With current developments in machine learning, deep learning, and artificial intelligence (AI), significant advances have been achieved across various fields, especially in the healthcare sector. Besides, watermarking techniques are also widely applied to store and secure data, allowing data to be seamlessly embedded into images that are invisible to the human eye, facilitating the transmission of secure messages. However, there is still no analysis and evaluation of the impact of watermarking techniques on the prediction results of models. In this study, the Least Significant Bit-based image watermarking technique is extensively analyzed for 1-bit, 2-bit, and 3-bit embedded planes with 4 different neural network models in predicting lung pathology, including Vanilla, VGG16, Densenet-121, and Capsule. The performance indexes are considered before and after embedding to conclude the impact of embedding techniques on the predictive ability of these models. First, the impact of embedding at different bit levels on the prediction ability of the original models is investigated. To better understand the impact of embedding on the supervised learning process, these models are retrained with 1-bit and 2-bit embedded training data; then the suitable thresholds are determined for each model to help improve the models’ ability to classify and predict diseases.