Maximisation of treatment efficiency in hospitals can lead to significant growth in terms of patient satisfaction, staff productivity and hospital revenue. Scheduling for the operating department is currently done manually, using standard values for duration that vary with procedure type. This implies that no patient characteristics nor historical data are used to personalise predicted surgical duration. Furthermore, schedules are not updated when procedures are delayed. A relevant step towards dynamic scheduling is the realisation of real-time analysis of surgical workflow, based on intraoperatively acquired data. As of yet, the amount research performed into the use of automatically generated online data for duration prediction is limited. This research project involves an analysis of the usability and characteristics of interventional X-ray data, for predicting the total duration of coronary angiograms intraoperatively. A random forest classification algorithm is used to analyse each acquisition within the dataset and classify the total duration of the corresponding procedure as being below 10 minutes, 10 to 20 minutes, 20 to 30 minutes or 30 minutes and over. An additional 22 features were generated to add data from prior acquisitions of the same patient. Recursive feature elimination was used to determine the optimal feature set, to be used in the final model. Based on out-of-bag validation, the overall accuracy of the classification model was found to be 92.8\%. Considering an average procedure duration of 11 minutes, the interventional X-ray data shows exceptional capabilities of classifying both standard and delayed procedures. Further analysis with respect to procedure time shows that some Class 1 acquisitions are overestimated as Class 2 acquisitions, but that overestimation of procedure duration rarely occurs for the other three phases. This implies that a prediction made beyond 10 minutes into the procedure can be perceived as the absolute minimum duration class of the procedure. Additionally, Class 3 and Class 4 predictions are found to always correspond to the minimum procedure time, independent of procedural progress. Class 2 and Class 3 acquisitions are underestimated up until 10 minutes and 20 minutes into the procedure, respectively. Class 4 acquisitions are correctly classified at a relatively earlier point with respect to progress and are never incorrect beyond 20 minutes. Given the fact that overestimation does not occur beyond 10 minutes, nor when Class 3 or Class 4 is predicted, the first prediction of a new duration class is concluded to be a reliable reference point. Further investigation has shown that most Class 3 and Class 4 detection occur within the first five minutes of a procedure. Therefore, the model is successful at predicting a total duration of 20 minutes or more at an early stage of the procedure. This could significantly benefit the hospital in terms of procedure planning and knowing when to request the next patient. For a deeper understanding of the model's potential, individual procedures were analysed to gain insight into its additional value in terms of procedure scheduling. In terms of prediction features, the acquisition frequency, longitudinal position of the operating table, cumulative procedure duration, patient age and cumulative cine acquisition time were found to be the most important. The implementation of these features in further research on CAG phases is recommended. Movement of the operating table seems to be particularly informative for workflow analysis. In terms of clinical application, further optimisation is required in order to enable accurate prediction for the duration of shorter procedures. Nevertheless, the model provides an accurate tool for the real-time monitoring of procedural workflow and detecting significant delay, which proves that the usability of the data goes far beyond machine maintenance and service only.