C: Cardiovascular system

Modeling notes

Endpoint. Continuous surrogates measured repeatedly: blood pressure, LDL cholesterol [4,5]. Both lag plasma concentration, because the drug acts on the turnover of the measured quantity [6,7].

Typical structure. Indirect response models [8,9], near-universal for LDL [10,11] and common for blood pressure. Effect-compartment models only where the delay is genuinely short, as for the acute heart rate effect of a beta blocker [12,13].

Counter-regulation (C02, C03, C09). Baroreflex activation, renin release and renal sodium handling push back against blood pressure effects [14], which is why feedback and tolerance terms appear in antihypertensive models [15].

Baseline is the trap. Blood pressure has a circadian rhythm, large measurement error and regression to the mean on enrolment [16,17], and placebo arms show genuine falls [18,19]. How baseline is handled changes bias and precision in the estimated drug effect, and normalising each observation by its own baseline is the worst common option [20,21].

Covariates worth pre-specifying. Renal function for diuretics and renally cleared renin-angiotensin agents [22,23], CYP2D6 for metoprolol and flecainide [24,25], SLCO1B1 for statins [26]. A data-driven search is underpowered at realistic dataset sizes [27,28].

Safety. A concentration-QTc analysis is expected under ICH E14 for essentially every new drug, as a separate pre-specified analysis [29,30] (see C-QTc modeling). Class-matched effects such as hypotension and bradycardia are exaggerated on-target pharmacology, and keep climbing after the efficacy curve flattens [31,32].

Approval rests on the surrogate. Approval often rests on the surrogate, with the outcome trial reporting years later [33,34]. The pharmacometric job is linking biomarker change to event hazard [35,36], and handling the adherence and dropout endemic to chronic outpatient dosing [37,38].

References