Abstract Background Patients with severe acute kidney injury (AKI) who require continuous venovenous hemodiafiltration (CVVHDF) in intensive care unit (ICU) are at high mortality risk. Little is known about clinical biomarkers for risk prediction, optimal initiation, and optimal discontinuation of CVVHDF. Methods This prospective observational study was conducted in seven university-affiliated ICUs. For urinary neutrophil gelatinase-associated lipocalin (NGAL) and plasma IL-6 measurements, samples were collected at initiation, 24 h, 48 h after, and CVVHDF discontinuation in adult patients with severe AKI. The outcomes were deaths during CVVHDF and CVVHDF dependence. Results A total number of 133 patients were included. Twenty-eight patients died without CVVHDF discontinuation (CVVHDF nonsurvivors). Urinary NGAL and plasma IL-6 at the CVVHDF initiation were significantly higher in CVVHDF nonsurvivors than in survivors. Among 105 CVVHDF survivors, 70 patients were free from renal replacement therapy (RRT) or death in the next 7 days after discontinuation (success group), whereas 35 patients died or needed RRT again (failure group). Urinary NGAL at CVVHDF discontinuation was significantly lower in the success group (93.8 ng/ml vs. 999 ng/ml, p