Data-driven predictive vehicle maintenance can in principle reduce the risk of costly breakdowns, damaged cargo, and increased emissions due to faulty components. However, implementing a cost efficient predictive maintenance policy is far from trivial. This thesis presents a machine learning-based method for predicting turbocharger component failure in heavy-duty trucks, called Survival-PLSTM. Survival-PLSTM uses a recurrent neural-network to model time-series data of a vehicle's operational history, and outputs an approximated cumulative hazard function. The method addresses a number of challenges related to modeling vehicle component failures, including right-censored data, class imbalance and uninformative features. The maintenance policy generated by Survival-PLSTM is shown to be more cost efficient than a corrective maintenance policy under feasible conditions. When predicting failure within a given time period as a binary outcome, Survival-PLSTM compares favorably with the Random Survival Forest (RSF) model — previously proposed for similar applications — in terms of sensitivity and specificity. The cost efficiency of a model-generated maintenance policy is shown to depend on the ratio between false-negative cost and false-positive cost, $\beta :=\frac{\text{false negative cost}}{\text{false positive cost}}$. Survival-PLSTM generates a more cost efficient policy when $\beta > 3.5$, while RSF is superior when $1 \leq \beta \leq 3.5$. Both models produce more cost efficient maintenance policies than a corrective maintenance policy.