Alena Simalatsar (1,2), Monia Guidi (2,3), Pierre Roduit (1), and Thierry Buclin (2)
(1) Institute of Systems Engineering, University of Applied Sciences and Arts - Western Switzerland, Sion, Switzerland, (2) Service of Clinical Pharmacology, University Hospital and University of Lausanne (CHUV), Lausanne, Switzerland, (3) School of Pharmaceutical Sciences, University of Geneva, University of Lausanne, Geneva, Switzerland
Objectives: The controlled delivery of intravenous (IV) anesthetic, such as propofol, aims at fast and safe achievement and maintenance of a suitable depth of hypnosis, by ensuring appropriate effect site (i.e. brain) exposure to the drug. Today, such drugs are regularly injected by Target Controlled Infusion (TCI) systems, piloted by an open-loop algorithm based on Pharmacokinetic (PK) models. However, clinical conditions may markedly alter propofol pharmacokinetics and actual concentrations could significantly differ from predicted ones, leading to important under- or over-exposure. The situation could be improved by closing the loop with sensors providing regular real measurements of the anesthetic concentration in body fluids. The aim of the present study was to develop a closed-loop algorithm based on the classic open-loop algorithm presented by Shafer et al [1] combined with a Kalman filter using real-time plasma measurements. We also perform stability analysis of this algorithm by accounting for realistic measurement noise, periods and delays.
Methods: Kalman filter is used to estimate the personalized plasma drug concentrations. The algorithm first uses the population PK model to produce a vector X’ of a priori estimates of the current concentrations in each compartment, with x’1 representing the plasma. Then it computes the estimate covariance P defining the model inaccuracy that we associate with the inter-patient variability of Eleveld et al [2]. Once a new measurement y is available, the measure residual d=x’1–y is computed. Kalman gain K, accounting for both model inaccuracy and measurement noise, is updated to compute a posteriori vector of concentrations X=X’+Kd. The estimate of K is updated with every new measurement. In turn, vector X is updated every second using the latest values of d and K to be used by Shafer’s [1] algorithm for continuous infusion rate adjustment.
A set of 1000 individual female patients (70 kg, 170 cm, 36 y) generated by the Eleveld et al model with inter-patient variability [2] and a fixed target concentration at 6 mg/L during 60 mins of surgical operation were chosen. To validate the algorithm the measurements for each individual were simulated with the corresponding inter-patient variability and two different intra-individual variability values (Eleveld et al 47% and twice smaller 23%) under the computed propofol infusion rate. The effect site Concentration-Time (CT) profiles for all individuals if they were administered the drug with the delivery rate computed using open-loop algorithm of classic TCI (Uav) and our algorithm (Uind) were computed. The 95% prediction intervals (PI95%) for the two different intra-individual variability values, various measurements periods (1, 5, 10, 15, and 30 sec) and delays between the time stamp of the blood sample and moment when it is available for processing (0, and 30 sec) were computed.
Results: The PI95% for individuals being administered with Uav, obviously, did not depend on measurement noise, period or delay. It remained constant in all experiments and exceeded the 40% accuracy area. In turn, the CT profiles of selected individuals administered with Uind computed assuming the 1 sec measurement period remained within 10% of accuracy area, thus insuring four times more precise effect site exposures. With the increase of measurement period PI95% was becoming larger and already exited the 20% accuracy area, however, still remaining within the 30% one with 30 sec measurement period. The measurement delay had a smaller effect on algorithm stability than the period and played a role only before the steady-state was reached.
The scenario with period of 15 sec and delay of 30 sec was considered as a realistic one. When assuming the maximum measurement noise the PI95% for individuals being administered with Uind had a tendency to be border line between 20% and 30%. The PI95% decreased lowering measurement noise such that the PI95% of virtual individuals administered with Uind stayed within the 20% accuracy area resulting in effect site CT profile variability being more than twice smaller than if they were administered with Uav.
Conclusions: We show that the approach based on Kalman filter ensures more precise and thus safer plasma and effect site exposures than currently achieved with open-loop TCI pumps. Reducing the measurement periods may provide up to four times better accuracy for effect site exposure.
References:
[1] Shafer, S.L. and K.M. Gregg, Algorithms to rapidly achieve and maintain stable drug concentrations at the site of drug effect with a computer-controlled infusion pump. J Pharmacokinet Biopharm, 1992. 20(2): p. 147-69. 2.
[2] E Eleveld, D.J., et al., A general purpose pharmacokinetic model for propofol. Anesth Analg, 2014. 118(6): p. 1221-37.
Reference: PAGE 27 (2018) Abstr 8467 [www.page-meeting.org/?abstract=8467]
Poster: Drug/Disease Modelling - CNS