Prediction of anticoagulation during hemodialysis by population kinetics and an artificial neural network

All systems currently used for routine hemodialysis require heparin administration to prevent blood clotting in the extracorporeal circuit. We tested the hypothesis that population-based statistical techniques can be used to predict heparin concentrations during routine hemodialysis. Two predictive models were developed, one based on nonlinear mixed effects modeling (NONMEM) and the other on a multilayer perceptron neural network. Serial clotting time data were obtained from forty-nine patients and used to develop the models. The models were used to predict the clotting times of 70 patients in a prospective test. We determined that the neural network provided greater precision, had fewer outliers in its predictions, and did not have the model misspecification in bolus administration that the NONMEM predictions demonstrated. A final NONMEM model was developed using all data from 119 patients to identify important covariates for predicting the heparin pharmacodynamic parameters, volume of distribution, and clearance. Both the volume of distribution and clearance increased following the initiation of dialysis and as the patient's baseline clotting time increased. The volume of distribution also increased as the patient's weight increased but was decreased by smoking and diabetes. Population-based statistical techniques may provide a useful alternative to existing methods for prescribing heparin.