Sparse sampling - collecting two to four blood samples per patient per cycle instead of the eight to twelve required for a full PK profile - has made therapeutic drug monitoring feasible in Phase II oncology trials where intensive sampling was never a realistic option. Maximum a posteriori (MAP) Bayesian estimation is the statistical engine that makes sparse sampling clinically useful: it combines the individual's observed concentrations with population PK priors to estimate that patient's individual clearance, volume, and AUC.
The combination works. Studies consistently show that two well-timed samples plus a good population model can estimate carboplatin AUC with precision comparable to full noncompartmental analysis in most patients. But "most patients" is not "all patients," and the failure modes are specific enough that understanding them should be mandatory knowledge for any clinical team designing a TDM protocol.
How MAP Bayesian Estimation Actually Works
The MAP Bayesian approach treats the problem as inference. We have a population PK model that specifies prior distributions for the individual's clearance (CL) and volume of distribution (V). These priors were estimated from a reference population and are parameterized in terms of typical values (theta parameters in NONMEM notation) and between-subject variability (omega matrix). We also have one or more observed drug concentrations in this individual patient.
The algorithm computes the combination of CL and V that maximizes the joint posterior probability given both the prior distribution (from the population model) and the likelihood of the observed data given those parameters. With a single concentration, the posterior is still heavily influenced by the prior. With two or more concentrations, the individual's data progressively dominate, and the posterior narrows around the individual's true parameters.
The AUC is then computed by integrating the concentration-time curve generated from the posterior CL and V estimates. This is the AUC that DoseMind reports for dose adjustment decisions.
Sampling Time Selection: The Three Errors That Break the MAP Fit
The single most impactful decision in a sparse sampling protocol is when to draw the samples. Three errors in sampling time selection consistently undermine MAP estimation quality, regardless of the software used.
Error 1: Both samples on the elimination phase. If both samples are drawn late post-infusion (for example, 4 hours and 24 hours for a drug with a half-life of 6 hours), they characterize elimination well but provide almost no information about the distribution phase. The MAP algorithm cannot distinguish between a patient with high clearance from one cycle to the next if both data points are on the same elimination slope. Adding one early sample (30 minutes to 1 hour post-infusion) dramatically improves Cmax estimation and volume of distribution precision, which in turn reduces AUC error.
Error 2: Sampling times that correspond to high inter-individual variability windows. In two-compartment models, the inflection point between distribution and elimination phases is where the concentration-time curve is most sensitive to individual variation in the inter-compartmental clearance rate constant (Q) and peripheral volume (V2). Placing a sample near this inflection point - rather than on the flat plateau of either phase - captures the parameter most uncertain in the population model. D-optimal sampling design calculations can identify these windows formally; clinical teams often skip this step and use round-number times that happen to cluster on the elimination slope.
Error 3: Ignoring actual infusion duration variability. Population PK models assume drug is administered as a constant-rate infusion over the protocol-specified duration. In clinical practice, infusion times vary - the nurse starts the carboplatin at 9:18 rather than 9:00, the infusion runs 42 minutes instead of 30 because the patient's IV line was flushed twice. If the nominal time is used rather than the actual time for the PK calculation, the sample timing is systematically wrong, introducing a structural error into every observation. TDM systems must capture actual infusion start and end times, not scheduled times.
Population Model Quality: The Invisible Variable
MAP Bayesian estimation is only as good as the population model it uses as the prior. The prior determines what the algorithm "believes" about typical clearance before seeing the patient's data. If the population model was built from a dataset that does not resemble your patient population, the prior will pull the posterior toward incorrect parameter values - and with only two samples, there may not be enough observed data to fully overcome a strongly miscalibrated prior.
Three questions to ask about any population PK model used for clinical TDM: First, what was the covariate structure? A model that only includes creatinine clearance as a covariate on CL may not account for age-related changes in tubular secretion that affect carboplatin beyond GFR. Second, what was the between-subject variability estimate on clearance? A model built from a relatively homogeneous population (e.g., good-performance-status patients in a Phase II trial) will have a narrower omega estimate than is appropriate for a Phase I population with impaired renal function, making the prior inappropriately confident. Third, has the model been externally validated? Internal NONMEM fit statistics (objective function value, goodness-of-fit plots) do not tell you how the model performs prospectively in a new cohort.
