Predicting phenytoin dosages using Bayesian feedback: a comparison with other methods

Pharmacokinetic interactions and dosing rationale for antiepileptic drugs in adults and children

AIMS

Population pharmacokinetic modelling has been widely used across many therapeutic areas to identify sources of variability, which are incorporated into models as covariate factors. Despite numerous publications on pharmacokinetic drug-drug interactions (DDIs) between antiepileptic drugs (AEDs), such data are not used to support the dose rationale for polytherapy in the treatment of epileptic seizures. Here we assess the impact of DDIs on plasma concentrations and evaluate the need for AED dose adjustment.

METHODS

Models describing the pharmacokinetics of carbamazepine, clobazam, clonazepam, lamotrigine, levetiracetam, oxcarbazepine, phenobarbital, phenytoin, topiramate, valproic acid and zonisamide in adult and paediatric patients were collected from the published literature and implemented in NONMEM v7.2. Taking current clinical practice into account, we explore simulation scenarios to characterize AED exposure in virtual patients receiving mono- and polytherapy. Steady-state, maximum and minimum concentrations were selected as parameters of interest for this analysis.

RESULTS

Our simulations show that DDIs can cause major changes in AED concentrations both in adults and children. When more than one AED is used, even larger changes are observed in the concentrations of the primary drug, leading to significant differences in steady-state concentration between mono- and polytherapy for most AEDs. These results suggest that currently recommended dosing algorithms and titration procedures do not ensure attainment of appropriate therapeutic concentrations.

CONCLUSIONS

The effect of DDIs on AED exposure cannot be overlooked. Clinical guidelines must consider such covariate effects and ensure appropriate dosing recommendations for adult and paediatric patients who require combination therapy.

Estimation of Population Pharmacokinetic Parameters of Free-Phenytoin in Adult Epileptic Patients

BACKGROUND

Serum concentrations of free-phenytoin (F-PHT) obtained in adult epileptic patients receiving PHT in monotherapy were analyzed to estimate the Michaelis-Menten pharmacokinetic parameters.

METHODS

Steady-state F-PHT serum concentrations, PHT dosing history, and associated information were collected prospectively. The maximum metabolic rate (Vm) and Michaelis-Menten constant (Km) of F-PHT and their interindividual variability data were estimated using nonlinear mixed effects modeling (NONMEM).

RESULTS

Twenty-nine patients with two or more available steady-state F-PHT serum concentrations (total of 63 dose/serum concentration pairs) met the inclusion criteria. Patients were taking PHT (100-500 mg/day) in monotherapy. The population estimates of F-PHT for Vm and Km were 9.1 mg/kg/day and 7.3 mg/L, respectively. The model was prospectively evaluated in a small group (seven) of additional patients.

CONCLUSIONS

The recommended daily dose in this population to achieve a F-PHT concentration of 1.5 mg/L is 6.1 mg/kg.

Three Michaelis-Menten pharmacokinetic dosing methods compared with physician dosing of phenytoin in an outpatient neurology practice

We compared predicted phenytoin serum concentrations using three Michaelis-Menten pharmacokinetic dosing methods with actual concentrations obtained from physician dosing in an outpatient neurology practice. Method 1 used population estimates for the Michaelis-Menten constant (Km) and maximum velocity (Vmax), method 2 used one dose and serum concentration pair to determine Vmax, and method 3 used two dose-concentration pairs to determine both Km and Vmax. In addition, physician doses were compared with pharmacokinetically calculated doses. Records of patients who received at least two phenytoin doses followed by two serum concentration determinations were reviewed. Data on age, gender, weight, physician doses, and resultant serum concentrations were collected. Pearson's correlation coefficient was used to compare physician maintenance doses with pharmacokinetically calculated predicted doses, whereas actual and predicted serum concentration data were used to determine precision and bias associated with each of the three methods. Actual serum concentrations fell into therapeutic range more frequently than predicted values in all but one comparison (method 3). Predicted and actual phenytoin doses were significantly correlated only with method 2. Only one of the three Michaelis-Menten pharmacokinetic dosing methods evaluated (method 3) was more predictive than physician phenytoin dosing.

Optimization of phenytoin therapy in adults with epilepsy in the Western Cape, South Africa

OBJECTIVE

To assess the extent to which adults with epilepsy were optimized and individualized on phenytoin monotherapy in the Western Cape, South Africa and to estimate the average optimized dose and serum phenytoin concentration, and the therapeutic range for this patient group.

METHODS

Patients were considered to be optimized on phenytoin if they were seizure-free or the best compromise was achieved between seizure reduction and side-effects.

