by using a threshold value of 0.01 from features extracted by processing the intermediate layer of VGG-16 performs better than the transfer learning models using pre-trained networks, such as VGG-16, ResNet-50, and MobileNetV2, and the learning methods that do not apply feature selection in distinguishing lumbar vertebral pathologies.