Patient-Specific PK-PD Models

• Patient-specific PK-PD models are mathematical systems that use compartmental ODEs, Hill-type dynamics, and fractional derivatives to capture individual drug responses.
• They integrate mechanistic, computational, and machine learning methods to estimate and personalize key parameters from high-dimensional data, enabling precision dosing.
• These models support closed-loop control and scenario-based optimization in clinical settings such as anesthesia, ICU care, and drug-eluting devices under safety-critical conditions.

Patient-specific pharmacokinetic-pharmacodynamic (PK–PD) models explicitly individualize the mathematical description of drug administration, distribution, effect, and response to a specific patient, exploiting both prior physiological knowledge and subject-level data. In contrast to population-averaged models, patient-specific PK–PD models are essential for precision dosing, personalized therapy, and individualized safety margins, especially in high-risk or dynamic clinical scenarios. Increasingly, such models integrate mechanistic, computational, and machine learning methods to leverage high-dimensional EHR, physiological, or molecular data.

1. Mathematical Foundations and Structural Components

The mathematical backbone of patient-specific PK–PD modeling remains the coupled compartmental system—linear or nonlinear ODEs (integer or fractional order) for PK, and a static or dynamic nonlinear mapping for PD. Patient idiosyncrasy is encoded in all parameter values: volumes, clearances, rate constants, transit times, and the functional form and parameters of PD effects.

Typical model structure:

• PK subsystem: $x$ (state, compartmental concentrations/masses), $u$ (input, dose events), governed by

$$\dot x(t) = A x(t) + B u(t),\qquad \bar y(t) = C x(t)$$

with $A$ comprising patient-specific rate constants, possibly time-varying or depending on demographic covariates (weight, age, etc.) or even time-warped via fractional derivatives (Zaitri et al., 2023, Daryakenari et al., 2024).

• PD subsystem: Static or dynamic mapping, frequently using Hill-type nonlinearities:

$$y(t) = \varphi(\bar y(t)) = \frac{E_{\max} \bar y(t)^\gamma}{C_{50}^\gamma + \bar y(t)^\gamma}$$

with individual-specific $E_{\max}$, $C_{50}$, $\gamma$, or, in some cases, full causal GP or learned neural mapping for nonparametric subject effects (Lu et al., 2020, Cheng et al., 2019).

Extensions:

• Adjoint or effect-site compartments: $C_{e}$ dynamics are linked to central PK and produce delayed, patient-dependent effects (Credico et al., 2024).
• PDE-based PK: For depot, stent, or tissue-targeted therapies, PK is governed by patient-specific tissue-diffusion and reaction models in 3D geometries (Manjunatha et al., 2023).
• Fractional-order PK: Fractional Caputo (or $\psi$-Caputo) derivatives introduce power-law memory and nonlocal tissue trapping, with patient-level fractional index $\alpha$ (Zaitri et al., 2023, Daryakenari et al., 2024).

2. Parameter Identification and Individualization

Direct estimation:

• Global optimization (e.g., branch-and-bound): Sought for nonconvex nonlinear regression involving ARX + Hill-type PD inversion (Credico et al., 2024). Guarantees exact patient-specific recovery if the mathematical structure admits unique inversion and the data are informative.
• Standard mixed-effects estimation: Empirical Bayes, hierarchical Bayesian models with population priors for all PK/PD parameters. Each patient's parameters are jointly inferred from their data within the cohort (Margossian et al., 2021).

Amortized/encoded estimation:

• Deep-learning encoders (biGRUs, MLPs, variational autoencoders) directly regress patient PK/PD parameters from short subject time-series as a summary statistic for personalized forecasting or dosing (Lu et al., 2020, Potosnak et al., 2023).

Scenario-based and robust optimization:

• Bayesian scenario generation: Enumerating a grid of plausible patient PK–PD parameters (e.g., $(\alpha, b)$ pairs for elimination and effect) and optimizing dosing across scenarios for worst-case or high-probability safety (He et al., 2023).

Advanced parameter types:

• Learned time-varying, piecewise, or fractional parameters (e.g., $k_{12}(t)$ time-dependent, $\alpha$ as a trained scalar) are fitted by PINN/fPINN frameworks to facilitate patient-specific anomalous diffusion, trapping, and adaptation (resistance/tolerance) (Daryakenari et al., 2024).

