Scientific, Medical, and Social Worlds › Physiology and Treatment Response

A physiological world model describes hidden biological state, measurements, and interventions under a chosen scope. Parameters and responses can vary among patients and within a patient over time. Randomized trials and controlled perturbations provide intervention evidence, but neither gives both mutually exclusive potential outcomes for one patient in the same episode. Missing state, treatment selection, and incomplete follow-up create additional identification problems. No simulation here is a dosing recommendation.

Consider an illustrative glucose-insulin model based on the minimal-model idea. Bergman and colleagues' 1979 study used measured insulin as an input to a model of glucose disappearance. The three-state system here is a control-oriented extension, not that original model verbatim. Its states are glucose concentration $G$, remote insulin action $X$, and insulin concentration $I$. An added insulin-balance equation and prescribed glucose appearance allow a delayed-action illustration. This is not the full UVA/Padova simulator or a validated model for individual treatment. With constant positive distribution volumes and specified basal values, write:

$$\begin{aligned} \frac{dG}{dt} &= -(p_1 + X)G + p_1 G_b + \frac{D(t)}{V_G} && \text{(glucose balance)} \\ \frac{dX}{dt} &= -p_2 X + p_3 (I - I_b) && \text{(remote insulin action)} \\ \frac{dI}{dt} &= -n(I - I_b) + \frac{u(t)}{V_I} && \text{(insulin balance)} \end{aligned}$$

Read these equations as a balance under the selected approximation. The term $-(p_1+X)G$ is nonpositive when $G\geq0$ and $p_1+X\geq0$, so it removes glucose in that scope. The signed action $X$ can be negative when insulin is below basal; if $p_1+X<0$ and $G>0$, this term instead adds glucose. The term $p_1G_b$ supplies a basal contribution, and $D(t)/V_G$ adds a glucose appearance rate divided by volume. $D(t)$ is glucose entering the modeled compartment, not raw meal mass; absorption needs its own model. Remote action grows with insulin above basal and decays at rate $p_2$. The insulin equation is a prescribed linear relaxation toward $I_b$ plus external input. These terms do not capture every feedback mechanism.

where:

• $G$: the plasma glucose concentration, the quantity the controller ultimately cares about
• $X$: a remote insulin action compartment representing insulin's delayed effect on glucose uptake
• $I$: the plasma insulin concentration
• $G_b$: the chosen basal glucose concentration
• $I_b$: the chosen basal insulin concentration
• $D(t)$: glucose mass appearance per unit time in the modeled compartment
• $u(t)$: exogenous insulin amount per unit time
• $V_G$: glucose distribution volume
• $V_I$: insulin distribution volume
• $p_1$: the insulin-independent glucose disappearance coefficient
• $p_2$: the remote-action decay coefficient
• $p_3$: the coefficient mapping insulin above basal into remote-action rate
• $n$: the insulin relaxation coefficient.

For consistent units, $p_1$, $p_2$, $n$, and $X$ have inverse-time units; $p_3$ has inverse-time-squared per insulin-concentration units. The represented state $s_t=(G_t,X_t,I_t)$ is only partly observed through $o_t$, such as a noisy glucose measurement. A belief $b_t$ should retain uncertainty about unmeasured action and parameters. A candidate controller must account for delayed response and the consequences of low glucose, but this set of equations alone does not establish a safe controller.

The example omits processes such as glucagon counterregulation and variation in meal absorption. Its intermediate action state is useful for explaining why the same current glucose measurement can coexist with different pending effects. A state estimate that includes recent input history can distinguish these situations better than a memoryless glucose-to-action rule under this model. It still needs parameter estimation, measurement validation, and a safe operating scope.

Delayed action can make repeated input based only on the current measurement problematic: effects from earlier inputs may arrive later. This is a reason to represent pending action, not a claim that every memoryless controller must fail or every model-based controller is safe. Filtering error, omitted counterregulation, and parameter mismatch can invalidate the apparent advantage. The model is a teaching example of partial observation and delay, not a clinical prescription.

The UVA/Padova simulator is a separate, more detailed example of scoped simulation evidence. Its authors report that FDA accepted S2008 for specified preclinical insulin-treatment and closed-loop algorithm evaluations. They subsequently found that patient trials had more hypoglycemia episodes than S2008 predicted and revised the simulator. Visentin and colleagues' 2014 comparison evaluates S2013 against glucose traces from a particular clinical trial and reports improved agreement.

