Author for correspondence: Bernard Vrijens, AARDEX Ltd., Statistical Research Centre, Rue des Cyclistes Frontière, 24, B-4600 Visé, Belgium (fax +32 4 374 86 22, e-mail firstname.lastname@example.org).
Abstract: Electronic compilation of ambulatory patients' or trial participants' dosing histories has revealed that a wide range of dosing patterns, markedly skewed toward underdosing, occur in virtually every disease and treatment situation so far studied. In planning ambulatory trials and their analyses, one should recognize that patients' variable exposure to test drugs, created by their diversely erratic execution of protocol-specified dosing regimens, is generally the single largest source of variance in drug responses. Trial subjects' erratic dosing behaviour may, if ignored, weaken the trial's assay sensitivity. In contrast, reliably compiled and soundly analyzed dosing histories may greatly inform the analysis of the trial. Dosing histories found to be associated with suboptimal clinical results can highlight particular dosing patterns that should be avoided. Thus begins the sequence leading from first observations, to repeat observations, to ethically possible experimental designs, to causal inference, i.e., learning and then confirming. With the broadening use of electronic monitoring to estimate longitudinal drug exposure, the need exists for an explicit discipline that concerns itself with `what the patient does with the drug'. It is called Pharmionics.
Population pharmacokinetics and pharmacodynamics study the variability between individuals in drug concentration profiles and pharmacological effects when standard drug dosage regimens are assigned. These twin disciplines are widely regarded as a scientific basis for the determination of the optimal dosage of a new drug. Population pharmacokinetics-pharmacodynamics can be performed on relatively sparse data, and has thus the great advantage of enabling the study of a representative sample of patients who take the drug over a possibly long period of time.
At least when patients comply well with assigned regimen of treatment, the combination of a steady-state assumption and the knowledge of when the last dose intake took place is sufficient to retrieve information from plasma concentration measures taken just a few times for each patient in a sample of the study population (Sheiner & Ludden 1992). However, as revealed by electronic monitoring data, patients often comply poorly with a dose-timing assignment, even when they take the prescribed number of doses (Urquhart 1997). In practice patients experience a wide range of concentrations by varying the dosing interval while keeping the dose constant. Fig. 1 compares a typical time course of drug concentrations for a hypothetical patient following strictly punctual dosing (grey curve) with an actual patient's data, showing suboptimal drug intake. If not accurately measured and taken into account in sound analysis, variable drug exposure (resulting from variable adherence to prescribed regimen) is a major or leading source of noise in pharmacokinetic studies, resulting in biased pharmacokinetic estimates and large residual variability (Girard et al. 1996; Vrijens & Goetghebeur 1999). The pharmacodynamic estimation becomes even more problematic as the model derives its input from estimated pharmacokinetic parameters. The assumption of steady-state prior to a last-taken dose can then lead to an erroneous assessment of the history of drug exposure and hence to biased estimates of pharmacodynamic parameters (Vrijens & Goetghebeur 2004).
How to measure and exploit variable drug exposure optimally
Central to the estimation of a patient's execution of a prescribed drug regimen is a reliably compiled drug dosing history. Electronic monitoring methods have emerged as the virtual ‘gold standard’ for compiling drug dosing histories in ambulatory patients. Other methods, which rely, in one way or another, on the patient as the source of such information, have proven unreliable, because of frequently exaggerated self-reports, or maneuvers, such as pill-dumping, that create a false record of complete or nearly complete dosing (Arnsten et al. 2001; Liu et al. 2001; Urquhart 2002; Wagner 2002; McNabb et al. 2003).
Reliable estimation of hierarchical nonlinear models can be obtained by using timing explicit pharmacokinetic-pharmacodynamic models with accurate information on a number of previous dose timings. In practice, electronic monitoring methods enable reliable estimation of ambulatory patients' drug dosing histories. Especially for non-linear pharmacodynamic estimation, not only can bias be reduced but higher precision can be retrieved from the same number of data points when drug intake times are irregular, instead of regular. This somewhat counter-intuitive finding is explained by the fact that regular takers experience a relatively small range of concentrations which makes it not possible to estimate any deviation from linearity in the effect model. In addition, accurate measures of dosing histories give a unique opportunity to observe and study the off-response to medication as dosing lapses occur in practice on the patients' own initiative. Estimators of pharmacokinetic-pharmacodynamic parameters can thus benefit greatly from information that enters through variation in the drug exposure process, provided that it is reliably measured (Vrijens & Goetghebeur 2004).
