Predicting slot lengths of MRI exams to decrease observed discrepancies between planning and execution

Abstract

This retrospective study aimed to reveal discrepancies between planned ($T_{\mathrm{plan}}$) and actual ($T_{act}$) slot lengths of abdomen MRI exams, and to improve $T_{plan}$ by predicting slot lengths via a machine learning algorithm. $T_{plan}$ and $T_{act}$ were retrieved from RIS and modality logfiles, respectively, covering 3038 MRI exams of 17 protocols performed at an abdomen department. Comparisons showed that 30% of exams exceeded planned slot lengths. On the other hand, exams completed within planning failed to manifest good adherence to schedule, as many of them were assigned with an unnecessarily long slot. While adjusting the planned exam duration by a fixed amount of time for each protocol could move $T_{plan}$ closer to the mean or median $T_{act}$, the large spread of $T_{act}$ would still be unaffected. This is why this study goes one step further, introducing a method to predict the required slot length not only per protocol, but for each individual exam. A Random Forest Regression model was trained on historic data to predict individual slot lengths ($T_{\mathrm{pred}}$) based on patient and exam context. The correlation between $T_{pred}$ and $T_{act}$ was found to be better than that of $T_{plan}$ and $T_{act}$, with Pearson correlation factors of 0.66 and 0.50, respectively. The overall adherence to schedule was also improved by the prediction, as seen by a reduction of both the root mean squared error (–28%) and the standard deviation (–16%) of the differences between planned/predicted slot times and $T_{act}$. To provide further insights into the discrepancies between planning and execution of MRI exams, nineteen exams from the Liver protocol with verified clinical information were selected. This case study showed that patient conditions, diagnostic purposes and the selection of sequences during exams could explain some variations of exam durations, but the potential for improving the exam time prediction by including this additional context is limited.