Aim: The aim is to investigate the predictability of pressure ulcers (PUs) incidence and its association with multiple parameters using machine learning (ML) methodology. The medical data were retrospectively collected within the Medical Information Mart for Intensive Care (MIMIC) project. Challenges presented by the problem's high-dimensional nature were addressed and the importance of model selection and data processing. Method: Health-related datasets contain many irregularly sampled time-variant and scarcely populated features, which may exceed the number of observations. ML techniques are effective when applied to such datasets. Successful utilization of ML-based PU prediction requires consistent reporting of clinical variable selection, data pre-processing, and model specifications. The ML techniques include regression algorithms, instance-based algorithms, ensemble algorithms, artificial neural network algorithms, and Bayesian algorithms. A custom database for PU prediction includes 7 time-invariant and 16 time-variant features for PU patients and a randomly sampled control group of the same size (4652 patients). Results / Discussion: The best performing random forest model yields an accuracy of 96%. The predictor importance may differ significantly in time for any given patient. The most important patient features are time-invariant (ICU length of stay, total input/ output). Conclusion: We uniquely explore the theoretical and practical considerations of applying six different classification models for predicting PUs using one of the most extensive medical databases (MIMIC-IV dataset). Acknowledgements: This work was supported by the Ministry of Health of the Czech Republic under grant no. NU21-09-00541, "The role of oxidative stress in pressure ulcers treatment in a patient with spinal injury". All rights reserved.