Predicting organ failure at 24 hours from early clinical data in an ovine pneumonia-sepsis model

Introduction

Having the ability to predict organ failure could impact treatment decisions and potentially lessen the consequences of acute lung injury (ALI) and sepsis. While Acute Physiology and Chronic Health Evaluation (APACHE II) and the Pneumonia Severity Index are useful in predicting the risk of morbidity and mortality, they do not predict the risk of developing organ failure. Currently no accepted practice allows for the prediction of the development of organ failure. The objective, therefore, of our study was to predict the development of organ failure at 24 hours using only the data available from the first 4 hours post inoculation.

Methods

This pneumonia-sepsis model included 19 sheep with ALI. Inoculation of ~2.5 × 10¹¹ colony-forming units methicillin-resistant Staphylococcus aureus (MRSA) induced pneumonia, while smoke injury was created through inhalation of cotton smoke. Four different groups were studied and are as follows: MRSA and smoke inhalation (M+S, n = 7), MRSA untreated (M, n = 3), MRSA treated (M+T, n = 3), and smoke inhalation only (S, n = 6). In order to use the injury group as a model input, all the sheep were modeled independent of group and a rank order of severity was determined. Additional inputs included a number of clinical and laboratory parameters. Only the first 4 hours of data were allowed to be used as an input. The model outputs were prothrombin time (PT) and mean arterial pressure (MAP) over the entire 24-hour time frame. To minimize overparameterization, only two inputs per output were used for prediction.

Results

The rank order of injury group from least to greatest severity was M+T, S, M, M+S. PT was best predicted by calcium and injury. The agreement between predicted and measured PT using only calcium as the input was r² = 0.24. Adding the second input, in this case injury group, improved the model's predictive ability (r² = 0.48). MAP was best predicted by lactate with an agreement between predicted and measured of r² = 0.64. Unlike PT, the model was not able to better predict MAP by adding a second input (r² = 0.64).

Conclusions

Our model was able to provide an accurate prediction of MAP using only the first 4 hours of data, while PT was less accurately predicted. However, this early study suggests that continued refinement of the progression model could provide a viable tool to predict organ failure in sepsis.