Why Cord-Blood Samples Test a Placental Model
Maternal–fetal PBPK models connect maternal drug concentrations to placental exchange and fetal concentrations. A cord-blood sample collected at delivery offers a direct observation with which to test a model’s prediction at a specific time. But it is a snapshot, not a measurement of the whole fetal concentration–time profile. A convincing comparison therefore depends on matching the prediction to the sample’s timing, matrix, and collection site, as well as on understanding which parts of the model the comparison actually tests.
The earlier material on Interpreting Maternal-Fetal Concentration-Time Profiles provides context for how maternal and fetal profiles may differ over time. Here the emphasis is narrower: evaluating model predictions against paired maternal and cord-blood samples. A measured maternal concentration can help describe the exposure at collection, but the analyst must decide whether it is an input to a conditional placental comparison or an outcome that the full PBPK model must predict too.
A valid comparison first establishes what “matched” means. The maternal and cord samples should be from the same dyad and collected at, or close enough in time to be treated as, the same time point. Record time relative to dose, gestational age, and whether the specimen is maternal plasma, whole blood, or another matrix. For cord samples, specify whether blood came from the umbilical vein or an umbilical artery: these sites need not have identical concentrations. The model output must correspond to the same sampling time and concentration basis.
There are two useful comparison designs. In a conditional placental comparison, the observed maternal concentration and sampling time may be supplied to the model, so the test focuses on the predicted cord concentration conditional on maternal exposure. In a full-model comparison, the model predicts both maternal and cord concentrations from the dosing history and physiological inputs; observed maternal values are then checked as well, rather than silently substituted. These designs answer different questions. A good cord prediction conditional on the observed maternal concentration does not by itself validate the model’s prediction of maternal exposure.
Quantifying Paired Prediction Bias
For \(n\) independent dyads, the average residual is \(\bar{e}=\sum_{i=1}^{n}e_i/n\). It estimates average observed-minus-predicted discrepancy in the sampled population under the design and assumptions used. A negative \(\bar{e}\) indicates average model overprediction; a positive value indicates average underprediction. The average alone is not enough: large positive and negative errors can cancel, so individual residuals and a measure of their spread should also be examined.
One inferential tool is the paired t procedure applied to the residuals. Its null hypothesis is \(H_0:\mu_e=0\), where \(\mu_e\) is the population mean residual for the target setting. A two-sided alternative is \(H_a:\mu_e\ne0\). The procedure assumes that dyads are independent and that residuals are drawn from a population that is approximately normal, especially when \(n\) is small. With larger samples, moderate departures from normality are often less influential, but outliers and dependence still require attention. Repeated samples from the same pregnancy are not independent dyads.
Here \(s_e\) is the sample standard deviation of the residuals. The test statistic measures the estimated mean discrepancy in standard-error units. A confidence interval for \(\mu_e\) is \(\bar e\pm t^*s_e/\sqrt{n}\), where \(t^*\) is the appropriate critical value with \(n-1\) degrees of freedom. This interval describes the range of mean biases compatible with the data and assumptions; it is not an interval for an individual future residual. A statistically detectable mean bias does not establish why the model is biased, and a nonsignificant test does not establish that predictions are accurate.
Worked Example: Testing Average Cord-Concentration Bias
In a hypothetical conditional comparison, four independent dyads have matched maternal and cord samples. The model predictions use each dyad’s observed maternal concentration and sampling time. Cord concentrations are in mg/L.
| Dyad | Maternal concentration (mg/L) | Predicted cord (mg/L) | Observed cord (mg/L) | Residual (mg/L) |
|---|---|---|---|---|
| 1 | 2.00 | 1.20 | 1.00 | -0.20 |
| 2 | 3.10 | 1.55 | 1.20 | -0.35 |
| 3 | 2.20 | 1.10 | 0.92 | -0.18 |
| 4 | 3.60 | 1.80 | 1.41 | -0.39 |
The residuals are observed minus predicted. Their mean is \((-0.20-0.35-0.18-0.39)/4=-1.12/4=-0.280\) mg/L. The mean is negative, so these data indicate average model overprediction of cord concentration by 0.280 mg/L, not underprediction.
The deviations from the mean are \(0.08,-0.07,0.10,-0.11\) mg/L. Thus \(s_e=\sqrt{(0.08^2+(-0.07)^2+0.10^2+(-0.11)^2)/3} =\sqrt{0.0334/3}=0.1055\) mg/L, and \(SE=0.1055/\sqrt{4}=0.05276\) mg/L. The test statistic is \(t=(-0.280)/0.05276=-5.307\), with \(df=3\). The two-sided p-value is approximately 0.013.
