Narrow Therapeutic Index Drugs: Regulatory Frameworks, Bioequivalence, and Model-Informed Precision Dosing

1. Introduction

Narrow therapeutic index (NTI) drugs are agents for which small changes in dose or blood concentration may result in serious treatment failure or life-threatening, persistent, or significantly disabling adverse effects [1,2]. Their clinical risks arise not only from the narrow separation between effective and toxic doses or concentrations, but also from steep exposure–response relationships, substantial interindividual pharmacokinetic and pharmacodynamic variability, and limited flexibility for dose adjustment [3]. NTI drugs are used across a broad range of therapeutic areas, including anticoagulation, cardiovascular disease, neurological disorders, respiratory disease, and immunosuppression; representative examples include warfarin, digoxin, phenytoin, lithium, theophylline, cyclosporine, and tacrolimus [4]. Accordingly, maintaining therapeutic efficacy while minimizing the risks of underexposure and overexposure remains a common challenge in the development, regulatory evaluation, and clinical use of these drugs.

Generic substitution is an important strategy for reducing drug costs and improving access to treatment; however, the pharmacological characteristics of NTI drugs make the assessment of interchangeability more complex than that for conventional drugs [5]. For these agents, even small differences in systemic exposure between formulations may raise concerns about reduced efficacy or increased toxicity, while sequential switching among different generic products may introduce a risk of so-called “biocreep” [5,6]. “Biocreep” arises when multiple generic products each meet bioequivalence criteria against the same reference product, but their exposure estimates occupy different positions within the acceptance range. Consequently, switching between these generic products may still produce substantial differences in drug exposure. For example, postapproval studies indicated that Accord Healthcare’s generic tacrolimus oral capsules might produce a higher peak blood concentration than the reference product, Prograf. Given the associated toxicity risk, the FDA changed the products’ therapeutic equivalence rating from AB to BX [5]. Potential biocreep therefore complicates regulatory oversight of narrow therapeutic index (NTI) drugs. Evaluations should extend beyond generic–reference bioequivalence to consider generic-to-generic interchangeability, complementary exposure metrics, and continued postapproval reassessment.

The approval of generic drugs typically relies on bioequivalence studies that bridge to the established efficacy and safety evidence for the reference product. Regulatory assessment must therefore establish that formulation differences do not produce clinically meaningful changes in drug exposure. For narrow therapeutic index (NTI) drugs, the margin between therapeutic and toxic exposure is small. Even modest changes in exposure may therefore compromise efficacy or safety, providing a scientific rationale for stricter or product-specific bioequivalence criteria. Accordingly, some regulatory authorities apply narrower bioequivalence acceptance ranges, whereas others use fully replicated crossover designs. These designs estimate within-subject variability for the reference product and compare the test and reference products in terms of both mean exposure and within-subject variability [7,8]. Nevertheless, substantial differences remain across countries and regions in the terminology, classification criteria, drug lists, study designs, and BE acceptance standards applied to NTI drugs. These inconsistencies limit the cross-regional comparability of study findings and complicate global drug development and clinical substitution decisions [9]. A recent analysis of generic drug applications submitted to the U.S. Food and Drug Administration (FDA) further showed that BE assessment frameworks for different NTI drugs vary in stringency, study pass rates, and statistical error control, indicating a continued need to balance scientific rigor, practical feasibility, and international regulatory harmonization [10]. Current regulatory frameworks lack a consistent cross-jurisdictional approach to narrow therapeutic index (NTI) drug designation, pharmacokinetic endpoint selection, study design, and bioequivalence acceptance limits. These regulatory decisions are not consistently aligned with drug-specific clinical risks.

