Causal inference using observational intensive care unit data: a systematic review and recommendations for future practice

AimTo review and appraise the quality of studies that present models for causal inference of time-varying treatment effects in the adult intensive care unit (ICU) and give recommendations to improve future research practice. MethodsWe searched Embase, MEDLINE ALL, Web of Science Core Collection, Google Scholar, medRxiv, and bioRxiv up to March 2, 2022. Studies that present models for causal inference that deal with time-varying treatments in adult ICU patients were included. From the included studies, data was extracted about the study setting and applied methodology. Quality of reporting (QOR) of target trial components and causal assumptions (ie, conditional exchangeability, positivity and consistency) were assessed. Results1,714 titles were screened and 60 studies were included, of which 36 (60%) were published in the last 5 years. G methods were the most commonly used (n=40/60, 67%), further divided into inverse-probability-of-treatment weighting (n=36/40, 90%) and the parametric G formula (n=4/40, 10%). The remaining studies (n=20/60, 33%) used reinforcement learning methods. Overall, most studies (n=36/60, 60%) considered static treatment regimes. Only ten (17%) studies fully reported all five target trial components (ie, eligibility criteria, treatment strategies, follow-up period, outcome and analysis plan). The treatment strategies and analysis plan components were not (fully) reported in 38% and 48% of the studies, respectively. The causal assumptions (ie, conditional exchangeability, positivity and consistency) remained unmentioned in 35%, 68% and 88% of the studies, respectively. All three causal assumptions were mentioned (or a check for potential violations was reported) in only six (10%) studies. Sixteen studies (27%) estimated the treatment effect both by adjusting for baseline confounding and by adjusting for baseline and treatment-affected time-varying confounding, which often led to substantial changes in treatment effect estimates. ConclusionsStudies that present models for causal inference in the ICU were found to have incomplete or missing reporting of target trial components and causal assumptions. To achieve actionable artificial intelligence in the ICU, we advocate careful consideration of the causal question of interest, the use of target trial emulation, usage of appropriate causal inference methods and acknowledgement (and ideally examination of potential violations) of the causal assumptions. Systematic review registrationPROSPERO (CRD42022324014)