Causal Bayesian Machine Learning for Assessing Heterogeneity of Treatment Effects by Dexamethasone Dose for Patients with COVID-19 and Severe Hypoxemia

This is a post-hoc exploratory analysis of the COVID STEROID 2 trial.⁷This was conducted according to a statistical analysis plan written after the trial’s pre-planned analysis was reported, but before the analysis reported in this manuscript was conducted (https: //osf.io/2mdqn/). This manuscript was presented following the Enhanced Reporting of Observational Studies in Epidemiology (STROBE) checklist.¹²Bayesian analysis reported according to the Bayesian reporting used in clinical studies (ROBUST) guidelines¹³.

Why Use the BART Methodology?

HTE means that due to inter-individual variability, some individuals respond differently (i.e., better or worse) than others who receive the same treatment. Most trials are designed to assess mean treatment effect. This is a summary of all individual effects for the test samples (see Supplementary Appendix for technical details). Her traditional HTE method examines patient characteristics one her at a time and attempts to identify differences in treatment effects according to individual variables. This approach may have limitations due to its low potency (due to multiple test adjustments) and its failure to take into account the fact that many traits under investigation are correlated and potentially synergistic. Well known. As a result, using conventional HTE approaches may miss more complex relationships between variables that better define individuals and thus better inform our understanding of variability in treatment response. Therefore, these data and statistical modeling challenges must be addressed to identify a true, clinically meaningful her HTE. BART is inherently an attractive method for this task. Algorithms hierarchically automate the detection of nonlinear relationships and interactions based on relationship strengths, reducing researchers’ discretion when analyzing experimental data. This approach also avoids model misspecifications and biases inherent in traditional interaction testing procedures. As we do here, BART can also be deployed within a counterfactual framework for studying HTE.^11,14,15which has shown superior performance to competing methods in extensive simulation studies^16,17These features make BART an attractive tool for investigators to explore HTE and inform more broadly future confirmatory HTE analysis in trials and hypothesis generation. Therefore, in this analysis, BART was used to assess the presence of multivariable HTE and to estimate conditional mean treatment effects across meaningful subgroups in the COVID STEROID 2 trial.

COVID STEROID 2 study

COVID STEROID 2 study⁷ was an investigator-led, international, parallel-group, stratified study conducted between 27 August 2020 and 20 May 2021 at 31 sites in 26 hospitals in Denmark, India, Sweden and Switzerland. It was a blinded, randomized clinical trial.^7,18This study was approved by the regulatory authorities and ethics committees of all participating countries.

This study included COVID-19 and severe hypoxemia (>10 L/min of oxygen, use of noninvasive ventilation (NIV), continuous use of continuous positive airway pressure (cCPAP), or invasive mechanical 1,000 adult patients admitted to the hospital with active ventilation (IMV) were enrolled. ). The patient reported primarily that he had previously used systemic corticosteroids for her COVID-19 for more than 5 days, that consent was not obtained, and that he had not received high doses for his non-COVID-19 indications. Excluded due to use of corticosteroids.^4,17Patients were randomized 1:1 to intravenous dexamethasone 12 mg/day or 6 mg/day once daily for up to 10 days.Additional details are provided in the primary protocol and study report^7,18.

The study protocol was institutionally approved by the Danish Medicines Agency, the Danish Metropolitan Ethics Committee and each study site. This trial was overseen by the Intensive Care Research Cooperation and the George Institute for International Health. A data and safety monitoring board oversaw the safety of study participants and conducted one planned interim analysis. Informed consent was obtained from the patient or his legal representative in accordance with national regulations.

reaserch result

(1) DAWOLS on day 90 (that is, the number of days observed without IMV, circulatory support, and renal replacement therapy and without assigning the worst possible value to a patient who died), and (2) 90 Dead for days. Binary mortality results were used to match the primary trial analyses. Time-to-event results also tend to be less robust in her ICU trials in general.¹⁹We chose DAWOLS on Day 90 instead of the study’s primary outcome (DAWOLS on Day 28) to match other analyzes of the study that sought to examine results over time. Both outcomes were evaluated in the complete intention-to-treat (ITT) population. This was her 982 after excluding patients without consent for data use.⁷There was no formal sample size calculation for this study, as the sample size is fixed.