3. Control, Dosing, and Closed-Loop Feasibility

Corridor control:

• The output corridor control formalism seeks PWM (pulse-width modulated) or impulsive (bolus-based) inputs such that the (nonlinear) PK–PD system output remains in a given safety band. The solvability hinges on the structure of the patient-specific model—especially the steepness ($\gamma$) of the PD Hill nonlinearity (Medvedev et al., 18 Jan 2026).
• Solvability conditions: For third-order positive systems with Hill PD, the existence, uniqueness, and explicit computability of bolus size and interval $(\lambda^*, T^*)$ sustaining the corridor are rigorously characterized; only sufficiently steep PD $\gamma$ permit feasible, safe dosing (Medvedev et al., 18 Jan 2026).

Model-based RL and MPC:

• Operationalizing patient-specific dosing in ICU settings, scenario-based model-predictive control (MPC) and RL optimize n-step dosing plans, using individualized models as simulators, subject to explicit safety bands. MILP-based approaches are used for systems with mixed-integer structure arising from piecewise PK and staged dosing (He et al., 2023).

Clinical translation:

• Real-time controllers rely on patient-specific fast identification, either from rich data or prior models; infeasibility (e.g., $\lambda^*$ or $T^*$ outside safe bounds) triggers escalation to continuous infusion or advanced monitoring (Medvedev et al., 18 Jan 2026).
• Rapid PD imaging: Sequential small-dose bolus paradigms, with time-series modeling (PK→PD→observed effect + drift), recover patient EC$_{50}$ and other sensitivity indices even in single short sessions, provided noise/signal permits (Black et al., 2013).

4. Data-Driven and Machine Learning Approaches

Hybrid mechanistic-ML frameworks:

• Neural ODEs and PINN/fPINN architectures embed classical PK–PD ODEs (integer or fractional) in neural networks, with physicochemical parameter constraints and neural field approximators for unknown or highly nonlinear components (Lu et al., 2020, Daryakenari et al., 2024). Patient-specificity is conferred by parameter embedding layers, subject-level inputs, or amortized encoders.

Global-local forecasting:

• In hybrid architectures, patient-specific low-dimensional PK parameters (e.g., absorption rate constants for insulin in glucose prediction) are treated as trainable embeddings, while the high-dimensional dynamics (glucose trajectories) are learned globally across a cohort, yielding improved performance over purely local or purely global models (Potosnak et al., 2023).

Latent force models with GPs:

• Convolutional/nonparametric models express the effect of medication as a causal latent force added to patient-specific GP baselines. Drug administration times are encoded as time-marked kernel events, with per-patient effect amplitude and decay hyperparameters (gains and decay rates) fit hierarchically (Cheng et al., 2019).
• Analytical cross-covariances enable efficient GP inference and partitioning of physiological variability into intrinsic patient idiosyncrasy and drug-induced causal response.

Approach/Framework	Patient-Specificity Mechanism	Distinctive Feature(s)
PWM Control via Corridor (Medvedev et al., 18 Jan 2026)	Direct-param identification, 1-cycle explicit solution	Solvability theory for safety
Bayesian Hierarchical (Margossian et al., 2021)	Patient-level priors, population-pooling	Full uncertainty, mixed effects
Neural ODE/Encoder (Lu et al., 2020, Potosnak et al., 2023)	Low-dim per-patient encoder; amortized inference	Predictive accuracy, regimen generalization
PINN/fPINN (Daryakenari et al., 2024)	Joint PINN+ODE/fractional order, time-varying params	Anomalous diffusion, resistance
Latent Force GP (Cheng et al., 2019)	Per-patient gain/decay in convolved kernel	Analytical tractability
MILP RL/MPC (He et al., 2023)	Scenario grid for patient PK–PD	Safety-critical dosing in ICU

5. Clinical Applications and Safety-Critical Contexts

Anesthesia:

• Three-compartment and fractional PK with effect-site and Hill-PD models dominate in anesthesia modeling for agents like propofol or atracurium. Individual $k_{ij}$, $C_{e50}$, $\gamma$, and fractional $\alpha$ are tuned to demographic details or empirical BIS data. Exact identification is critical for automated closed-loop infusion (Credico et al., 2024, Zaitri et al., 2023).
• Safety bounds (e.g., max bolus size, min inter-dose interval) are imposed as hard constraints in identification/control, and corridor infeasibility is often due to shallow Hill slopes (low PD $\gamma$) (Medvedev et al., 18 Jan 2026).

ICU anticoagulation:

• Personalized PK modeling (Michaelis–Menten, piecewise-linear) for unobservable heparin concentration with individualized PD mapping to coagulation (e.g., aPTT). Scenario-MPC ensures high-probability maintenance within therapeutic ranges (He et al., 2023).

