Long ago, Boyd et al. used logistic regression to develop the first trauma and injury severity score (TRISS) for predicting the mortality of injured patients [1]. However, determining how to develop a reliable daily prediction model of survival at an intensive care unit (ICU) is still a complex problem. How should the existing clinical and dynamically changing pathophysiological conditions of each patient be factored into the daily prediction equation to produce an estimate of likely survival or mortality in different ICUs?

The authors of this paper investigate the reliability of using variables that fluctuate daily in discrete-time event forecast models to estimate the likelihood of mortality or survival of individual patients admitted to ICUs. Initially, they compute the change in blood count ratios and variation types over consecutive days. Next, they use the appropriate Chi square test and student t-test to compare the daily hemoglobin, white blood cell and platelet counts, and the blood count change and variation ratios of surviving and dying patients, in order to identify time-dependent covariate variables for data analysis and modeling with a reputable logistic regression analysis [2].

The results from the two daily mortality prediction models exhibit decent standardization and percipience of patients dying over various days in an ICU. This research highlights: the need to collaborate with biostatisticians to validate clinical data, the importance of eradicating outliers in patient data from logistic regression, and the relevance of eliminating colinear variables of death from a reliable probability model. Given that the probability of mortality has a wide range with a significant increase over time for some cohort of patients, the need exists to explore new reliable prediction models that consider the daily changes in the conditions of ICU patients as transient factors. The authors recognize the limited generalizability of the mortality prediction model results, due to the lack of external validation with a variety of facilities. I recommend that biostatisticians and bioinformatics researchers read this paper and weigh in on its insightful mortality prediction models.