Why Use Particles for Changing PK Parameters?
An individual's clearance or other PK characteristics may change over time. A stable-parameter update can use earlier concentrations to learn one fixed parameter value, but that assumption is unsuitable when the parameter itself evolves. A particle filter instead carries a distribution of possible current parameter values forward, changes those values according to a specified evolution model, and reweights them when a concentration is observed.
This is useful when concentrations are sparse and the posterior distribution is not well represented by a single estimate. The filter does not make sparse data informative by itself: its conclusions depend on the PK model, the assumed pace of parameter change, the observation-error model, and how well the particle set represents the distribution.
State and Observation Models
A filter requires two linked models. The state-transition model describes how the latent PK state changes between times. The observation model describes the probability of a measured concentration conditional on the state and known dosing and sampling history. A common assumption is that the state is Markov: given the current state and known history, the next state does not depend on earlier latent states. Observations are conditionally independent given the states and relevant dosing history.
For example, a positive clearance parameter can evolve as a random walk on the log scale:
Here \(CL_t\) is clearance at time \(t\), \(q\) controls the rate of change per unit time, and \(\Delta t\) is the time between states. This model allows gradual proportional changes while keeping clearance positive. It is an assumption, not a conclusion inferred automatically from the observations; the choice of \(q\) must be clinically and empirically defensible.
For an illustrative one-compartment IV bolus model with fixed volume \(V\), the concentration after a dose \(D\) may be written \(C(t)=D/V\exp(-CL_t t/V)\) when clearance is treated as constant over that interval. For observed concentration \(y_t\), one possible observation model is \(y_t\mid\theta_t\sim N(C_{\text{model}}(t;\theta_t),\sigma^2)\). This assumes additive, normally distributed measurement and residual error with known or separately estimated variance \(\sigma^2\). A lognormal observation model may be more appropriate when errors are proportional; the likelihood used for weights must match the chosen model.
Sequential Weighting and Resampling
At a new time, the filter first propagates each particle through the transition model. It then compares the new observation with the concentration predicted under each propagated state. In the bootstrap particle filter, particles are drawn from the transition model itself; consequently, their unnormalized weights are their previous weights multiplied by the observation likelihood.
For a Gaussian observation model, the likelihood is proportional to \(\exp\{-[y_t-C_{\text{model}}(t;\theta_t^{(i)})]^2/(2\sigma^2)\}\). The normalizing constant cancels when weights are normalized, provided the same observation variance applies to all particles. With a different proposal distribution, the importance weight must also account for the ratio of transition density to proposal density; omitting that correction generally targets the wrong distribution.
Repeated weighting can leave most particles with negligible weight. A useful summary is the effective sample size:
It ranges from 1 to \(N\): equal weights give \(ESS_t=N\), while one particle carrying nearly all weight gives an ESS near 1. If ESS falls below a chosen threshold, resampling draws a new particle population with replacement, with selection probabilities given by the normalized weights. The resampled particles usually receive equal weights. Resampling reduces weight degeneracy, but duplicates high-weight particles and discards low-weight ones; it does not create new information or parameter diversity.
Worked Example: Weighting Particles After a Concentration
Suppose four equally weighted particles represent possible clearances of 2, 3, 4, and 5 L/h. A hypothetical patient receives a 400 mg IV bolus, and \(V=50\) L, so the initial concentration is \(D/V=8\) mg/L. At 6 hours, a concentration of 5.20 mg/L is observed. Assume clearance is constant over this interval for each particle and the observation model is normal with \(\sigma=0.50\) mg/L. These assumptions specify the calculation; they do not claim that the true patient follows this model exactly.
For each particle, calculate the predicted concentration as \(8\exp(-CL\cdot6/50)\), then calculate a likelihood proportional to \(\exp[-(y-C_{\text{model}})^2/(2(0.50)^2)]\).
The unnormalized weights sum to \(0.0917+0.7480+0.8827+0.2697=1.9921\). Dividing by this sum gives weights approximately \(0.0460,\ 0.3755,\ 0.4431,\ 0.1354\), respectively. The weighted clearance mean is \(2(0.0460)+3(0.3755)+4(0.4431)+5(0.1354)=3.668\) L/h. This is a particle approximation to the posterior mean, not a claim that clearance is known to that precision. Among these four candidates, the observation favors the 3 and 4 L/h particles.
Propagation, Resampling, and Forecasting
After an update, the filter must represent the next state distribution, even if no concentration is measured at the next scheduled time. Each particle is propagated through the transition model, such as the log-clearance random walk. When another observation arrives, the propagated states are weighted by its likelihood. If an observation is missing, there is no likelihood update at that time; the state can still be propagated, and uncertainty can grow according to the transition model.
Worked Example: Effective Sample Size and Resampling
Use the normalized weights from the preceding hypothetical example: \(0.0460,\ 0.3755,\ 0.4431,\ 0.1354\). Their squared sum is approximately \(0.0460^2+0.3755^2+0.4431^2+0.1354^2=0.3578\), so \(ESS=1/0.3578=2.795\) particles. With \(N=4\), this is below a chosen resampling threshold of 3.
For an explicit systematic-resampling illustration, use cumulative weights \(0.0460,\ 0.4215,\ 0.8646,\ 1.0000\), and equally spaced points \(0.10,\ 0.35,\ 0.60,\ 0.85\). The first two points select the 3 L/h particle, and the final two select the 4 L/h particle. The resampled population is therefore \(3,3,4,4\) L/h, each with weight \(1/4\). The threshold and resampling scheme are implementation choices; systematic resampling is one standard option.