PK–PD imaging and challenge experiments:

• Rapid quantitative PK–PD imaging deploys patient-specific single-compartment PK (multiple small boluses), fit to per-voxel imaging time-series via sigmoidal Emax models, yielding individualized, region-specific EC$_{50}$ with explicit statistical testing (Black et al., 2013).

Drug-eluting devices and tissue models:

• Patient coronary geometry, tissue-level diffusion-reaction-advection PDEs, and personalized stent release kinetics underpin in-silico prediction and tuning of restenosis outcomes and optimal drug load in stented vessels (Manjunatha et al., 2023).

6. Sensitivity Analysis, Identifiability, and Limitations

Sensitivity to patient-specific PD parameters:

• The feasibility, safety, and dynamic controllability of patient-specific PK–PD systems are dramatically affected by nonlinear PD steepness ($\gamma$): shallow curves (small $\gamma$) may preclude holding effect within a corridor via any feasible bolus/interval choice—even as kinetic acceleration (increased $\alpha$) offers limited relief (Medvedev et al., 18 Jan 2026).
• Fractional-order parameters ($\alpha$) in fPINN/fractional-Caputo models modulate system memory and can delineate subjects with anomalous drug trapping, resistance, or delayed effect, offering mechanistic axes for individualization (Zaitri et al., 2023, Daryakenari et al., 2024).

Identifiability:

• Rich excitation (enough dynamic range and time points) is necessary for unambiguous identification of individual PD parameters; uninformative data (e.g., monotonic response, low signal/noise) can result in non-identifiable or multiply feasible patient parameterizations (Credico et al., 2024).
• In imaging or high-dimensional time-series settings, statistical tools (F-tests, model selection, cross-validation) and hierarchical Bayesian shrinkage protect against overfitting or misattribution of drug effects (Black et al., 2013, Margossian et al., 2021).

Model structure and computational limits:

• The curse of dimensionality in BnB and MILP identification can inhibit real-time deployment for complex PD maps or high-order ARX models. PINN/fPINN and neural-ODE frameworks trade-off explicit mechanistic interpretability for data-driven flexibility but require large and diversified subject-level time-series (Daryakenari et al., 2024, Lu et al., 2020).

7. Outlook: Integration, Adaptation, and Future Directions

Comprehensive patient-specific PK–PD frameworks are trending toward hybrid architectures: integrating explicit compartmental, stochastic, and time/fractional-adaptive components with neural surrogates or GP-based flexible mappings. Real-time adaptation, uncertainty quantification, and robust safety optimization are increasingly emphasized for deployment in closed-loop settings (e.g., automated anesthesia, ICU titration), as well as for adaptive, in-silico trial design (Lu et al., 2020, Daryakenari et al., 2024, Potosnak et al., 2023, He et al., 2023).

Key challenges remain: rapid and robust identifiability from sparse or noisy real-world data; the clinical validation of these models in diverse, high-acuity populations; and interpretability and safety guarantees under algorithmic adaptation. Emerging methodologies—physics- and data-informed neural architectures, scenario optimization, and hierarchical Bayesian regularization—provide powerful tools for achieving individual precision without sacrificing model-grounded safety and reliability.

References:

• "Solvability of The Output Corridor Control Problem by Pulse-Modulated Feedback" (Medvedev et al., 18 Jan 2026)
• "Deep learning prediction of patient response time course from early data via neural-pharmacokinetic/pharmacodynamic modeling" (Lu et al., 2020)
• "Model Based Reinforcement Learning for Personalized Heparin Dosing" (He et al., 2023)
• "Pharmacokinetic/Pharmacodynamic Anesthesia Model Incorporating psi-Caputo Fractional Derivatives" (Zaitri et al., 2023)
• "Global Deep Forecasting with Patient-Specific Pharmacokinetics" (Potosnak et al., 2023)
• "CMINNs: Compartment Model Informed Neural Networks -- Unlocking Drug Dynamics" (Daryakenari et al., 2024)
• "Rapid quantitative pharmacodynamic imaging by a novel method" (Black et al., 2013)
• "A Branch and Bound method for the exact parameter identification of the PK/PD model for anesthetic drugs" (Credico et al., 2024)
• "Computational modeling of in-stent restenosis: Pharmacokinetic and pharmacodynamic evaluation" (Manjunatha et al., 2023)
• "Flexible and efficient Bayesian pharmacometrics modeling using Stan and Torsten, Part I" (Margossian et al., 2021)
• "Patient-Specific Effects of Medication Using Latent Force Models with Gaussian Processes" (Cheng et al., 2019)