The choice of parameters for bioequivalence assessment should reflect the exposure characteristics that formulation differences may alter and the clinical relevance of those changes. AUC_0–t quantifies cumulative systemic exposure from dosing to the last quantifiable concentration, whereas C_max captures peak exposure and is influenced by the absorption rate. These measures therefore serve as the primary endpoints in conventional pharmacokinetic bioequivalence assessments. AUMC is the area under the first moment curve of the concentration–time profile, whereas MRT is generally calculated as the ratio of AUMC to AUC. After oral administration, MRT reflects both drug absorption and subsequent disposition. Although AUMC and MRT characterize the temporal profile of systemic exposure, neither directly substitutes for measures of total or peak exposure. Their estimates are also particularly sensitive to inadequate terminal sampling and the proportion of the curve requiring extrapolation. They are therefore generally treated as supplementary pharmacokinetic parameters rather than primary endpoints for routine bioequivalence decisions. When early exposure, onset of action, or steady-state trough concentration is clinically relevant, partial AUC, t_max, C_τ, or other appropriate measures may be included. The choice should reflect the drug, formulation, and specific clinical question. For narrow therapeutic index (NTI) drugs, endpoint selection and the definition of acceptance limits are related but distinct decisions. Endpoint selection should identify the exposure dimension requiring comparison on the basis of the exposure–response relationship. Acceptance limits should instead reflect the clinical consequences of small exposure differences, within-subject variability, and applicable product-specific guidance. Suitable endpoints and acceptance criteria can be defined only after the clinically relevant exposure differences have been identified. This risk-based approach allows bioequivalence assessment to address clinically meaningful differences rather than pharmacokinetic similarity alone.

In clinical practice, therapeutic drug monitoring (TDM) and empiric dose adjustment remain important approaches to controlling exposure to NTI drugs [11,12,13]. However, a single concentration measurement or fixed dosing rule is often insufficient to account for the combined effects of genetic background, organ function, concomitant medications, pathophysiological status, and formulation characteristics on drug exposure [2,14]. Model-informed precision dosing (MIPD) integrates population pharmacokinetic (PopPK) or pharmacokinetic–pharmacodynamic (PK–PD) models with patient characteristics and individual concentration or biomarker data. Bayesian methods can then be used to update individual parameter estimates over time and translate TDM results into quantitative dosing recommendations [15,16,17]. Physiologically based pharmacokinetic (PBPK) models further incorporate drug physicochemical properties, human physiology, metabolic and transport pathways, food and drug interactions, and formulation dissolution characteristics, thereby supporting extrapolation to special populations, virtual bioequivalence trials, and the development of clinically relevant quality standards [18,19,20,21]. Machine learning can identify complex nonlinear relationships in high-dimensional clinical and real-world data and has been applied to blood concentration prediction, dose estimation, treatment-response assessment, and toxicity risk prediction. Reinforcement learning extends this framework by formulating dose optimization as a sequential decision-making problem that is continuously updated as the patient’s clinical state evolves [22].

Although these modeling approaches offer new avenues for the individualized management of NTI drugs, their clinical translation remains limited by the lack of standardized criteria for model quality assessment, insufficient representativeness of training data, limited external generalizability, poor algorithmic interpretability, and a paucity of prospective evidence on clinical outcomes [23,24]. Existing studies [4] have largely addressed the identification criteria for NTI drugs, BE assessment of generic products, international regulatory differences, or the clinical application of individual modeling approaches in isolation. Consequently, the continuum linking drug regulation, formulation performance, individual exposure, and dynamic dose management has not yet been considered within a unified framework [25]. It is therefore necessary to connect premarketing classification, BE assessment, and formulation quality control with postmarketing TDM, model-informed dosing, and risk management [26].

In response to these limitations, this structured narrative review summarizes definitions of narrow therapeutic index (NTI) drugs, their identifying characteristics, and jurisdiction-specific drug lists. It compares the bioequivalence endpoints, study designs, and acceptance criteria used by major regulatory agencies for generic NTI products. It also examines how these regulatory requirements relate to drug-specific clinical risks. The review then evaluates applications of population pharmacokinetic (PopPK), pharmacokinetic–pharmacodynamic (PK–PD), physiologically based pharmacokinetic (PBPK), machine-learning, and hybrid models. These applications include exposure prediction, dose optimization, extrapolation to special populations, virtual bioequivalence assessment, and clinical risk prediction. By distinguishing product-level regulatory assessment, exposure-level quantitative analysis, and patient-level dose management, this review aims to clarify how these components interact throughout the pharmaceutical lifecycle. It also identifies unresolved challenges in regulatory harmonization, model validation, and clinical translation.

