Monitoring with Sparse Data

Why Two Models Can Disagree on a Single Trough

The discrepancy you are seeing represents a classic pitfall in pharmacokinetic (PK) modeling known as parameter unidentifiability due to sparse sampling.

The short answer is that both models can perfectly fit your single trough data point (resulting in a low Sum of Squared Errors and a favorable Akaike Information Criterion), but they make vastly different mathematical assumptions about what happens during the unmeasured distribution phase right after the drug is infused.

The Illusion of a "Good Fit" (Low SSE and AIC)
  • Mathematical flexibility: With only a single measured concentration (the trough), both a 1-compartment and a 2-compartment model have more than enough mathematical flexibility to adjust their parameters so that the predicted curve passes exactly through that single point.
  • Deceptive residuals: Because the curve perfectly intersects your only data point, the residual error at that point is near zero, yielding a very low Sum of Squared Errors.
  • Flawed AIC comparison: The Akaike Information Criterion penalizes a model for having too many parameters. However, if a 2-compartment model fits that single point perfectly, the structural penalty of having 4 parameters instead of 2 isn't large enough to offset the "perfect" fit. The AIC scores will therefore look clean and comparable, masking the reality that the 2-compartment model is mathematically over-parameterized for a trough-only dataset.
The "Unseen" Distribution Phase (Alpha Phase)

Vancomycin naturally exhibits multicompartment kinetics with a distinct distribution (α) phase that lasts 1 to 2 hours post-infusion.

  • 1-compartment model: Assumes instantaneous distribution. It draws a single exponential decay from the end of the infusion straight to the trough. Because it ignores the initial high concentration spike of the distribution phase, it generally calculates a lower, smoother Area Under the Curve (AUC).
  • 2-compartment model: Attempts to account for both a central volume (Vc) and a peripheral volume (Vp). Because it expects a sharp distribution phase, it projects a much higher, steeper peak immediately following the infusion before dropping down to meet the trough.

Because AUC is the total integrated area under the concentration-time curve (∫ C · dt), the large, unmeasured concentration spike projected by the 2-compartment model right after each maintenance dose heavily inflates its calculated AUC relative to the 1-compartment model.

Total Dependence on Bayesian Priors

When sparse data (a single trough) is fed into PK software, the software relies almost entirely on its Bayesian population priors to guess the rest of the curve.

  • In the 1-compartment model, the prior population estimates for Clearance (CL) and Volume of Distribution (Vd) are highly stable when anchored by a trough.
  • In the 2-compartment model, the software has to guess four variables (Vc, Vp, CL, Q) using only one real-world anchor point. If the underlying population model used by the software has a high distribution variance, the software will generate an unstable, unverified curve between the doses, leading to wildly divergent AUC calculations.
Structural Comparison of the Two Fits
Feature 1-Compartment Model (with Trough-Only) 2-Compartment Model (with Trough-Only)
Early post-infusion curve Smooth, lower exponential curve Sharp, high concentration spike (α-phase)
Calculated AUC Moderately conservative (may underestimate by ~10%) Frequently and significantly overestimated
Parameter stability High; easier to identify with sparse data Low; mathematically unidentifiable
Clinical Takeaway

Pharmacokinetic literature explicitly warns that two-compartment models built from trough-only data should be avoided.

While vancomycin is biologically a 2-compartment drug, a 1-compartment model is significantly more reliable and clinically acceptable when dealing with sparse, trough-only data. To utilize a 2-compartment model accurately and capture the true AUC, a second data point is required — specifically a peak sample drawn 1–2 hours post-infusion — to anchor and verify the distribution phase.

References

Links open in a separate, closable popup window. Ctrl/Cmd-click, middle-click, or right-click → "Open link in new tab" to open a reference in a new tab instead.

  1. Evaluated effects of reduced structural pharmacokinetic sampling on vancomycin PK parameter estimation. PMC9220236.
  2. Vancomycin dosing modeled with one vs. two compartments. EMCrit PulmCrit.
  3. Practical and resource-related challenges of obtaining additional sampling. PMC12465766.
  4. Comparison of vancomycin AUC using sparse data (peak-trough and trough-only datasets). PMC8899690.
  5. Pharmacokinetic model most widely used for vancomycin's multi-compartment profile. RxKinetics.
  6. Ideal vancomycin dosing regimens and AUC correlation with efficacy. RxKinetics.
  7. Optimizing vancomycin therapy: a review of AUC calculation techniques. Contagion Live.
  8. Vancomycin AUC estimation methods and their limitations. InsightRX.
  9. Vancomycin pharmacokinetics in septic patients, noncompartmental analysis. PMC5125998.
  10. Evaluation of the PK/PD index of vancomycin, reduced one-compartment model. PMC9220236.
  11. Vancomycin Bayesian modeling for clearance and volume of distribution. ClinCalc.
  12. Primary issues of using trough-only data with population models. PMC8518113.
  13. A one-compartment Bayesian method described using trough-only sampling. PMC10269041.
  14. Vancomycin dosing calculator for AUC-dosing in critically ill patients. DoseMe.
  15. Comparison of AUC for vancomycin from one- and two-compartment models using sparse data. ResearchGate.
  16. Guidance on monitoring accomplished through a PK model based on richly sampled vancomycin data. UNMC.