Example in HIV-infected patients
We illustrate the proposed methodology through a study involving 35 antiretroviral-naive, HIV-infected patients prescribed a tri-therapy including lopinavir/ritonavir (QD: 800/200 mg or BID: 400/100 mg), stavudine, and lamivudine. Intensive sampling of lopinavir concentrations was done at week 3, and 4 additional trough concentrations were measured at intervals during the next 12 months. Using electronically monitored-compiled times of the last dose taken prior to blood sampling explained 40% of the residual, within-patient variability in the concentrations of lopinavir in plasma that arose when the electronically monitored-compiled dosing data were ignored and the samples were assumed to have been drawn at the trough. Overall variability was reduced by 55% when electronically compiled dosing history data were used, compared to the usual assumption about the trough (Vrijen et al. 2003a&b). In this closely managed, phase II trial, an average 95% of prescribed doses were taken, but considerable variability in dose timing was evident. Greater variability in dose-timing than in dose-taking is a consistent finding in published studies. Thus, switching from patient-reported to electronically-monitored dosing histories improved information derived from this population pharmacokinetic study.
By combining electronically recorded dosing histories with individual pharmacokinetic parameters of a single patient, one can project individual drug concentrations over long periods of time. This process is illustrated in fig. 1, which shows the model-based projection, based on the electronically compiled dosing history data, of a single patient's drug concentrations during 100 days. Drug holidays, which occur in many patients, are seen in fig. 1, and lead to drug concentrations falling below the expected steady state boundaries. From this type of picture one logically questions the relevance of monthly therapeutic drug monitoring when dosing histories previous to the sampling are unknown.
Monthly observations on plasma viral load (copies/ml) were categorized into 4 clinically meaningful states, 0–49, 50–399, 400–1999, and ≥2000 copies/ml. A time-dependent continuation ratio model was used to analyze longitudinal ordinal responses. Patient improvement and deterioration in viral load through the defined categories were modelled separately and both were significantly associated with lopinavir concentrations in plasma (P=0.0002 and P<0.0001, respectively). The most significant pharmacokinetic-derived predictor, the time that the concentrations fell below the EC50, appeared to be insensitive to between-patient variability in pharmacokinetic parameters. Changes in viral load were most significantly driven by within-patient dose-timing errors. In this setting it was thus possible to supplement patient-specific pharmacokinetics with a common summary of patient-specific dosing history patterns. This summary variable, called the timing error, is related to the 3rd moment of the distribution of interdose intervals. These analyses showed the particularly high impact of multi-day intervals between doses, i.e. drug holidays. Avoidance of these long inter-dose intervals should be a priority in efforts to maintain viral suppression.
Explanatory power of dose-timing data varies from one drug and treatment situation to another, with some antiretroviral drugs being much less forgiving than others. Comparisons among available drugs can include their relative degrees of forgiveness for serially-omitted doses. While the time that the concentrations fall below the EC50 allows for a direct interpretation of the internal drug exposure, the Timing Error has its own advantage for comparing the impact of variable adherence on viral load, independent of the kinetic properties of the drug studied.
Implementation of this sensible approach is illustrated in fig. 2 for two protease inhibitors. After initial viral suppression (viral load<400 cp/ml), the protease inhibitor represented by the dotted line (lopinavir/ritonavir) is less dependent on Timing Errors, compared to the one represented by the solid line (nelfinavir). In other words, for a given error in timing of drug intake, the probability of losing viral control is larger for the protease inhibitor represented by the solid line. For patients with perfect dose timing, the likelihood of losing viral control is very low. This type of analysis reveals thus a potential reason for virologic failure and can also guide the practitioner in determining when to consider an adherence intervention strategy.
A new terminology
Variable drug exposure due to variable adherence to prescribed therapy and its clinical consequences are sufficiently important, and have so many different aspects, that the need exists for an explicit discipline that concerns itself with `what the patient does with the drug', falling in line as a 3rd subdiscipline of biopharmaceutics. The other two subdisciplines are well-known: ‘pharmacokinetics’, what the patient's body does to the drug; ‘pharmacodynamics’, what the drug does to the patient's body. The 3rd subdiscipline is called …
Pharmionics is the discipline concerned with the ways in which prescription drugs “go” into use – in the broadest sense of the word “go”. This new field subsumes matters that meant little when prescription drugs had little therapeutic power, and were usually used singly rather than in complex combinations. Pharmionics has gained in importance as drugs have gained in both therapeutic strength and potential for harm if misused. Pharmionics is akin to avionics, which became essential as flight gained in power and speed, exceeding the unaided pilot's ability to control flight reliably. The common Greek root in both terms is ionics, from the verb to go. Today, we need to quantify reliably how the drug ‘goes’ in its intended use, as with avionics in flight.