This history does not make every virtual patient or treatment question validated. It identifies a version, intended use, comparison dataset, and later correction. Acceptance for that preclinical scope is not general certification of clinical accuracy, and the cited report does not establish a universal reason why regulators accepted the simulator.

Pharmacokinetics (PK) describes drug distribution and removal; a pharmacodynamic (PD) component relates exposure to a modeled effect. A two-compartment balance gives a useful unit check. Let $A_p$ and $A_t$ be drug amounts in central and peripheral compartments, with fixed volumes $V_p,V_t>0$. Concentrations are $C_p=A_p/V_p$ and $C_t=A_t/V_t$. An input $R(t)$ supplies drug amount per time to the central compartment, where elimination also occurs.

With first-order transfer rates $k_{12},k_{21}$ and elimination rate $k_{el}$, amount balances are $A_p'=-(k_{el}+k_{12})A_p+k_{21}A_t+R(t)$ and $A_t'=k_{12}A_p-k_{21}A_t$. Dividing by the respective volumes gives concentration dynamics:

$$\begin{aligned} \frac{dC_p}{dt} &= -(k_{el}+k_{12})C_p + k_{21}\frac{V_t}{V_p}C_t + \frac{R(t)}{V_p},\\ \frac{dC_t}{dt} &= k_{12}\frac{V_p}{V_t}C_p-k_{21}C_t. \end{aligned}$$

Transfer moves amount, not equal concentration increments between differently sized compartments. The volume ratios are dimensionless; each right-hand term has concentration-per-time units. Total modeled mass is $M=V_pC_p+V_tC_t$, so summing the amount balances gives

$$\frac{dM}{dt}=R(t)-k_{el}V_pC_p.$$

Transfers cancel. With no input or elimination, total mass stays constant. For example, if $V_p=1$, $V_t=2$, $C_p=1$, $C_t=0$, $k_{12}=1$, and all other rates and input are zero in consistent units, then $C_p'=-1$, $C_t'=1/2$, and $M'=-1+2(1/2)=0$. Omitting the volume ratio in $C_t'$ would incorrectly create mass.

This balance assumes constant volumes, nonnegative input, and the specified first-order flows. It does not incorporate saturable clearance, delays, additional organs, or changing volumes. Its mass invariant is a check on the representation, not evidence that it accurately predicts a particular patient's exposure.

A separate illustrative PD function maps central concentration to a saturating effect. For $C_p\geq0$, $EC_{50}>0$, $n>0$, and $E_{\max}\geq0$, the Hill expression below gives zero at zero concentration, half the modeled maximum at $EC_{50}$, and approaches $E_{\max}$ at high concentration. Those mathematical properties do not establish that a real response is monotone or saturating over the queried range:

$$E(C_p) = E_{\max}\frac{C_p^{\,n}}{EC_{50}^{\,n} + C_p^{\,n}}$$

where:

• $C_p,C_t$: central and peripheral drug concentrations, amount per volume
• $k_{el},k_{12},k_{21}$: nonnegative elimination and transfer coefficients, inverse time
• $V_p,V_t$: fixed positive compartment volumes
• $R(t)$: supplied drug amount per time
• $E(C_p)$: modeled effect at central concentration
• $E_{\max}$: asymptotic maximum in this illustrative response model
• $EC_{50}$: positive concentration giving half that maximum
• $n$: positive, dimensionless Hill exponent.

Between-patient variation motivates population models and individual parameter estimates. A hierarchical prior can pool information when measurements are sparse; it can also impose an inappropriate population assumption. Online filtering can update states or parameters when data support it. Neither calling a model a digital twin nor assigning physical units to its parameters establishes personalized predictive accuracy. Validation must match the patient group, intervention, horizon, and observation process.

Tumor models may describe growth and treatment response using several different mechanisms, including resistant subpopulations. The Norton–Simon hypothesis is a distinct phenomenological relation: treatment-induced regression is proportional to the unperturbed growth rate of a tumor at that size. Simon and Norton's own 2006 account describes its origin in laboratory and clinical observations and the testing of scheduling predictions in trials. It is not exclusively a resistant-clone competition model, and it does not assert that dose density always outranks dose intensity for every tumor, regimen, or toxicity constraint. The lesson here is to separate a specified model, its empirical tests, and its scope, not to infer a treatment recommendation.

