Background: Despite the wealth of information carried, periodic brain monitoring data are often incomplete with a significant amount of missing values. Incomplete monitoring data are usually discarded to ensure purity of data. However, this approach leads to the loss of statistical power, potentially biased study and a great waste of resources. Thus, we propose to reuse incomplete brain monitoring data by imputing the missing values - a green solution! To support our proposal, we have conducted a feasibility study to investigate the reusability of incomplete brain monitoring data based on the estimated imputation error.
Materials and methods: Seventy-seven patients, who underwent invasive monitoring of ICP, MAP, PbtO (2) and brain temperature (BTemp) for more than 24 consecutive hours and were connected to a bedside computerized system, were selected for the study. In the feasibility study, the imputation error is experimentally assessed with simulated missing values and 17 state-of-the-art predictive methods. A framework is developed for neuroclinicians and neurosurgeons to determine the best re-usage strategy and predictive methods based on our feasibility study.
Results/conclusion: The monitoring data of MAP and BTemp are more reliable for reuse than ICP and PbtO (2); and, for ICP and PbtO (2) data, a more cautious re-usage strategy should be employed. We also observe that, for the scenarios tested, the lazy learning method, K-STAR, and the tree-based method, M5P, are consistently 2 of the best among the 17 predictive methods investigated in this study.