Machine Learning Prediction of Parameters of Early Warning Scores in Intensive Care Units

Overview

A large number of different organ functions are recorded in real time for patients being monitored in an intensive care unit. On the one hand, the measured values collected are used for continuous monitoring of vital parameters, e.g. blood pressure, heart rate and respiratory rate, but are also evaluated several times a day in conjunction with other data as part of ward rounds. In both cases, continuous monitoring from a limited number of parameters, but also in the distinct evaluation with a more extensive set of analyzable parameters, there are limitations in the evaluability even with all the care and expertise available: In continuous analysis, interpretation is limited by the restricted number of continuously recorded parameters described above. Although a large number of such measurements are possible, and at least theoretically a larger number of parameters could be measured, patient-specific limits such as patient cooperation, medical limits such as the significance of the measured values in specific situations, but also economic limits are often decisive in this context. Although accurate conclusions can be drawn from the continuous and therefore complete representation of aspects of human physiology, the limitation of the available parameters reduces the interpretability of the synthesis of different statuses. In the broader, more comprehensive assessments during visits at specific points in time, on the other hand, there are limitations due to, among other things, point recordings of individual measured values and the predefined visit times. Even if limit values are (or can be) defined for the measured data, and a consequence, e.g. a therapy step, is initiated if these values are exceeded or not reached, this alert can only be initiated retrospectively if these values are exceeded and a consequence can only be initiated retrospectively. In this situation, a pathophysiological change is already so far advanced that in many cases a compensation mechanism no longer functions adequately and turns into a decompensation situation. In this situation, the patients affected in an intensive care unit are in many cases in mortal danger. Both situations, continuous recording of a limited number of parameters and the evaluation of extensive data in the form of a snapshot could be optimized despite the limitations mentioned. Without changing the collection of data (time, scope, etc.), the possibilities for optimizing their interpretation and the consequences that can be derived from the interpretation remain. The interpretation of the data is primarily determined by the interpreters as the method of interpretation. Current approaches attempt to use machine learning (ML) methods to predict individual situations that recognize adverse events in the given data and at the same time allow alarms to be triggered pre-emptively, i.e. before a life-threatening situation occurs. Furthermore, there are already studies on the change of early warning scores in time series, which are, however, limited in their informative value for longer prediction periods.