2. Literature Search Strategy

For this structured narrative review, we searched PubMed, Embase, Web of Science Core Collection, Scopus, and the China National Knowledge Infrastructure (CNKI) on 4 September 2026. We also searched the official websites of relevant regulatory agencies. Literature was eligible if published on or before 30 June 2026. The database searches followed two complementary strategies designed to capture both general and drug-specific evidence. The first combined terms related to narrow therapeutic index (NTI) drugs with terms for regulation, bioequivalence, modelling, or precision dosing. The second paired representative drug names with terms for modelling or precision dosing. Each search strategy was adapted to the controlled vocabulary, search fields, and syntax of the corresponding database. Complete database-specific search strategies are provided in Supplementary Methods S1.

The inclusion criteria were as follows: (1) official documents and studies addressing the definitions, identification criteria, drug lists, or regulatory policies for NTI drugs; (2) studies evaluating bioequivalence, interchangeability, within-formulation variability, or switching-related risks of generic NTI products; (3) studies applying PopPK, PK–PD, PBPK, physiologically based biopharmaceutics modeling (PBBM), machine learning, reinforcement learning, or hybrid models to exposure prediction, dose optimization, extrapolation to special populations, virtual bioequivalence assessment, or clinical risk prediction; and (4) methodological, clinical, and external validation studies directly relevant to these topics. Modeling studies were required to report the data source, modeling approach, validation procedures, and outcomes of clinical or regulatory relevance. Duplicate publications, conference abstracts, commentary articles lacking essential methodological information, studies unrelated to human drug dosing or regulatory evaluation, and nonofficial secondary online sources for which original regulatory documents were available were excluded.

The searches of the five databases yielded 23,927 records, which two reviewers independently screened against the inclusion and exclusion criteria. Disagreements were first resolved through discussion; unresolved cases were adjudicated by a third reviewer. Screening yielded 92 eligible journal articles, and searches of official regulatory agency websites identified 21 additional regulatory documents. In total, the review included 113 sources; the complete selection process is shown in Figure S1.

3. Overview of Narrow Therapeutic Index Drugs

The definition and terminology of NTI drugs remain inconsistent across jurisdictions. Different countries and regions use different designations for this drug category. For example, the United States and China commonly use the term “narrow therapeutic index drugs” (NTI drugs), whereas the European Union often uses “NTIDs” (narrow therapeutic index drugs), Japan uses “NTRDs” (narrow therapeutic range drugs), Canada adopts the term “critical dose drugs” (CDDs), and Korea refers to “active substances with a narrow therapeutic index” [4]. Under Title 21 of the Code of Federal Regulations, Section 320.33 (21 CFR 320.33) [27], a drug is considered to have a narrow therapeutic ratio when the median lethal dose (LD₅₀) is less than twice the median effective dose (ED₅₀), or when the minimum toxic concentration (MTC) in blood is less than twice the minimum effective concentration (MEC). This definition has been widely cited and has influenced the development of related criteria in other jurisdictions. Previous studies [3,8,28] have further suggested that the identification and assessment of NTI drugs should generally consider the following characteristics: (1) close proximity between effective and toxic doses, or between effective and toxic concentrations; (2) a risk of treatment failure when blood concentrations fall below the effective range and of serious adverse effects when concentrations exceed the toxic threshold; (3) a need for TDM based on pharmacokinetic or pharmacodynamic measures; (4) low-to-moderate within-subject variability (≤30%); and (5) relatively small dose adjustments in clinical practice, typically <20%. The indicators described above may support preliminary screening for narrow therapeutic index (NTI) drugs or provide supplementary evidence, but they are generally insufficient to establish NTI status on their own. Classification should also consider the exposure–response relationship, therapeutic window, potential clinical consequences, formulation characteristics, and applicable jurisdiction-specific regulatory requirements. NTI drugs should also be distinguished from highly variable drugs and highly variable drug products. NTI status is a pharmacodynamic and clinical-risk property, reflecting the narrow margin between therapeutic and toxic exposure and the consequences of small exposure changes. By contrast, high variability is a pharmacokinetic and statistical property, typically defined by a within-subject coefficient of variation exceeding 30% for reference-product AUC or C_max. An NTI drug is not necessarily highly variable, because small exposure changes may remain clinically important despite low AUC or C_max variability when its therapeutic window is narrow. Conversely, a highly variable drug does not necessarily have a narrow therapeutic index. Its pharmacokinetic parameters may fluctuate substantially across dosing periods without necessarily causing serious consequences for efficacy or safety. These two dimensions should therefore be distinguished explicitly in clinical practice.