Now suppose the log-clearance transition has standard deviation \(0.10\) over the next interval. For illustration, use hypothetical standard-normal draws \(-0.5,\ 0.2,\ 0.7,\ -0.1\) for the four resampled particles. Applying \(CL_{\text{new}}=CL_{\text{old}}\exp(0.10z)\) gives approximately \(2.854,\ 3.061,\ 4.290,\ 3.960\) L/h. These are propagated possibilities, not four new measurements. With a different set of random draws, the particular values would differ, while the intended distribution is determined by the transition model.
Why the Update Works—and What Can Go Wrong
The weighting step follows directly from Bayes' rule. Given the previous observations, the transition model supplies the predicted distribution of the current state. The new observation then multiplies that predicted density by its conditional likelihood; normalization makes the result a probability distribution. Particles sampled from the transition distribution approximate that predicted density, and their likelihood weights approximate the multiplication by the observation likelihood.
This argument depends on a correctly specified transition and observation model, valid likelihood evaluations, and enough particles to represent the relevant regions of the distribution. A state-space formulation and the general logic of sequential Bayesian updating are the essential earlier material; the particle approximation is the numerical method for carrying out the update when direct calculation is inconvenient.
Particle estimates of posterior summaries are weighted sums. For example, the estimated posterior mean of clearance is \(\sum_i w_t^{(i)}CL_t^{(i)}\). A forecast is obtained by propagating particles forward through the state and concentration models, retaining their weights unless a new observation leads to reweighting. Thus a forecast distribution reflects both uncertainty about the current state and future state evolution.
Worked Example: Updating Again After State Evolution
Continue the hypothetical calculation above. After resampling and state propagation, suppose the four equally weighted clearances at the next interval are \(2.854,\ 3.061,\ 4.290,\ 3.960\) L/h. Take 5.20 mg/L as the concentration at the start of this interval and predict the concentration six hours later by multiplying it by \(\exp(-CL\cdot6/50)\). The resulting predictions are approximately \(3.692,\ 3.601,\ 3.108,\ 3.234\) mg/L.
Suppose a second hypothetical observation is 3.40 mg/L and the conditional observation standard deviation remains 0.40 mg/L. Its likelihoods, omitting the common normalizing constant, are \(\exp[-(3.40-C_{\text{model}})^2/(2(0.40)^2)]\). Substitution gives approximately \(0.766,\ 0.881,\ 0.766,\ 0.917\). Their sum is 3.330, so the normalized weights are approximately \(0.230,\ 0.265,\ 0.230,\ 0.275\). The updated weighted mean clearance is \(0.230(2.854)+0.265(3.061)+0.230(4.290)+0.275(3.960)=3.543\) L/h.
This calculation shows how an observation updates a distribution that has already evolved. The update is conditional on the chosen propagation draws and models; with a larger particle set, repeated runs can better represent the intended posterior and reduce Monte Carlo variation.
Diagnostics, Alternatives, and Limitations
Monitor ESS over time, the diversity of particle values after resampling, and whether predictions under the particles are compatible with observed concentrations. A sustained collapse of ESS indicates that the particle population is poorly matched to the likelihood or that the particle count is inadequate. Repeated resampling can produce particle impoverishment: many particles become copies, so later observations cannot readily recover states that were lost. Increasing the particle count, resampling only when needed, or using a proposal that incorporates the new observation can help, but each requires computational and modeling judgment.
A second failure mode is an unsuitable evolution variance. If \(q\) is too small, the filter may respond too slowly to genuine parameter changes; if it is too large, it may attribute ordinary measurement noise to implausible PK shifts and produce unnecessarily variable forecasts. Sensitivity to plausible values of \(q\), observation-error assumptions, dosing records, and sampling times should be examined. A persistent mismatch between observed and particle-predicted concentrations may indicate model misspecification, data problems, or too few particles; it is not automatically evidence that clearance changed.
| Approach | Useful when | Main limitation |
|---|---|---|
| Particle filter | The state changes over time and the posterior may be nonlinear or non-Gaussian. | Monte Carlo error and particle degeneracy can be substantial. |
| Kalman filtering | State transitions and observations are linear with Gaussian errors. | PK models are often nonlinear or constrained, so its assumptions may not hold directly. |
| Recursive update for stable parameters | Individual parameters are assumed constant across observations. | It does not represent genuine time evolution unless that evolution is added to the model. |
A particle filter is not automatically more reliable than these alternatives. It trades restrictive distributional assumptions for simulation error and dependence on particle design. It also requires a defensible model for parameter evolution; the data alone generally cannot distinguish rapid biological change from noisy observations without such structure. The particle filter updates PK information only here; combining it with pharmacodynamic observations requires a joint observation model.
Check Your Understanding
Try each question, then compare with the brief solutions below.
- In a bootstrap particle filter, what calculation changes particle weights when a new concentration arrives?
- For four particles with equal weights, what is the ESS? What does an ESS near 1 indicate for four normalized weights?
- Why can setting the transition variance \(q\) too high lead to misleading PK updates?
- If a scheduled concentration is missing, should the state transition stop? Explain.
Solutions. (1) Multiply each previous weight by the conditional likelihood of the observation under that particle, then normalize the products to sum to one. (2) Equal weights of \(1/4\) give \(ESS=1/[4(1/4)^2]=4\). An ESS near 1 means one particle carries nearly all the weight. (3) Excessive transition variance allows large parameter changes, so the filter may interpret residual noise as PK evolution and overstate state uncertainty. (4) No. Propagate the state according to the transition model, but do not apply an observation-likelihood update until a concentration is available.