Assessing clinical outcomes of nosocomial pneumonia patients with a pharmacometric multistate model
Diego Vera-Yunca, Lena E. Friberg
Department of Pharmacy, Uppsala University, Uppsala, Sweden
Objectives: In the field of antibiotics, pharmacometric methods are usually applied to preclinical data that mostly contains longitudinal information. However, when clinical data is analysed, main outcomes (clinical or microbiological cure/failure) lack an assessment of longitudinal information. The COMBINE project, part of the AMR Accelerator programme from Innovative Medicines Initiative (IMI), aims to develop or improve tools to translate preclinical results into clinical outcomes more reliably. In particular, the goal of this work was to build a multistate model for pneumonia clinical data that assess the relationship of clinical outcomes (cure/not cured) over time (at end-of-treatment and at end-of-study) and disease progression by evaluating the effect of early predictors on the transitions between clinical states.
Methods: Data from a phase IV clinical trial comparing linezolid and vancomycin was available[1]. A multistate modelling approach was applied, which considered states that corresponded to the clinical outcomes: either failure (S1) or cure (S2) at end-of-treatment and end-of-study, and death (S3). Patients were assigned to state S1 at baseline and then they were allowed to transit to S2 and vice versa. Besides, patients can die at any time point by transiting to the death state (S3). Transition rates λij from state i to state j were used to describe probabilities of patients transiting between states over time. Those rates were modelled as parametric hazard functions (exponential or Weibull). When the probability of death was computed, patients without a death event were considered as censored.
Baseline demographics or disease-related covariates were tested as predictors on the transition rates: Minimum Inhibitory Concentration (MIC) for either linezolid or vancomycin (depending on the study arm the patient belonged to), sex, age, weight, creatinine clearance (calculated using the Cockcroft-Gault equation, in mL/min), Clinical Pulmonary Infection Score (CPIS), Acute Physiology and Chronic Health Evaluation II (APACHE II) score and total white blood cell counts (cells/mm3). Model selection was based upon the objective function value (OFV) and inspection of visual predictive checks (VPCs).
Results: A total of 329 patients with 896 observations were analysed. The transition from failure to cure (λ12), cure to failure (λ21) and cure to death (λ23) followed an exponential function, while failure to death (λ13) was best explained by a Weibull function. A step function described patients during the treatment period up to 8.5 days. Moreover, the transition rates from failure to cure (λ12) and from cure to failure (λ21) up to the end of the treatment period were set to the same value as no significant drop in OFV was found. Besides, due to the low number of patients that transited from cure to death during the treatment period, the transition from cure to death (λ23) was set to zero.
The transition rate from failure to death (λ13) was 14 times higher than from the cure state to death (λ23). A high APACHE II score (29) with respect to the median value (17) decreased the probability of going from failure towards cure (λ12) [hazard ratio (HR)=0.72, 95%CI=0.56-0.92] and increased the probability of dying for patients who had been defined as cured (λ23, HR=3.57, 95%CI=1.43-8.89). Older patients (in the 97.5th percentile, 86 years) showed a higher λ23 (HR=2.94, 95%CI=2.61-3.32) compared to the median age (65 years). Low creatinine clearance values (in the 2.5th percentile, 17 mL/min) increased λ13 (HR=1.61, 95%CI=1.16-2.24) in comparison to the median (85mL/min). Model evaluation by VPCs showed good description of the data.
Conclusions: The developed multistate model successfully described pneumonia clinical outcomes, showing that the time course of death follows different functions depending on the patient state. Significant early predictors related to demographics or patient status were also found. This is an initial step towards a framework which eventually aims to translate quantitative drug effect information (i.e., PKPD-predicted bacterial load) from preclinical results to improve design and prediction of clinical trials.
References:
[1] Wunderink RG, Niederman MS, Kollef MH, et al. Linezolid in Methicillin-Resistant Staphylococcus aureus Nosocomial Pneumonia: A Randomized, Controlled Study. Clin Infect Dis. 2012;54(5):621-629. doi:10.1093/cid/cir895