Because regulatory approaches and clinical criteria differ across jurisdictions, international NTI drug lists remain heterogeneous. EMA guidance also acknowledges that NTI drugs cannot be defined using a single universal criterion and should instead be assessed on a case-by-case basis, taking into account the clinical characteristics of the individual drug [29]. A systematic comparison of the classification features and drug lists used for NTI drugs is therefore important for informing regulatory decision-making, guiding generic substitution, and safeguarding clinical use. To preserve distinctions among source types, we classified documents related to narrow therapeutic index (NTI) drugs into three categories. These comprised definitions and designation criteria in regulations or general guidance documents, drug- or product-specific regulatory guidance, and clinical classifications or management recommendations from expert consensus statements. The first category describes the general regulatory framework within the relevant jurisdiction. Drug- or product-specific guidance applies only to the drug, dosage form, or product it addresses. Expert consensus statements, by contrast, reflect professional judgement on clinical risk management. National and regional NTI drug lists are summarized in Table 1. Existing lists show that NTI drugs span multiple therapeutic areas, including hematology, cardiovascular medicine, endocrinology, anti-infective therapy, oncology and immunomodulation, neurology, and respiratory medicine. Warfarin [30,31], digoxin, levothyroxine, tacrolimus, cyclosporine, phenytoin, lithium, and theophylline appear repeatedly across multiple jurisdictions. In contrast, some anti-infective, antidiabetic, and anticancer agents are included only in selected regulatory lists or expert consensus documents, indicating that NTI classification continues to be influenced by regional regulatory practice, perceptions of clinical risk, and differences in the available evidence base.

4. Regulatory Requirements and Policy Frameworks for Narrow Therapeutic Index Drugs

Regulatory authorities generally define a generic drug as a medicinal product with the same qualitative and quantitative composition of active substance and the same pharmaceutical form as the reference product, for which bioequivalence has been demonstrated through appropriate bioavailability studies [44]. When applicable regulatory requirements are met, demonstration of bioequivalence generally supports the expectation that the generic and reference products will have comparable efficacy and safety [45]. The core framework used by regulatory authorities to assess bioequivalence for conventional generic drugs is broadly similar across jurisdictions. The area under the concentration–time curve to the last measurable time point (AUC_t) and the maximum plasma concentration (C_max) are typically used as the primary pharmacokinetic endpoints. The 90% confidence intervals for the geometric mean ratios of the test and reference products, calculated after logarithmic transformation, are then compared with prespecified bioequivalence acceptance limits. For most drugs, the conventional acceptance range for both AUC_t and C_max is 80.00–125.00% [4]. For narrow therapeutic index (NTI) drugs, by contrast, even small changes in exposure may affect efficacy or safety. Regulatory frameworks address this risk through different control strategies. Narrower AUC acceptance limits primarily constrain differences in overall exposure between the test and reference products. Whether C_max acceptance limits are similarly narrowed depends on the clinical relevance of peak exposure to efficacy, toxicity, or therapeutic drug monitoring. Requirements for the same drug or product may also differ across jurisdictions because of the applicable regulatory documents, dosage form and strength, or reference product. Further differences may arise from fasting or fed conditions, pharmacokinetic endpoints, and statistical criteria. Table 2 separates general provisions from product-specific requirements to clarify the product-level risk-control considerations reflected in these regulatory differences. The EMA specifies that the acceptance range for AUC_t should be narrowed to 90.00–111.11% for NTI drugs [46]. The conventional range of 80.00–125.00% may generally be retained for C_max; however, when C_max is particularly relevant to safety, efficacy, or TDM, the acceptance range should also be narrowed to 90.00–111.11%. Similar approaches have been adopted by the WHO [47], Singapore [48], the Gulf Cooperation Council (GCC) [41], and the Egyptian Drug Authority (EDA) [39].