Preselected prognostic HTE factors

BART is a data-driven approach that can scan for interdependencies between any number of factors, but pre-selected We only examined the heterogeneity of a set of factors. Preselected variables that were included in this analysis are listed below with the scales used in parentheses. Continuous covariates were standardized to have a mean of 0 and a standard deviation of 1 before analysis.Detailed variable definitions are available in the study protocol¹⁸.

1. 1.

   age of participants (continuous),

2. 2.

   limitations of care (yes, no),

3. 3.

   Level of respiratory support (open system vs. NIV/cCPAP vs. IMV)

4. Four.

   interleukin-6 (IL-6) receptor inhibitors (yes, no),

5. Five.

   When using dexamethasone for up to 2 days and when using 3-4 days before randomization,

6. 6.

   participant weight (continuous),

7. 7.

   diabetes mellitus (yes, no),

8. 8.

   ischemic heart disease or heart failure (yes, no),

9. 9.

   chronic obstructive pulmonary disease (yes, no), and

10. Ten.

    – Immunosuppression (yes, no) within 3 months prior to randomization.

statistical analysis

HTE was assessed on an absolute scale (that is, mean day difference in 90-day DAWOLS count and risk difference in 90-day mortality).The analysis was divided into two stages^14,20,21,22In the first stage, conditional mean treatment effects were estimated according to each participant’s covariates using the BART model. DAWOLS results were treated as continuous variables and analyzed using standard His BART, whereas binary mortality results were analyzed using Logit BART. In the second step, a “fit-the-fit” approach was used to identify different treatment effects for covariate-defined subgroups using the estimated conditional mean treatment effect as the dependent variable of the model. rice field.This second stage used classification and regression tree models^{twenty three}, the maximum depth was set to 3 as an a posteriori decision to aid interpretability. Because fit-the-fit reflects estimates from his BART model, the resulting overall treatment effects (such as risk differences) differ slightly from the raw trial data.

BART models are often fitted using a total of 200 trees and specifying a base prior of 0.95 and a power of 2 prior. This penalizes the net branch growth within each tree.¹⁵These default hyperparameters tend to work well in practice, but were likely suboptimal for this data. Therefore, the hyperparameters were evaluated using 10-fold cross-validation to compare the predictive performance of the models under 27 pre-specified possibilities. That is, all combinations of power priors equal to 1, 2, or 3 and basic priors equal to 0.25, 0.5, or 0.95. , and the number of trees equal to 50, 200, or 400. Prior probabilities corresponding to the smallest cross-validation error were used in the final model. Each model used a Markov Chain Monte Carlo procedure consisting of 4 chains with 100 burn-in iterations each and a total of 1100 iterations. The posterior convergence of each model was assessed using the diagnostic procedure described in Sparapani et al.^{twenty four}Model diagnostics were good for all models.All parameters appeared to converge within the burn-in period and Geweke’s convergence diagnostic z-score^{twenty five} It was almost normal. All BART models were fitted using the R statistical computing software v. 4.1.2.²⁶ Using “BART” package v. 2.9^{twenty four}all CART models were adapted using the ‘rpart’ package v. 4.1.16²⁷.

Analyzes were performed under the ITT paradigm. Compliance issues were considered minimal. As with the primary analysis of trials, the small amount of missing outcome data was ignored in the primary analysis. Sensitivity analyzes were performed under best/worst and worst/best case imputations. For best/worst case imputation, all missing mortality outcome data for the 12 mg/day group were set to 90-day survival and all missing mortality outcome data for the 6 mg/day group were set to 90-day mortality. , then repeated the entire estimation procedure. day to day. All days with missing life support data were then set to survival without life support in the 12 mg/d group and vice versa in the 6 mg/d group. Under the worst/best case imputation, the estimation procedure was repeated under opposite conditions. For example, set all missing mortality outcome data in the 12 mg/day group to death at 90 days, and set all missing mortality outcome data in the 6 mg/day group. The group lives in 90 days.

The decision trees (for the 90-day mortality results and for the 90-day DAWOLS results) from each of the above fitting analyzes were output (non-standardized continuous variables, i.e. original scale). Similarly, decision trees for each outcome after best-case and worst-case imputation were output for comparison with the full record analysis. All statistical code is available at https://github.com/harhay-lab/Covid-Steroid-HTE.