Under the paired-t assumptions, the data provide evidence against zero average bias. Using \(t^*=3.182\), the 95% confidence interval is \(-0.280\pm3.182(0.05276)\), or approximately \([-0.448,-0.112]\) mg/L. The interval suggests average overprediction in this hypothetical setting. With only four dyads, the normality assumption is difficult to assess and the estimate is imprecise; the result should not be generalized beyond the represented setting without further evidence.
Comparing Prediction Error and Candidate Models
Mean bias is a useful diagnostic but can hide poor individual predictions. A complementary summary is the root mean squared error, \(RMSE=\sqrt{\sum_i e_i^2/n}\). Squaring prevents positive and negative residuals from cancelling and gives greater weight to large errors. The RMSE has the same units as concentration. It is sensitive to outliers and should be reported alongside, not instead of, residual patterns and uncertainty. Mean absolute error, \(\sum_i|e_i|/n\), is another summary that is less dominated by the largest errors.
Model comparisons are most informative when each candidate’s predictions are made for the same observed dyads and times, using parameters fixed independently of those observations. If parameters were adjusted to minimize error on these same samples, the resulting error is a measure of fit, not an unbiased estimate of performance on new samples. A held-out validation set or an appropriately designed resampling strategy can help assess predictive performance, although either approach still requires representative data and careful handling of repeated dyads.
Worked Example: Comparing Two Fixed Placental Models
Two hypothetical models, A and B, were specified before being compared with two independent paired samples. Their predictions are in mg/L. For these displayed predictions, calculate the sum of squared residuals (SSE) and RMSE for each model.
| Sample | Observed cord (mg/L) | Model A prediction (mg/L) | Model B prediction (mg/L) |
|---|---|---|---|
| 1 | 0.120 | 0.150 | 0.1821 |
| 2 | 0.225 | 0.260 | 0.2737 |
For model A, the residuals are \(0.120-0.150=-0.030\) and \(0.225-0.260=-0.035\) mg/L. Therefore \(SSE_A=(-0.030)^2+(-0.035)^2=0.000900+0.001225=0.002125\ (\mathrm{mg/L})^2\). Its \(RMSE_A=\sqrt{0.002125/2}=0.03260\) mg/L.
For model B, the residuals are \(0.120-0.1821=-0.0621\) and \(0.225-0.2737=-0.0487\) mg/L. Therefore \(SSE_B=(-0.0621)^2+(-0.0487)^2=0.00385641+0.00237169=0.00622810\ (\mathrm{mg/L})^2\), or approximately \(0.00623\ (\mathrm{mg/L})^2\). Its \(RMSE_B=\sqrt{0.00622810/2}=0.05581\) mg/L. Model A has the smaller squared error for these two samples. This small descriptive comparison does not establish that A is generally superior or that its placental mechanism is correct.
Using Sampling-Time Patterns to Investigate Mechanism
Residual patterns across time can help prioritize model checks. A transfer limitation changes how rapidly and how much drug crosses the placenta; fetal elimination changes how drug is removed from the fetal compartment. Their effects on a cord concentration depend on the maternal input history, placental exchange, fetal distribution, and gestationally varying physiology. A single low cord value relative to the maternal value cannot identify which process is responsible.
Consider samples taken at different times relative to a maternal dose, each paired with a maternal sample. If observed cord concentrations are lower than predictions during the rising maternal phase, that pattern may be consistent with slower transfer or an overestimated transfer capacity. If early predictions agree but observations fall below predictions during the decline, excessive persistence in the fetal compartment—or inadequate fetal elimination in the model—may be a candidate explanation. These are clues, not unique diagnoses: changing dose timing, maternal clearance, binding, gestational age, or sample handling can produce related patterns.
