Knee Osteoarthritis (OA) stands out as a one of a common joint condition in the 21st century, primarily arising from wear and tear in knee joints. Timely identification and anticipation of OA severity can significantly contribute to early intervention and improved disease management. In our study, we present a machine learning-centered strategy for detecting knee OA and predicting its severity, employing three well-established algorithms: Convolutional Neural Network (CNN), K-Nearest Neighbor (KNN), and Random Forest (RF). The CNN algorithm serves the purpose of feature extraction from input images, determining the disease’s severity. Its architecture encompasses 2 fully connected layers and 5 convolutional layers. Meanwhile, the KNN and RF algorithms are tasked with categorizing input images into five distinct levels: Normal, Doubtful, Mild, Moderate, and Severe. For measuring the performance of our proposed method, employment of various metrics or parameters, including precision, accuracy, recall, Matthews Correlation Coefficient (MCC), and F1 score. The CNN algorithm showcases a noteworthy accuracy of 89.93%, surpassing the KNN and RF algorithms, which achieve accuracies of 60% and 55.45%, respectively. Notably, the CNN algorithm demonstrates superior precision, recall, and F1 score compared to both KNN and RF algorithms. Our proposed approach exhibits promise as a screening tool for knee OA detection and severity prediction, offering potential implications for early intervention and improved management strategies for this prevalent joint condition.