As we discuss in our article on what clinical teams need to know about the population PK model running their dosing software, interrogating the model's provenance is not optional for clinical TDM applications.
When MAP Bayesian Fails: Misspecified Model Structure
Model misspecification - using a one-compartment model for a drug with a meaningful distribution phase, or vice versa - is a systematic error that sparse sampling cannot correct. With three samples from a two-compartment drug analyzed in a one-compartment framework, the algorithm will fit the three points but will misattribute the early concentration decline to elimination rather than distribution, leading to AUC underestimation.
For carboplatin, a two-compartment model is appropriate for most patients. For drugs with enterohepatic recirculation (certain bile acid derivatives, some anthracyclines in the preclinical literature), even a two-compartment model may be inadequate - the secondary absorption peak creates a non-monotonic concentration-time profile that violates MAP assumptions about model structure. Sparse sampling TDM is not recommended for drugs with complex absorption pharmacology without formal prospective validation of the MAP approach in that specific context.
Practical Limitations in the Pediatric Population
Sparse sampling is particularly attractive in pediatric oncology because reducing sample volumes and venipunctures directly impacts patient welfare. But pediatric MAP estimation faces specific challenges. Population PK models with robust pediatric covariate data are far less available than adult models. The allometric scaling and maturation function parameterization used in pediatric models adds uncertainty, particularly for neonates and infants where scaling assumptions are least validated.
The recommendation is to use pediatric-specific population models where available and to interpret MAP estimates in very young patients with wider credible intervals than would be used in adults. The clinician's decision threshold for dose adjustment should account for this additional uncertainty - a point estimate of AUC 4.2 from a robust adult model in a 55-year-old patient carries different reliability than the same point estimate from a pediatric model in a 4-year-old with substantially different maturation function assumptions.
Minimum Sample Requirements for Reliable AUC Estimation
The empirical literature on minimum sample requirements is fairly consistent. For carboplatin with a standard two-compartment population model and typical between-subject variability of 20-25% on CL:
One sample: AUC estimation precision of approximately 20-30% CV%. Acceptable as a check for grossly aberrant exposure but not sufficient for dose adjustment decisions.
Two well-timed samples (one early, one late): AUC precision of approximately 8-12% CV% in prospective validation studies. This is the practical minimum for carboplatin TDM-guided dosing.
Three samples: AUC precision improves to approximately 5-8% CV% and allows more reliable estimation of individual volume of distribution. Recommended when the population model has high between-subject variability or when the patient population is expected to differ substantially from the model training dataset.
The Timing Precision Requirement for Clinical Workflows
One underappreciated practical constraint: MAP estimation requires that sample timing be accurate to within approximately 5-10 minutes for early samples (within the first hour post-infusion). At early time points on a typical carboplatin concentration-time curve, concentrations are declining steeply. A 15-minute error in a nominally 30-minute post-infusion sample time corresponds to a concentration that may differ from the true 30-minute value by 20-30%.
This precision requirement has direct implications for workflow design. Blood draw timing must be documented at the actual draw time, not the scheduled time. Training CRCs on accurate time documentation is not a formality - it is a data quality requirement for the PK calculation. TDM systems that accept scheduled rather than actual sample times will produce systematically biased AUC estimates in real-world deployment.
Conclusion: Sparse Sampling Works If You Engineer the Workflow for It
MAP Bayesian estimation from two or three sparse samples is reliable for carboplatin AUC-guided dosing in adults with a well-characterized population PK prior. The failure modes are not inherent to the statistical method - they are engineering problems: poorly chosen sampling times, nominal rather than actual infusion and sampling times recorded in the system, population models not validated in the target population, and model structures mismatched to the drug's pharmacokinetics.
A TDM system that captures actual event times, uses a validated population model, and presents the AUC estimate with appropriate confidence intervals addresses all of these failure modes in the clinical workflow. Contact the DoseMind team at hello@dosemind.com to discuss how sparse sampling can be integrated into your trial's dose optimization protocol.