RESULTS

538 (233 black and 305 coloured) adult people with epilepsy were treated at nine epilepsy clinics as outpatients. Of these patients, 332 (226 male and 106 female, 149 black and 183 coloured) were included in the data analysis as they were considered to have reliable phenytoin levels. Phenytoin doses and steady-state serum concentrations were predicted using the Michaelis-Menten equation. Patients attended a clinical pharmacokinetic service for 7.7+/-5.3 (range 1-22) months. The average optimized dose was 305.8 (range 100-500) mg/day and the average optimized level was 62.7+/-23.9 (range 15-133) micromol/l. Most patients (61.9%) were optimized in the therapeutic range 40-79 micromol/l; 21.1% were optimized above and 17% below this range. In 1.6% of patients serum concentrations above 120 micromol/l were required. Dosage adjustments were made in 47.0% of patients, increased in 31.9% and reduced in 15.1%.

CONCLUSION

These findings indicate that many patients (47%) attending outpatient clinics were not optimized on phenytoin therapy.

Evaluation of a bayesian pharmacokinetic program for phenytoin concentration predictions in outpatient population

The present work evaluates the performances of a Bayesian program (PKS) for phenytoin concentration predictions in an outpatient population. The retrospective study involved 19 epileptic adults receiving oral phenytoin. The program was used to predict estimated serum concentrations from 0, 1, 2 or 3 feedback concentrations. Measurements of prediction bias (ME) decreased as soon as one steady-state concentration (Css) was used for estimations. Precision (MAE) was significantly improved with 1 Css and was even better and stable with 2 and 3 Css. Likewise, RMSE (composite of bias and precision) regularly decreased when the number of Css used increased. On a clinical way, 12% of the estimations were unacceptable (prediction error > 5 mg/l) with 1 Css and less than 3% with 2 or 3 Css. This number of rejected estimations increased to 45% when no feedback concentration was used. Besides, the program was able to predict important rises of serum levels in spite of relative low increase of the dose when 1 Css at least was known. Thus, the phenytoin dosing program has acceptable performances when at least 1 Css is known, and represents a potential tool to assist the clinician in the particular condition of outpatient population.

Bayesian parameter estimation and population pharmacokinetics

The widespread application of Bayesian parameter estimation in the area of therapeutic drug monitoring (TDM) has prompted the need for well conducted population studies to obtain relevant prior pharmacokinetic parameter estimates. In many cases the population has consisted of a relatively small number of subjects. This may be unavoidable for drugs used in cancer chemotherapy or in small, specific populations of patients. In contrast, information about drugs which are used extensively, such as the aminoglycosides, can be obtained by population studies which involve a large number of individuals. Indeed, this technique has proved particularly useful for determining parameter estimates which can be employed in neonatal TDM. Bayesian parameter estimation has been most frequently used for drugs with narrow therapeutic ranges such as the aminoglycosides, cyclosporin, digoxin, anticonvulsants (especially phenytoin), lithium and theophylline. However, the technique has now been extended to cytotoxic drugs, Factor VIII and warfarin. Bayesian methods have also been used to limit the number of samples required in more conventional pharmacokinetic studies with new drugs. Further advances in the use of these methods are likely to include measures of drug response and toxicity requiring population studies which also include relevant pharmacodynamic information.

Dosage optimization methods applied to imipramine and desipramine in enuresis treatment

Three methods for estimating maintenance dosage requirements of imipramine were compared retrospectively in 146 enuretic patients. The dosing methods evaluated included individual (serum levels data) and/or population (average pharmacokinetic parameter) information. The use of imipramine and desipramine serum concentrations, as opposed to average population parameters only, improved forecast precision and accuracy for dosage individualization. The clinical acceptability of this was achieved through knowledge of a single serum concentration. No significant differences were seen between non-linear regression and the Bayesian method, this is in agreement with the high contribution of the patient's data to the Bayesian fitting (FF = 0.8). When one or two serum level data were available, a better performance was obtained by estimating pharmacokinetic parameters than level:dose ratios.

One-point feedback control method for phenytoin dosage adjustment

Routine clinical pharmacokinetic data collected from patients receiving phenytoin have been analysed to propose a new and simple equation to aid the dosage adjustment of this drug. The data were analysed using NONMEM, a computer program designed for population pharmacokinetic analysis that allows pooling of data. The rate equation for the elimination of phenytoin can be written as Do = kCssn, which fits the steady-state serum concentration (Css) and daily dose data (Do). The parameter n is the kinetic order and the parameter k is an arbitrary rate constant. From the above equation, D2 = D1C1 -nC2n can be derived, which forms the basis of predicting the dosage, D2, to obtain a desired Css, C2, using one initial Css, C1, obtained with an initial dose, D1, and using a population value of n. The value of n for phenytoin was estimated to be 0.312 in this study. The predictive performance of this equation was compared with the Richens and Dunlop nomogram and Bayesian feedback method using two or more steady-state concentration/dose pairs from each of 78 outpatients. This equation allowed the prediction of a dose needed to produce a desired steady-state concentration with errors comparable with the Bayesian feedback method for therapeutic drug monitoring.