Worked Example: Reading a Time-Dependent Residual Pattern
In a hypothetical study, maternal and cord samples are paired at two post-dose times in pregnancies with comparable gestational ages. The model and two prespecified variants predict cord concentrations conditional on the measured maternal concentrations. The values below are in mg/L.
| Time after dose | Observed maternal | Observed cord | Baseline prediction | Lower-transfer variant | Higher-fetal-elimination variant |
|---|---|---|---|---|---|
| 1 hour, rising phase | 2.0 | 0.18 | 0.30 | 0.19 | 0.28 |
| 6 hours, declining phase | 0.8 | 0.10 | 0.24 | 0.12 | 0.11 |
At 1 hour, the baseline residual is \(0.18-0.30=-0.12\) mg/L, whereas the lower-transfer variant residual is \(0.18-0.19=-0.01\) mg/L. At 6 hours, the baseline residual is \(0.10-0.24=-0.14\) mg/L; the higher-elimination variant residual is \(0.10-0.11=-0.01\) mg/L, and the lower-transfer variant residual is \(0.10-0.12=-0.02\) mg/L.
The early comparison is more consistent with the lower-transfer prediction, while both variants are close to the late observation. These two snapshots do not distinguish the mechanisms: both variants could improve fit at the later time. The appropriate next check is whether each mechanism, simulated with the same maternal input and physiological assumptions, reproduces the full pattern across well-timed samples and relevant gestational ages. The data shown are illustrative, not evidence that either mechanism is correct.
Diagnostics, Alternatives, and Limitations
Before interpreting a summary statistic, inspect residuals against predicted and observed cord concentrations, maternal concentration, time after dose, and gestational age. A consistent sign suggests bias; an increasing spread at higher concentrations suggests concentration-dependent error. A few extreme residuals may dominate RMSE or the paired t result. Also check whether predictions systematically miss one sampling site, such as umbilical artery versus vein, or one matrix. Investigate assay limits, sample timing, transcription, and units before changing physiological parameters.
A paired t interval addresses average residual bias under its assumptions, not the accuracy of every prediction. RMSE and mean absolute error summarize overall error but do not provide a causal explanation. Plots and stratified summaries can reveal patterns but may be unstable in small samples. Model-based alternatives can account for measurement error or correlated repeated observations, but require an explicit error and dependence structure. Whatever method is used, distinguish external validation from model calibration and report the number of independent dyads, sampling conditions, uncertainty, and any excluded or below-quantification samples.
Cord samples are typically available at delivery, so they may be sparse in time and concentrated in a limited range of gestational ages. They cannot alone reconstruct fetal exposure throughout pregnancy. Ratios such as \(C_{c}^{obs}/C_{m}^{obs}\) can be descriptive, but a ratio is not a universal measure of transfer: it depends on timing, maternal concentration, distribution, and potentially measurement error in the denominator. In particular, when maternal concentrations are near an assay’s quantification limit, ratios can be unstable. A low ratio is not proof of a placental barrier, just as a good prediction at delivery does not validate the entire fetal concentration–time profile.
Confirm dyad, sampling time, dose history, gestational age, matrix, umbilical vessel, and concentration units.
State whether maternal observations are model inputs for a conditional placental test or outcomes in a full-model evaluation.
Report observed-minus-predicted errors, average bias, uncertainty, and an error summary such as RMSE; check patterns by time and concentration.
Use time-dependent patterns to guide mechanistic checks, but do not infer transfer limitation or fetal elimination uniquely from sparse cord observations.
Check Your Understanding
Answer each question using the definitions and assumptions in this tutorial. Brief solutions follow the questions.
- If observed cord concentration is 0.40 mg/L and predicted concentration is 0.55 mg/L, what is the residual and which direction is the model’s error?
- For residuals \(-0.10\), \(0.02\), and \(-0.04\) mg/L, calculate the mean residual.
- What does a paired t test of mean residual equal to zero assess, and name two key assumptions.
- Why can a low cord-to-maternal concentration ratio not by itself establish limited placental transfer?
- Two prespecified models have RMSE values of 0.08 and 0.12 mg/L on the same validation samples. What can be concluded, and what cannot?
Solutions. (1) \(e=0.40-0.55=-0.15\) mg/L; the prediction exceeds the observation, so the model overpredicts. (2) The mean is \((-0.10+0.02-0.04)/3=-0.12/3=-0.040\) mg/L, indicating average overprediction in these observations. (3) It assesses evidence that the population mean observed-minus-predicted discrepancy differs from zero; assumptions include independent dyads and approximately normal residuals, particularly for small \(n\). (4) The ratio depends on sampling time, dose history, distributions, and the maternal denominator; a snapshot does not uniquely identify a transfer mechanism. (5) The first model has lower RMSE on these samples, indicating smaller root mean squared prediction error for this comparison. This alone does not prove general superiority, mechanistic correctness, or acceptable error in other populations.