A retrospective study analyzed data from the eICU database (2014–2015) in 20,543 adult surgical ICU patients admitted through the emergency department. Change in BMI was calculated as BMI at discharge minus BMI at admission. The 28-day mortality in the ICU was 4.70%. Change in BMI showed a non-linear U-shaped association with mortality: the risk decreased with increasing BMI to a value of -1.75 kg/m², then increased sharply. Each additional kg/m² above this threshold increased mortality by 9% (HR 1.09, 95% CI 1.05–1.12, p < 0.0001). Patients in the quartile with the highest BMI gain had a 52% higher risk than those in the lowest quartile (HR 1.52, 95% CI 1.27–1.82, p < 0.0001). Dynamic BMI change outperformed static BMI or weight in prediction (AUC 57.9). A slight reduction in BMI (-1.75 kg/m²) was associated with the lowest risk of mortality.