To distinguish prediction from effect estimation, define patient-level potential outcomes $Y_i(1)$ and $Y_i(0)$ for specified treatment and control strategies. The binary assignment $A_i\in\{0,1\}$ selects one strategy. Under consistency, the observed outcome is:

$$Y_i = A_i Y_i(1) + (1 - A_i) Y_i(0)$$

For this patient and episode, we observe the potential outcome corresponding to the received strategy, not both alternatives. The average treatment effect in a stated target population is:

$$\mathrm{ATE} = \mathbb{E}[Y_i(1) - Y_i(0)]$$

where:

• $Y_i(1)$: the potential outcome for patient $i$ under treatment
• $Y_i(0)$: the potential outcome for patient $i$ under control
• $\mathbb{E}[\cdot]$: the expectation taken over the population of patients

Identification from observational data needs conditions, not simply a good outcome predictor. One standard identification route requires well-defined strategies and consistency, conditional exchangeability given sufficient pretreatment covariates, and positivity on the target population's support. Treatment interference must also be addressed by the estimand and design. Conditional exchangeability states:

$$\{Y_i(1), Y_i(0)\} \perp A_i \mid \mathbf{X}_i$$

where:

• $Y_i(1)$: the potential outcome for patient $i$ had they received treatment, a latent quantity that is observed for a patient only when $A_i = 1$
• $Y_i(0)$: the potential outcome for patient $i$ had they received control, a latent quantity that is observed for a patient only when $A_i = 0$
• $A_i$: the treatment indicator for patient $i$, equal to 1 if treated and 0 if not
• $Y_i$: the observed outcome for patient $i$, equal to whichever potential outcome corresponds to the received treatment
• $\mathbf{X}_i$: the vector of measured covariates for patient $i$, conditioned on for exchangeability

If baseline severity affects both assignment and outcome, a naive treated-versus-untreated comparison can confound effect with selection. The direction is not universal; it depends on the mechanisms and estimand. When a larger outcome value means harm, a harmful-looking effect is positive. Outcome discrimination and calibration among observed treatments do not, by themselves, identify the alternative-treatment contrast. They remain useful predictive diagnostics, but they answer a different question.

The synthetic cohort below makes one possible sign reversal numerical. Higher severity raises both treatment probability and baseline outcome, while assignment lowers the outcome by a fixed amount. An unadjusted group difference therefore mixes severity selection with treatment response. Regression adjustment uses the actual observed severity and the correct additive specification in this example; that privileged setup is why its result can be compared with the stipulated effect. It is not a claim that ordinary adjustment always recovers a clinical effect.

A target-trial approach specifies eligibility, strategies, assignment, follow-up, outcome, causal contrast, and analysis before attempting an observational emulation. For sustained treatment, confounding can vary over time and be affected by prior actions; an appropriate longitudinal analysis must address that structure. For the binary, single-time setting here, positivity requires:

$$0 < P(A_i = a \mid \mathbf{X}_i) < 1$$

where:

• $A_i$: the treatment indicator for patient $i$
• $a$: a particular treatment level
• $\mathbf{X}_i$: the vector of measured covariates for patient $i$
• $P(A_i = a \mid \mathbf{X}_i)$: the probability that patient $i$ receives treatment level $a$ given their covariates

for both treatment levels on covariate values in the target population's support. This is a condition on the population distribution, not a promise of many samples in every finite-data stratum. If an option never occurs for a relevant group, its effect there cannot be learned nonparametrically from that group's observed assignments alone. Extrapolative assumptions or other evidence can change the analysis, but they should be stated rather than hidden.

Sequential treatment introduces trajectory-level requirements. In an off-policy comparison, the observed assignment process must support the target actions at histories the target policy can reach, and sequential causal assumptions must hold. Importance weights or a fitted simulator cannot create support where none exists. State construction matters: an insufficient summary $s_t$ may leave confounding even when the full measured history would be adequate. This extends the planning discussion in Part VII without establishing the safety or efficacy of a learned clinical policy.

A treatment-effect claim should state the target population, strategies, identification assumptions, and sensitivity to departures from them. Prospective evidence is valuable where feasible, but it must also match the intended use. Outcome prediction alone is not a substitute for this argument. The causal estimates, their uncertainty, and the eventual action under a stated loss are three separate objects. Hernán and Robins' textbook provides the potential-outcome and longitudinal framework.