Comparison of methods for the prediction of phenytoin concentrations

A comparison was made of different methods for the prediction of the serum concentrations of phenytoin (PHT) at steady-state with a view to determining which of them had the best predictive performance. The methods employed calculated the predicted concentrations based on a dose steady-state concentration pair. Two of the methods used involved solving the Michaelis-Menten equation, determination of a single parameter in each individual and maintaining the Km (Method A) or Vmax (Method B) values at a constant. Methods C and D were Bayesian techniques that used population parameters determined in a population studied by us (Method C) and parameters drawn from the literature (Method D). Calculation of bias and precision suggests that Method C is the most suitable of those studied, with a mean prediction error (ME) of 0.56 +/- 2.16 mg/litre, a mean absolute error (MAE) of 1.76 +/- 1.31 mg/litre and a root mean squared prediction error (RMSE) of 2.17 mg/litre. Method C was also the method that showed the lowest percentage of underestimation (5.26%) and overestimation (10.53%).

Phenytoin dosage individualization--five methods compared in the elderly

Five methods to predict phenytoin dosage have been compared in nine continuous care elderly patients. For each patient three steady-state plasma concentrations were obtained at three different doses. The data were used to estimate the 'optimum' dose for each patient by direct linear plot. The optimum dose for each patient was predicted from each plasma concentration using five dosage prediction methods based on the Michaelis-Menten equation using: (i) the population mean Vmax, (ii) the population mean KM, (iii) the linearized Bayesian method, (iv) the Rambeck nomogram, and (v) two plasma concentration-dose data pairs to estimate both Vmax and KM. The predictive precision was similar for each of methods (i-iv). Ninety-six out of 126 dosage predictions with the five methods were within 25 mg of the optimum dose. Methods (ii) and (iv) tended to overpredict dosage, particularly when used to interpret low plasma phenytoin concentrations.

Application of Bayesian estimation for the prediction of an appropriate dosage regimen of amikacin

A Bayesian approach was developed to determine an amikacin dosage regimen to achieve the desired plasma concentrations for each patient. Statistical characteristics of pharmacokinetic parameters were first evaluated in a group of patients (reference population), which when combined with three individual plasma concentrations of drug led to a Bayesian estimation of individual pharmacokinetic parameters. By using these parameters, an individual dosage regimen was then established to avoid residual and peak amikacin concentrations of up to 3 and 25 micrograms/mL, respectively. In a test group of 33 patients, adapted amikacin dosage regimens ranged from 4 to 43 mg/kg/d, with schedules requiring up to four infusions per day. Infusion time varied from 40 min to 4 h. These differences in drug administration protocol result from the wide interindividual variability of amikacin pharmacokinetic parameters. Performance of the developed methodology was evaluated by computing bias and precision of the estimated total body clearance and of the trough and peak amikacin concentrations that were reached after dosage regimen determinations.

Predictive performance of pharmacokinetic methods for phenytoin dosing: a multi-center evaluation in 282 patients with epilepsy

A retrospective study was conducted in 282 patients with epilepsy to assess the predictive performance of pharmacokinetic methods for individualizing dosage of phenytoin. Two population-based dosing methods (population clearance method and bayesian feedback method) and one individual-based method (the so-called linearized Michaelis-Menten method) were evaluated, when applicable, for single-point and/or 2-point dose predictions of phenytoin. In single-point predictions, we found a generally low percentage of dose calculations falling inside the +/- 10% range (48.9% and 51.1% for the population clearance and the bayesian methods, respectively). In 2-point predictions, the bayesian method was 'accurate' (dose within the +/- 10% range) in approximately 54.3% or 55.0% of cases (depending on the particular method of implementation adopted). An even worse percentage of 'accurate' dose predictions (38.3%) was obtained by using the linearized Michaelis-Menten method. Our data do not confirm results from previous studies indicating a generally good performance of pharmacokinetic methods for predicting phenytoin dosage.

