Continuous urea monitoring in hemodialysis: a model approach to forecast dialytic performance. Results of a multicenter study

Background: A urea biosensor, inserted into the ultrafiltrate collection-line of paired filtration dialysis (PFD), not only allows on-line dialysis quantification, but also forecasts final (Cend) and 30 min equilibrated urea concentration (Ceq), the most reliable value for calculating dialysis efficiency. The urea biosensor processes plasma ultrafiltrate continuously, delivering a large amount of data to the computer, which estimates the parameters by a mathematical model, thus predicting the whole urea profile with rebound. Methods: A multicenter randomized trial on 41 patients was conducted to ascertain the ability of a two-pool variable-volume urea model to forecast Cend and Ceq at 60 and 90 min after the start of dialysis. Two alternative dialytic treatments, A or B, were chosen, the latter being more efficient. Each treatment included six serial PFD. The accuracy of forecasting was evaluated through four indices based on forecast errors, calculated as the difference between observed and forecasted urea values: mean percent error (MPE) (%), mean absolute deviation (MAD) (mg/dl), mean absolute percent error (MAPE) (%) and root mean squared error (RMSE) (mg/dl). Results: Forecasted urea concentrations were lower than those measured by the biosensor. WE for Cend was negligible in A (+1.2%) and much higher in B (+7.2%); both values improved at 90 min, +1.0% and +5.8%, respectively. MAD for Cend was similar in both treatments and improved slightly at 90 min, ranging from 4.9 to 5.9 mg/dl. MPE for Ceq was +4% in A and and more than doubled in B (+11.5%); both values improved at 90 min, +3.7% and +9.7%, respectively. MAD for Ceq was 7.5 mg/dl in A and 8.5 mg/dl in B; both improved at 90 min, 6.7 and 7.4 mg/dl, respectively. The other indices, MAPE and RMSE, showed similar results. Comparison between the errors of the two treatments with analysis of variance (ANOVA) for repeated measures gave no significant results. Conclusions: Our model forecasts of urea concentrations were overall lower than the measured ones: the bias was negligible for A-Cend, greater for the A-Ceq and when the more efficient treatment B was used. The 60 min predictions improved at 90 min. The comparison between the prediction errors in the two treatments were not statistically significant. The recirculation measurement would probably reduce the bias if it were properly incorporated into the model.