Master thesis : Personalised Nutrition vs. Standardised Guidelines in Critical Care: A Target Trial Emulation of the STAR Protocol (including introduction to research methodology)

Abstract

In the intensive care unit (ICU), the importance of glycaemic control for the survival of critically ill patients has been demonstrated, as a result of the common occurrence of stress-induced hyperglycaemia. While the main actor of this glycaemic control is insulin, another important component for critically ill patients is the nutritional support. Historically, the paradigm of clinical nutrition has often relied on standardised recommendations (a "one-size-fits-all" approach), but the concept of personalised nutrition has recently emerged, aiming to tailor the caloric intake to the evolving metabolic state and tolerance of each individual patient. In this context, the Stochastic TARgeted (STAR) protocol embodies this shift toward highly personalised management. Designed as a combination of physiological and stochastic models clinically validated for glycaemic control, STAR continuously evaluates patient-specific insulin sensitivity. By employing stochastic forecasting to anticipate future patient evolutions, the algorithm recommends a combination of insulin and nutrition. This approach optimises caloric delivery for each patient while ensuring the maintenance of strict normoglycaemia, thereby illustrating a specific approach for personalised nutrition in critical care. The objective of this work is to causally evaluate the impact of several standardised nutritional strategies (such as the ESPEN guidelines or aggressive early feeding) compared to the unrestricted and adaptive management of the STAR protocol. While conducting a randomised controlled trial is the gold standard to answer such causal questions, this study uses observational data. Utilising retrospective clinical data, several challenges have to be overcome, such as the verification of the identifiability conditions inherent to all causal inference studies. However, the main concern arises from the lack of randomisation encountered in observational studies. Consequently, the methodology of this work is rooted in the rigorous Target Trial Emulation framework. Regarding the statistical methods, the parametric g-formula was implemented to model these interventions while adjusting for baseline and time-varying confounders and accounting for competing events. The results demonstrate that restricting the STAR protocol to strictly follow international guidelines (formulated by the ESPEN) offers no significant advantage over the unrestricted STAR protocol, thereby validating the efficacy and safety of the unrestricted STAR protocol. Furthermore, the study confirms the clinical risks associated with an early high-nutrition strategy, which led to an increase in mortality at both the ICU and hospital levels. Finally, a comprehensive analysis of the starvation scenario highlighted the mathematical limitations of causal inference algorithms when faced with a lack of longitudinal data (positivity violations), underscoring the importance of rigorous clinical interpretation when applying predictive models.