Prior knowledge of item characteristics, such as difficulty and response time, without pretesting items can substantially save time and cost in high-standard test development. Using a variety of machine learning (ML) algorithms, the present study explored several (non-)linguistic features (such as Coh-Metrix indices) along with MPNet word embeddings to predict the difficulty and response time of a sample of medical test items. In both prediction tasks, the contribution of embeddings to models already containing other features was found to be extremely limited. Moreover, a comparison of feature importance scores across the two prediction tasks revealed that cohesion-based features were the strongest predictors of difficulty, while the prediction of response time was primarily dependent on length-related features.