Plasma protein binding (PPB) affects the relationship between total and unbound drug concentrations, but does not directly justify narrower bioequivalence acceptance limits for narrow therapeutic index (NTI) drugs. The 90.00–111.11% acceptance range is intended to constrain differences in absorption rate and extent between the test and reference products. The specific limits are determined primarily by the potential efficacy and safety consequences of altered exposure. In crossover bioequivalence studies, both formulations are compared within the same participants under standardized conditions. If protein binding remains approximately linear across the study concentration range and the unbound fraction is stable, unbound and total exposure ratios are generally similar. Current regulatory assessments therefore usually calculate AUC and C_max from total parent-drug concentrations measured in an appropriate biological matrix. When binding is saturable or concentration-dependent, or the unbound fraction is affected by albumin, organ function, metabolites, or concomitant medications, unbound concentrations may provide supplementary evidence. Such measurements may also aid clinical interpretation, but their inclusion in bioequivalence assessment should depend on current drug- and product-specific evidence [49]. This issue warrants particular attention when drug–drug interactions (DDIs) are present. DDIs may alter systemic exposure by affecting protein binding or drug disposition. Bioequivalence assessment should therefore determine whether interaction magnitude depends on the formulation or mode of administration and whether evaluation under specific conditions is warranted. A study in seven healthy participants found that valproate coadministration increased the unbound fraction of phenytoin from 9.6% to 15.6%. The effect involved both displacement from protein-binding sites and metabolic inhibition, suggesting that total concentrations may not fully capture changes in unbound exposure [50]. A four-period crossover study in 18 healthy participants found that voriconazole increased the AUC of immediate-release and extended-release tacrolimus 6.02-fold and 2.62-fold, respectively [51]. This finding suggests that formulation release characteristics may modify the magnitude of a drug interaction. Together, these studies can inform the need for supplementary assessment under specific coadministration conditions and guide therapeutic drug monitoring and dose adjustment. However, current evidence primarily supports drug- or product-specific risk management and does not justify uniform PPB- or DDI-specific bioequivalence acceptance limits for NTI drugs.

The National Institute of Health Sciences (NIHS) of Japan [52] generally retains the conventional bioequivalence acceptance range of 80.00–125.00% for NTI drugs, while applying additional, more stringent criteria. When the initial bioequivalence study includes at least 20 participants (n = 10 per group), or when the combined sample size of the initial and supplementary studies is at least 30, and the test and reference products exhibit highly similar dissolution profiles, AUC_t and C_max may additionally be required to fall within the narrower range of 90.00–111.11%. The Korean Ministry of Food and Drug Safety (MFDS) [53] narrows the acceptance range for AUC_t to 90.00–111.11% for NTI drugs, while retaining 80.00–125.00% for C_max. Health Canada [37] specifies an acceptance range of 90.00–112.00% for AUC_t, while maintaining 80.00–125.00% for C_max.

Compared with regulatory approaches based primarily on fixed acceptance ranges, the NMPA/CDE and the FDA apply more complex study designs and multidimensional assessment criteria to selected NTI drugs. Most jurisdictions generally use a standard 2 × 2 two-sequence, two-period crossover design, whereas China [54] and the United States [32] commonly require a fully replicated four-period crossover design for specific NTI drugs (Figure 1). In a 2 × 2 crossover study, participants are randomly assigned to two treatment sequences and receive the test (T) and reference (R) products in successive periods separated by an adequate washout interval. Because each participant receives each product once, participants serve as their own controls, thereby reducing the influence of between-subject variability. This design is simple, relatively short, and less costly, and is therefore widely used in conventional generic-drug BE studies. However, because T and R products are each administered only once, the 2 × 2 design generally permits estimation only of pooled residual variability and does not allow robust, separate estimation of the within-subject standard deviations for T (SWT) and R (SWR). Consequently, this design may provide insufficient evidence when variability is high or when regulatory assessment requires direct comparison of within-subject variability between the two products.