Phenytoin dosage predictions in paediatric patients

Phenytoin dosing in paediatric patients is complicated both by alterations in patient requirements due to growth and maturation changes and by the capacity-limited characteristics of phenytoin metabolism. This study examines 2 pharmacokinetic methods to adjust phenytoin dosage based on a single dosing-rate/steady-state serum phenytoin concentration pair. A Bayesian forecaster and a fixed parameter [rate of metabolism (Vmax)] method were examined with previously published sets of a priori parameter estimates. The fixed Vmax method was utilised with the parameter derived from native Japanese (method 1), US Caucasian (method 2) and European (method 3) patients. The Bayesian forecaster used a priori parameter estimates obtained from native Japanese (method 4) and European (method 5) patients. Each method was examined retrospectively in 34 paediatric patients with a total of 48 predictions possible. Measures of absolute predictability, bias (mean error, % dose) and precision (root mean squared error, % dose), were -3.58/12.2, -1.51/12.2, 4.06/9.96, -4.38/13.2, and -3.10/11.5, for methods 1, 2, 3, 4 and 5, respectively. There was no significant difference among the 5 methods. However, the Bayesian algorithm tended to be more robust over a broad range of situations, providing predictions in all cases. The fixed Vmax methods could not provide predictions in every case. Finally, all methods had a significant number of overpredictions of dosage. Poorer results were observed when prediction of steady-state serum concentrations were performed, partly due to the retrospective nature of the study. We conclude that close monitoring of patients, regardless of the method chosen to adjust dosage, is recommended.

Comparison of single-point phenytoin dosage prediction techniques

To dose the anticonvulsant phenytoin (PHT) in a clinical situation is difficult because of the non-linear metabolism of the drug. Therefore many techniques have been advocated to aid dosage adjustments based on a single-point PHT concentration determined at steady-state (ss). We retrospectively investigated six methods in a population of 130 out-patients treated with PHT. The dose needed to achieve a desired PHT concentration at ss was calculated based on an observed ss dose-concentration pair using a Bayesian feedback method (B), the Richens and Dunlop nomogram (RD), the Rambeck nomogram (R), the Martin nomogram (M), a population clearance equation (PC), and the Wagner equation (W). The mean prediction error (ME), mean absolute error (MAE), and root mean squared error (RMSE) were separately calculated for each method, and served as a measure of prediction bias and precision. The MEs for B, RD, R, M, PC, and W, respectively, were -0.4, 1.2, 6.9, 3.4, -1.8, and 0.0 mg/day. The MAEs were 33.9, 38.5, 44.3, 50.4, 43.5, and 53.7 mg/day. The RMSEs were 43.7, 53.1, 65.2, 63.5, 56.0, and 68.2 mg/day. The MAE and RMSE showed lowest values for method B, followed by method RD. Therefore, we assume that method B is the most accurate in making routine PHT dosage adjustments.

A comparison of graphical nomogram methods with a computerized Bayesian analysis method in the interpretation of serum phenytoin concentrations

A study was performed to compare the predictability of a reported Bayesian graphical method, the drug nomogram used in the Bayesian Computer Method and the Driessen Nomogram with that of a computerized Bayesian analysis method in the interpretation of serum phenytoin concentrations. It was found that the results generated by the graphical method were similar to those of the computer with a mean prediction error of 3.9 mg/day in the dose to achieve a concentration at steady-state of 20 ml/l. Overall the results of the graphical method were less biased and had more precision with a significant improvement in relative precision (P less than 0.01) than the initial estimate or Driessen methods.

A comparison of phenytoin dosing methods in private practice seizure patients

In an attempt to evaluate the accuracy of phenytoin (PHT) pharmacokinetic dosage adjustments in a private practice setting, three single dose-concentration pair methods and three multiple point PHT pharmacokinetic dosing methods were studied. Dose and concentration data from 28 patients seen in a private neurology practice were utilized for the study. From a comparison of these methods in private practice seizure patients, it appears that the Bayesian feedback method may be the most accurate in making routine PHT dosage adjustments, perhaps by minimizing the contribution of unknown variables within the Bayesian approach.

Performance of Bayesian feedback to forecast lidocaine serum concentration: evaluation of the prediction error and the prediction interval

The prediction performance of the Bayesian feedback method was evaluated with respect to accuracy and precision, and efficacy and safety (width of the prediction interval) on the basis of 90 predictions in 30 patients treated with lidocaine. The mean of the prediction error (PE) and the root mean squared error (RMSE) served as a measure of accuracy and precision. The variance of the standardized prediction error (SPE) was used to evaluate the estimate of the standard deviation of the prediction error. SPE was defined as PE divided by the standard deviation of the predicted concentration. The standard error of RMSE and of the variance of SPE was determined by bootstrap. The results indicate that the lidocaine serum concentration at 12 hr (C2) after starting continuous infusion can be predicted with high accuracy and precision with a single feedback measurement obtained 2-4 hr (C1) after commencement of treatment: RMSE = 20.6%. Prediction at 24 hr (C3) was less accurate: RMSE = 31.4%. Using both C1 and C2 to predict C3 improved precision (RMSE = 23.4%). The evaluation of the prediction interval revealed that the current algorithm produces an upward biased estimate, probably due to a positive bias in the estimate of the covariance matrix of the parameter estimates. It is suggested that evaluation of prediction performance should include the estimate of the prediction interval.