We carried out this ancillary study in three intensive care units (ICUs; ClinicalTrials.gov, identifier NCT03292120). The initial study was approved by the Ethics Committee of the Nice University Hospital, France (agreement number 2016-A00533–48). The requirement for written informed consent was waived due to this study’s strict observational design according to French law 17. However, we obtained the patient or relatives consent for using these data.
As reported previously in our first trial, we screened all patients with septic shock during their first 6 hours after ICU admission from June 2016 to November 2018. Septic shock was defined according to the international sepsis definitions available at the time of the study 18. We excluded patients in whom septic shock was diagnosed after 6 hours, those younger than 18 years old, pregnant women and patients with a do-not-resuscitate order. Moreover, patients who experienced premature death within 24 hours were excluded from the analysis, as the prognostic factors were consequently irrelevant. The three ICUs were organized as described by Leone and al. 19, and their clinical practices were in agreement with international and French national guidelines.
Measurements and study design
We prospectively collected the patients’ demographic variables from their electronic charts. Time 0 (T0) was set as soon as hemodynamic monitoring was set, that is, 1 hour after ICU admission.
The data were gathered upon ICU admission (T0), 6 hours after admission (T6), 12 hours after (T12), 24 hours after (T24), 48 hours after (T48) and 72 hours after (T72). Hemodynamic variables were included as in our previous study 20. The Simplified Acute Physiology Score (SAPS II) was collected upon admission 21. Sequential Organ Failure Assessment (SOFA) score was calculated at T0 and T48, as was 28-day mortality rate 1,22. Clinical worsening was defined as an increase in SOFA score ≥ 1 within 48 hours (ΔSOFA ≥ 1).
Several variables were included in the univariate analysis to assess their association with ΔSOFA ≥ 1 and 28-day mortality rate: age, gender, body mass index, SAPS II, SOFA, mean arterial pressure (MAP), diuresis, fluid balance, norepinephrine dose, serum lactate concentration, P(v-a)CO2, serum lactate concentration clearance from T0 to T6, ScvO2, continuous cardiac index (CCI) and adjuvant inotrope treatment. Then, the multivariate analysis was built.
Then, we determined different phenotypes using five variables measured during the patients’ ICU stay: norepinephrine dose, serum lactate concentration, MAP, P(va)CO2, CCI (Table 3). We recognized three total phenotypes: safe, intermediate and poor. The safe phenotype encompassed patients with favourable outcomes, as opposed to the intermediate and poor phenotypes (Table 3). The patients were classified into one of the phenotypes at three timepoints: T0, T6 and T24. Of note, each patient’s phenotype pattern could evolve during these three timepoints.
To this purpose, we used the multi-profile hidden Markov model (HMM) 23,24. The HMM allowed us to investigate the dynamics of the patients’ trajectories during their ICU stay. The HMM is a statistical model that can be used to describe the evolution of observed events which depend on internal factors that cannot be directly observed. It refers to the observed event as a symbol and the hidden factor underlying the observation as a state. The HMM also consists of two stochastic processes, namely, an invisible process of hidden states and a visible process of observable symbols. The hidden states form a Markov chain, and the probability distribution of the observed symbol depends on the underlying state. For this reason, the HMM is also called a doubly embedded stochastic process 25. Thus, the HMM aims to recover hidden groups or patterns from observed data. It is similar to clustering techniques, but this method is more flexible. Moreover, the HMM provides transition probabilities overtime across different profiles to show the probability of transitioning from profile “I” to profile “J”.
Modelling observations in these two layers, one visible and the other invisible, is particularly useful, since many real-world problems involve classifying raw observations into a number of categories, or class labels that are more meaningful to us. The HMM has therefore been applied in a variety of fields, such as diabetes 26, breast cancer 27 and public health 28.
For example, in our study, the observed data were the relevant variables of septic shock patients recorded during their first 24 hours after ICU admission. The hidden groups (states) were the patients’ latent phenotypes. The variables included in the HMM were primarily based on domain knowledge and a literature review, in addiction with those collected for our study as defined previously 15,20,29,30. We chose to build the HMM for 3 phenotypes for it was the most relevant for the patient outcomes analysis.
Finally, we used two outcomes: favourable and poor, with the latter defined as the persistence of the intermediate or poor phenotypes after T0.
The primary outcome was to assess the variables associated with 28-day mortality rate at T0 and T6. The associations between the variables and ΔSOFA ≥ 1 were then studied to assess the secondary outcomes. Moreover, we examined the clinical course of our three phenotypes (safe, intermediate and poor) and two probability outcomes (favourable and poor).
Additionally, we identified cut-off values for the clinical and biological variables at T0 to predict a favourable or poor outcome. Then, we developed an algorithm including these variables to facilitate clinical management in the early stages of septic shock.
We expressed continuous variables as mean (standard deviation) or median values (interquartile range) as appropriate, and compared between the different septic shock phenotypes using analysis of variance (ANOVA), Student’s t-test for continuous variables and the Pearson Chi-square test or Fisher’s exact test for discontinuous variables. The associations between the variables and potential risk factors were initially assessed using univariate analysis. The variables associated with 28-day mortality rate and/or ΔSOFA ≥ 1 in the univariate analysis, with a p-value < 0.05 in at least one comparison, were included in the multivariate analysis. Odds ratios (OR) were displayed with a 95% Confidence interval (CI). All statistical analyses were performed using DMGM software and IBM SPSS v20. A p-value less than 0.05 was considered statistically significant.