Dynamic Temperature Modeling Predicts Mortality in Murine Sepsis

Body temperature generally correlates with mortality in mouse sepsis. However, single temperature measurements fail to accurately discriminate whether mice will subsequently live or die following a septic insult, limiting their usefulness as an endpoint in pre-clinical sepsis studies. Temperature trajectories using serial measurements have been demonstrated to have predictive capacity in septic patients. The purpose of this study was to determine if temperature trajectories could predict subsequent mortality in septic mice. A cohort of 511 C57Bl/6 mice from three different laboratories undergoing survival experiments had serial body temperature measured every 12 hours from baseline to 7 days following the onset of sepsis. To optimize generalizability, the cohort was intentionally heterogeneous, including 6-31 week old mice with an equal number of females and males undergoing three different models of sepsis with different comorbidities prior to sepsis, and different genetic variants. A training cohort comprised the first 80% (n=417) and a validation cohort comprised the final 20% (n=94). To quantitatively predict mortality risk following sepsis, a Bayesian joint model was developed incorporating longitudinal temperature measurements using data collected at baseline and 12 and 24 hours after the onset of sepsis to predict probability of death in the subsequent 12-72 hours. Predictive accuracy improved progressively with longer horizons with AUCs increasing from 0.864 at 12 hours (predicting 24–36 hour mortality) to 0.932 at 72 hours (predicting 24–96 hour mortality) in the training cohort. Model performance was not improved by incorporating type of sepsis model, age or sex as covariates. While specificity was >0.92 at all timepoints in the training cohort, it was only 0.796 at both 24 and 36 hours in the validation cohort. Since a specificity <1 means that some mice predicted to die would instead survive leading to errors in survival curve, an additional pre-specificed analysis was independently evaluated to identify irreversible terminal states where no mouse ended up surviving. At 24, 36 and 48 hours, temperature thresholds of 27.0 °C, 28.2 °C and 30.2 °C identified mice that would subsequently die with a specificity of 1. Temperature trajectories identifying longitudinal hypothermia therefore represent a powerful predictor of mortality in a large heterogeneous population of murine sepsis. Whereas a single temperature measurement has limitations in prediction of subsequent mortality except at extreme hypothermia, dynamic Bayesian modeling enables individualized mortality risk estimation as criteria that may be used for humane endpoint determination.