Researchers have introduced a novel approach to harnessing medical data to estimate when patients in intensive care units may be approaching death. The study involved analyzing data from 54,000 adult ICU patients, incorporating vital signs, laboratory results, electrocardiograms, chest X-rays, and notes from physicians. All of this information, including textual data, was transformed into images and subsequently condensed into a singular "visual memory" of each patient, represented as a vector comprising 1,024 numerical values. With this representation, the AI system aimed to forecast changes in a patient's condition over the following 3 to 48 hours. The model excelled at predicting mortality within a 48-hour window, achieving a score of 0.852, where a score of 0.5 indicates random chance and a score of 1 signifies perfect accuracy. Additionally, the AI was capable of foreseeing risks associated with acute kidney injury and organ dysfunction, as well as predicting the timing of ICU discharge. However, the significance of this research extends beyond the AI's ability to predict death with high accuracy, which a traditional statistical model can also achieve. Rather, the study highlights that a single AI model can simultaneously process vastly different types of data. In principle, this means that even more test results could be integrated, resulting in a comprehensive predictive output.
Informational material. 18+.