In a fully replicated four-period crossover design, participants are likewise randomized to two treatment sequences, but each participant receives both T and R twice. Repeated administration of each product enables separate estimation of SWT and SWR within the same study and allows direct comparison of their within-subject variability, thereby providing more precise and robust BE evidence for NTI drugs. Its main limitations are the greater number of dosing and washout periods, longer study duration, higher demands on participant adherence, and substantially increased costs and resource requirements. Because even small differences in exposure may have clinical consequences for NTI drugs, China and the United States use replicated designs for selected products to improve the precision and reliability of BE assessment. The preceding overview outlines the basic structures and intended uses of conventional two-period crossover and fully replicated four-period crossover designs. Specific trial conditions, including the study population, administered dose and product strength, fasting or fed conditions, sampling schedule, analyte selection, and safety monitoring, vary by drug and product. They should be determined from the drug’s pharmacokinetic properties, clinical risks, formulation type, and relevant product-specific guidance. Because these conditions differ across drugs and products, and some case-specific assessment details are not fully disclosed publicly, this review does not prescribe uniform trial conditions. Individual protocols should instead be developed for the specific product and in accordance with the regulatory requirements of the relevant jurisdiction.

For these NTI drugs, China and the United States generally assess three criteria simultaneously. First, reference-scaled average bioequivalence (RSABE) is applied. RSABE is an approach that scales bioequivalence assessment to the within-subject variability of the reference product. It is used primarily when the reference product shows substantial within-subject variability. The method evaluates differences between the log-transformed mean exposure metrics of the test and reference products relative to within-subject variability estimated from replicate reference administrations. This comparison determines whether the difference in mean exposure is acceptable relative to the reference product’s own variability. RSABE therefore calibrates average bioequivalence assessment to reference-product variability rather than simply increasing sensitivity to exposure differences. Second, average bioequivalence (ABE) is used as an additional constraint, requiring the 90% confidence intervals for the geometric mean ratios of AUC_t and C_max to fall within 80.00–125.00%. Third, within-subject variability is compared between T and R, typically requiring the upper bound for the ratio of within-subject standard deviations (σWT/σWR) to be ≤2.5. T and R are considered bioequivalent only when both AUC_t and C_max satisfy all three criteria: RSABE, ABE, and the within-subject variability comparison. This multidimensional framework evaluates not only the similarity of mean exposure but also whether the test product introduces substantially greater within-subject variability than the reference product, thereby providing additional safeguards for the efficacy and safety of NTI drug therapy. It should be noted that BE requirements for NTI drugs in Taiwan and the Hong Kong Special Administrative Region differ from those in mainland China. The Taiwan Food and Drug Administration (TFDA) [55] retains the conventional acceptance range of 80.00–125.00%, whereas the Drug Office of the Hong Kong Department of Health [42] narrows the acceptance range to 90.00–111.11%.

Differences in regulatory policies for the bioequivalence assessment of NTI drugs largely reflect how individual authorities balance the risks of generic substitution against evidentiary reliability, study feasibility, and drug accessibility. On the one hand, because NTI drugs have narrow therapeutic windows, overly simple study designs or permissive acceptance criteria may fail to detect clinically meaningful differences in exposure. On the other hand, highly complex replicated designs and multiple decision criteria can substantially increase study duration, cost, and operational burden, thereby reducing the efficiency and accessibility of generic NTI drug development. Nevertheless, inconsistent regulatory standards may limit the comparability and cross-jurisdictional applicability of bioequivalence findings [14]. The same generic NTI product may yield different regulatory conclusions across regions because of differences in acceptance ranges, study designs, or statistical methods, complicating global development, multinational submissions, and clinical substitution decisions. Accordingly, a major challenge for international regulatory harmonization is to establish a tiered bioequivalence framework that accounts for therapeutic window, exposure–response relationships, within-subject variability, pharmacokinetic characteristics, and clinical risk, while maintaining both scientific rigor and practical feasibility. However, publicly available information does not fully capture product-specific adjustments made during regulatory review or regulatory considerations that are not publicly documented. Accordingly, assessment of a specific product must integrate the drug’s properties, product characteristics, applicable regulatory policies, and relevant guidance documents.

