The obesity paradox and nutrition in critically ill patients
Survival analysis and multivariable modelling in a multicentre ICU study to assess obesity, mortality and nutritional support
A controversial clinical hypothesis that only a rigorous analysis could test
Does obesity improve survival in critically ill patients? The phenomenon, known as the "obesity paradox", has been debated for years. Testing it meant going beyond comparing means between groups.
The study drew on data from a prospective multicentre registry of ICUs across Spain, covering patients who required artificial nutritional support. The question had a catch: differences in survival between body mass index groups may be due not to obesity itself but to unadjusted baseline characteristics, and the effect of nutritional support — dose and timing of initiation — overlaps with all of this.
Answering it properly required three things that a superficial analysis does not provide: adjusting for confounders, modelling time to event in a way that respects the follow-up structure, and checking the assumptions of each model instead of taking them for granted. The aim was not to confirm the paradox, but to determine rigorously whether the data supported it.
525 patients
from a prospective multicentre registry in ICUs across Spain, stratified by body mass index
3 approaches
multivariable analysis, logistic regression and survival analysis, checked against one another
1 critical assumption
proportional hazards tested with Schoenfeld residuals, with a parametric alternative evaluated when it did not hold
Three families of methods on the same question, with every assumption checked
Multivariable analysis and associated factors
Characterisation of the body mass index groups and analysis of associated factors through adjusted comparisons. Multivariable modelling with regularisation and variable selection methods (ridge, lasso, elastic net and stepwise selection) to assess which factors explained the differences between groups, with reproducible preprocessing and missing data imputation.
Mortality prediction and comparison with machine learning
Logistic regression models to estimate the probability of death from the clinical and nutritional variables, evaluated by sensitivity and specificity. Those models were compared against a battery of machine learning algorithms, which did not provide a substantial improvement over classical statistical methods on these data.
Survival analysis with assumption checking
Estimation of survival by subgroup with Kaplan-Meier and the log-rank test, followed by Cox modelling. When Schoenfeld residuals showed that one covariate violated the proportional hazards assumption, the Cox model was ruled out as the final model and a parametric Weibull alternative was evaluated, independently checking how well it fitted the data.
A solid answer to a controversial hypothesis, and interpretable predictive models
A clear answer, even if not the one expected
The analysis found no statistically significant differences in survival between body mass index groups, and no signs supporting the obesity paradox in this cohort. It was a negative result, but a consistent one: the question was assessed on a solid methodological basis.
Prediction and survival models that do fit
Beyond body mass index, the work delivered a logistic regression model, with internally assessed performance, to estimate the probability of death in the ICU from a reduced set of variables, and a parametric survival model whose fit to the data was independently checked. Analytical results that can be reused regardless of the starting hypothesis.
Findings on nutritional support
The analysis showed differences in calorie and protein requirements between body mass index groups, with obese patients receiving an intake that deviated from their theoretical requirements. A clinically relevant result for reviewing the adequacy of nutritional support in patients with obesity, separate from the question of the paradox.
What this case shows about the way we work
Rigour lies in checking, not confirming
What set this work apart was not choosing the fashionable model, but checking its assumptions. When the Cox proportional hazards assumption did not hold, it was not ignored: the Cox model was ruled out as the final model and a parametric alternative was evaluated, checking its fit to the data instead of taking it for granted. That discipline is what separates a defensible result from one that falls apart at the first review.
A well-conducted negative result is still a result
We did not force the data to say what the hypothesis expected. The fact that the obesity paradox did not appear in this cohort adds evidence to a controversy that is still open. The value lies in the reliability of the answer, not in its sign.
Several methods, one question
Approaching the same question through multivariable analysis, logistic regression and survival analysis was not about piling up techniques, but about testing how stable the conclusion was across complementary approaches. That convergence reduces dependence on a single analytical approach and strengthens the consistency of the conclusion.
Trabajar con Julia es muy fácil: lo entiende todo a la primera y es capaz de transmitir respuestas fáciles a preguntas mucho más complejas. Además, querría señalar que detrás de esta gran profesional hay una mejor persona, que no sólo ha demostrado una gran capacidad, sino que a nivel humano ha entendido desde un primer momento la importancia de nuestra área de investigación, mostrándonos su apoyo y entusiasmo incondicional durante toda la colaboración.
Dr Juan Carlos López Delgado
Intensive Care Medicine Specialist at Hospital Universitari de Bellvitge and Clinical Researcher at IDIBELL
Resulting scientific output
Relationship between the obesity paradox and enteral nutrition in critically ill patients
Capabilities deployed in this project
Shall we talk?
Tell us what you are working on and we will give you a straight answer on how we can help.