Researchers from the Massachusetts Institute of Technology have developed a new machine-learning model that groups patients into subpopulation according to their health status and predicts the patient’s risk of dying during their stay in intensive care unit (ICU). This model is effective enough that it outclasses global mortality prediction models and exposes performance disparities of those models across specific patient subpopulation.
The patients in the ICU contains wide range of health conditions and are dependent on the way treatment that will be provided to them. Patients undergo various physiological tests such as checking vital signs and bloodwork, to decide if the patient is at immediate risk of dying if not treated actively.
In last few years, various models have been developed to predict mortality in ICU, depending on the different health factors seen during their stay. Various models either have drawback with the patient subpopulation or have limited data for testing and training.
But, in the paper presented by MIT researchers at Proceedings of Knowledge Discovery and Data Mining conference, showed that machine-learning model by them can perform better in both the conditions. It has the ability to share data across all subpopulation to get better predictions and trains specifically on patient subpopulation. Thus, by having both the functions the model can predict patient’s risk of mortality within the first two days of their stay in ICU. The model has been compared to strictly global and other models.