In addition, some healthcare professionals remain cautious about the safety of generic substitution for NTI drugs [44,56]. Conventional ABE primarily compares mean exposure between the test and reference products and does not fully account for differences in product variability, between-subject variability, or within-subject variability. It therefore provides limited information on whether exposure will remain stable in an individual patient after switching formulations. Population bioequivalence (PBE) and individual bioequivalence (IBE) were proposed in the late 1990s as extensions of average bioequivalence (ABE) [57]. PBE evaluates between-formulation differences in both mean exposure and total variability, theoretically addressing population-level “prescribability.” IBE additionally incorporates within-subject variability and subject-by-formulation interaction, theoretically addressing within-person “switchability.” Both methods characterize pharmacokinetic distributions from a statistical perspective. However, meeting their respective criteria does not, by itself, establish the clinical interchangeability of narrow therapeutic index (NTI) drugs across patients. The limited regulatory use of PBE and IBE partly reflects uncertainties in their statistical interpretation and evidentiary basis. Both rely on composite criteria, allowing differences in formulation means and variance components to offset each other within a single statistic. Results may also be sensitive to the scaling constant, model and distributional assumptions, outliers, and the precision of variance and interaction estimates. Moreover, their decision thresholds have not been adequately validated against clinical outcomes. FDA guidance issued in May 2026 states that IBE is not currently used in regulatory assessment. PBE is used mainly in selected in vitro bioequivalence studies requiring comparisons of both means and variability. For example, FDA’s product-specific guidance for fluticasone propionate nasal spray applies a modified PBE approach to compare the means and total variability of cascade-impactor measurements [58]. By contrast, the replicate-design EQUIGEN study [59] compared branded lamotrigine with two generic products in 49 evaluable patients with epilepsy. Comparisons among the three formulations met the ABE, scaled ABE, and IBE criteria, with no detected subject-by-formulation interaction. The former illustrates the use of PBE for in vitro product-performance comparisons. The latter shows how IBE can assess within-subject variability and interaction effects in replicate pharmacokinetic studies. For in vivo applications, these approaches generally require replicate crossover designs, complex statistical models, larger samples, and stringent data-quality controls. These requirements increase both study costs and operational complexity, partly explaining their limited adoption in routine regulatory standards. Future work should refine the statistical methods and study designs, define appropriate use cases and clinical interpretations, and develop a tiered framework integrating mean exposure, within-subject variability, subject-by-formulation interaction, and clinical risk. Such advances could reduce uncertainty surrounding generic substitution of NTI drugs and support greater international harmonization of bioequivalence regulation. Nevertheless, neither PBE nor IBE can currently replace clinical outcome evidence obtained after actual formulation switching.

5. Applications of Modeling Approaches to Narrow Therapeutic Index Drugs

This review proposes a three-level conceptual framework for the lifecycle risk management of narrow therapeutic index (NTI) drugs. The framework links product-, population-, and individual-level evidence to support decisions across different stages of drug development and clinical use. Its scientific validity and clinical utility, however, require confirmation through further clinical studies and real-world data. At the product level, assessment focuses on NTI classification, formulation performance, and bioequivalence between test and reference products. It examines whether product differences or formulation changes could shift systemic exposure beyond a prespecified acceptable range. At the population level, clinical pharmacokinetic analyses, population pharmacokinetic (PopPK) modelling, and physiologically based pharmacokinetic or biopharmaceutics modelling (PBPK/PBBM) characterize exposure distributions. These approaches examine the effects of formulation factors, drug interactions, and population characteristics. They also identify high-risk scenarios not adequately captured by conventional bioequivalence studies. At the individual level, therapeutic drug monitoring, patient covariates, and Bayesian forecasting guide dose adjustment using measured drug concentrations or pharmacodynamic responses. These three evidence streams serve distinct decision functions, with systemic exposure providing the common quantitative link between them. Premarketing assessments and studies following formulation changes primarily control product-related differences in exposure. Postmarketing surveillance and model updating instead identify and manage between-patient and within-patient exposure variability. Evidence generated during clinical use can, in turn, refine drug interaction management, formulation risk assessment, and product-specific regulatory requirements. Figure 2 summarizes the major factors and pathways that may increase toxicity risk during treatment with NTI drugs.

