The machinery health management system required for developing maritime autonomous surface ships can be realised by employing prognostics and health management (PHM) methods. Pertinent PHM models are typically trained by using datasets corresponding to limited operating conditions and are subsequently employed to analyse a wide envelope of conditions. This study employs a PHM model that consists of a Deep Neural Network (DNN) submodel and an Auto-Regressive Integrated Moving Average (ARIMA) submodel for predicting the health indicator of a marine four-stroke engine. In specific, this study aims to quantify the accuracy of this PHM model predictions. The PHM model is developed by employing limited datasets and subsequently validated by employing extended datasets. The extended datasets reflect practical operating conditions including ambient temperature variations, stochastic degradation trends, several engine loads, and multiple simultaneous degradations. The results demonstrate that, when the testing dataset is employed, the PHM model predicts the engine exhaust valve health indicator for future time slices with high accuracy of R-squared values of 0.998. However, the model accuracy deteriorated reaching R-squared values of 0.707 when validation datasets representing extended operating envelope are used. This study’s results emphasise that the PHM model accuracy is affected by the available datasets for training, necessitating the generation of trustworthy datasets and scientific methods for developing trustworthy PHM models.