Dose banding- weighing up benefits vs risks

Background: Dose banding is the allocation of patients into a pre-specified dose group based on the patient’s value of a covariate and is commonly recommended in drug labels.  The simplest form of dose banding assigns all patients to receive the same dose irrespective of their characteristics.  The corollary to “one dose fits all” is where each patient is essentially their own dose group (i.e. a continuous dose adjustment, e.g. mg/kg of body weight). The benefits of dose banding is the simplicity for adjusting doses based on a patient’s characteristics, for instance all patients with eGFR > 30 ml/min might be recommended to receive the same dose and those less than 30 to receive a different (or no) dose.  There are two main issues with dose banding: (1) dose banding results in loss of granularity in dosing by applying a discrete set of doses rather than selecting from a continuous dose range which may reduce the probability of success and (2) patients may be harmed if they have a set of characteristics that is just above or below the dose banding cut-off such that they receive an increased dose or reduced dose, respectively.

Aims: This study aims to (1) determine the influence of dose banding vs “one dose fits all” vs continuous covariate based dosing on probability of target attainment (i.e. success), (2) explore optimisation of dose banding on the probability of study success, (3) determine the characteristics of optimising dose banding to reduce harm for those at the dose banding cut-point and (4) define a utility that maximises success while reducing harm.

Methods: The aims were addressed by a simulation study using a simple pharmacokinetic application. The model was a 1-parameter steady-state model defined by the parameter CL (mean=1 L/h, CV%=30) and dose (1 mg). An influential covariate with positive linear correlation (Z; mean=1, CV%=30) was chosen. Success (target attainment) was determined as a steady-state concentration of between 1 and 2 mg/L.  Harm was determined using a linear loss function associated with the distance of the predicted concentration from the target for subjects that had a covariate value that was just above or just below the dose band cut-off.  All simulations were performed using MATLAB.  Optimisation was performed using a global adaptive random search. The following scenarios, relating to the 4 aims were considered.  The design consisted of the dose banding cut off (i.e. the value of the covariate at which the dose would be changed) and the value of the dose for the band just below the cut off and above the cut off.  The following simulations were constructed (i) the probability of target attainment for “one dose fits all” [aim 1], (ii) the probability of target attainment for a continuous covariate based dosing regimen [aim 1], (iii) the optimal banding for target attainment with 2 or more dose levels [aim 2], (iv) the optimal banding for minimising harm for patients at the cut-off value of Z [aim 3] and (v) the performance of a utility that balances success against harm [aim 4].

Results: The lowest success is seen with a “one dose fits all” scenario and for (i) an optimal dose of 1.3 mg provided a probability of target attainment of 0.59. In contrast a dose that is based continuously on the covariate Z (ii) yielded a success of P=0.72.  Optimising the probability of target attainment with two dose levels yielded (iii) a dose band cut-off at the 52th percentile of the covariate Z and dose levels of 1.1 and 1.8 mg.  To minimise harm (iv), i.e. the loss associated with being dose reduced at the cut-off value of Z, then the best harm aversion method was to revert back to “one dose fits all”.  During the search either the dose for each dose band were the same or the cut-off value of Z went to the extrema (e.g. 100thpercentile).  Finally a linear utility of both probability of target attainment and the loss associated with harm resulted in (v) a narrowing of the differences between the two dose levels.  The dose band cut off was lower at the 36thpercentile of Z and the dose values were 1.1 and 1.3 mg.

Conclusion: Exploring dose banding, using a simple pharmacokinetic model, showed a clear benefit of continuous covariate based dosing for achieving success. Dose banding can be optimised to achieve higher success rates of attaining the target, but this also increases the risk of harm to patients at the cut-points.  If dose banding is to be considered then consideration of both target success as well as loss associated with dose adjustment need to be considered on a case by case basis.