Precision dosing converts patient-specific information and longitudinal monitoring data into actionable dosing decisions. Current strategies can be grouped according to their decision basis and mode of implementation. These include stratified dosing based on prescribing information and patient characteristics, empirical titration guided by efficacy or toxicity, and therapeutic drug monitoring-guided dose adjustment. Pharmacogenomics-guided dosing and model-informed precision dosing (MIPD) represent two further strategies. For narrow therapeutic index drugs, MIPD integrates population-level priors, patient covariates, and longitudinal concentration or biomarker data. It can estimate individual pharmacokinetic parameters and exposure, compare candidate regimens, and iteratively update dose recommendations through Bayesian forecasting. The appropriate model depends on the decision problem, with population pharmacokinetic models primarily supporting individual parameter estimation and adaptive dose adjustment. Physiologically based pharmacokinetic and biopharmaceutics models (PBPK/PBBM) characterize how physiology, pharmacokinetic drug interactions, and formulation factors shape exposure. Machine-learning models can identify complex nonlinear relationships in high-dimensional data, whereas hybrid models combine approaches to address specific questions (Table 3). MIPD should therefore complement rather than replace integrated clinical judgement. It provides quantitative, iterative support for dose optimization when exposure or response targets are defined, models are adequately validated, and reliable individual-level data are available.

For narrow therapeutic index drugs, model performance should be evaluated against drug-specific target exposure ranges or clinical decision thresholds. Acceptable error margins, including the tolerated magnitude and direction of prediction error, should be prespecified. Evaluation should separately quantify prediction bias and precision, together with the probability that model-informed dosing produces exposure below or above the target range. Statistical measures such as mean absolute error (MAE) and root mean square error (RMSE) can quantify overall prediction error. However, these measures alone cannot establish whether a model is clinically acceptable. This judgement should also consider external validation, sensitivity analyses under clinically relevant scenarios, and prospective clinical evaluation. The relevant decision thresholds should be derived from drug-specific evidence linking exposure to efficacy and toxicity, the therapeutic exposure range, and dose-adjustment objectives.

6. Conclusions

Narrow therapeutic index drugs remain challenging across regulatory evaluation, generic substitution, and clinical dose management because small changes in exposure may compromise efficacy or safety. Current regulatory frameworks differ substantially in drug classification, bioequivalence study design, acceptance criteria, and variability assessment, limiting cross-jurisdictional consistency. PopPK models coupled with Bayesian forecasting currently have a comparatively mature evidence base for therapeutic drug monitoring and individualized dose adjustment. PBPK models primarily support mechanistic extrapolation, drug-interaction prediction, virtual bioequivalence assessment, and evaluation of formulation performance. Machine-learning and hybrid models show potential for predicting exposure and dose requirements, but current evidence comes mainly from retrospective analyses and simulation studies. Their prospective clinical benefit and suitability for regulatory use therefore remain to be established. Future efforts should focus on harmonizing risk-based bioequivalence requirements, standardizing model development and validation, and integrating regulatory, formulation, pharmacokinetic, and patient-level information throughout the drug lifecycle. Such an integrated framework may improve the reliability of generic substitution and support safer, more precise use of narrow therapeutic index drugs.

Author Contributions

Conceptualization, M.L. and L.Z.; methodology, M.L. and Y.C.; literature search, M.L., Y.C., Y.J., Z.W., L.X. and H.X.; validation, M.L., Y.C. and L.Z.; formal analysis, M.L., Y.C., J.L., Y.J. and Z.W.; investigation, M.L., Y.C., Y.J., Z.W., L.X. and H.X.; resources, L.Z.; data curation, M.L., Y.C., Y.J., Z.W., L.X., J.L. and H.X.; writing—original draft preparation, M.L.; writing—review and editing, Y.C., Y.J., Z.W., L.X., H.X. and L.Z.; visualization, M.L., Y.C. and Y.J.; supervision, L.Z.; project administration, M.L. and L.Z.; funding acquisition, L.Z. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding Statement

This research was funded by the China Medical Education Association, grant number [2026] 523.