Purpose: To develop a disease progression model of the clinical course of disease in patients with pneumonia-related sepsis that represents hospitalized patients as they progress through varying states of sepsis until death or hospital discharge.

Methods: We describe an empirically based, Monte Carlo microsimulation model used to represent individual patients with pneumonia-related sepsis and their clinical condition over time, enhancing realism beyond that practical in Markovian, state-based models. Patients are characterized by level of illness as represented by the Sepsis-related Organ Failure Assessment (SOFA) scores rather than course categories such as sepsis, severe sepsis and death. Missing data in the study was interpolated through rules based on clinical expertise. Individual patient histories from the Genetic and Inflammatory Markers in Sepsis (GenIMS) trial were decomposed into overlapping 3-day triplets containing clinical information and serial SOFA scores (and SOFA components) creating groups of short-term histories indexed by time in hospital. Simulated histories were constructed by recursively finding a patient triplet that was “similar” in terms of level and/or direction of SOFA score and using that triplet's next day values to update the index patient. By retaining information on the day of hospitalization, identical SOFA scores may imply different prognoses at different times during hospitalization. The empiric methods modes the longitudinal progression of component SOFA scores dependently, maintaining the inherent associations found in the progression of sepsis across organ systems. Longitudinal records of deaths, discharges and SOFA scores were compared between the average of 100 model runs and actual data with the Kolmogorov-Smirnov distribution test.

Results: Models that incorporating the duration of hospitalization (non-stationary) and that incorporate both the level and direction of SOFA generally perform better than those that do not: model-based estimates of deaths and SOFA at death were different from actual (p=0.16) but discharges, SOFA at discharge, and SOFA of remaining cohort were not., whereas all parameters except the number of discharges were significant when models did not include measures of progression or history, but matches only on level of SOFA score.

Conclusions: Sepsis is a non-stationary disease and it is more appropriate to model this disease incorporating time and the duration of hospitalization. Future work will incorporate diagnostic and therapeutic decisions into our simulation model that will allow for comparison of various treatment alternatives.