Tutorials › Biostatistics › Accelerated Titration Designs

Home  /  Tutorials  /  Phase I Study Design

Phase I & Dose Escalation Design

Accelerated Titration Designs: A Worked Example

A computational guide to accelerated titration designs for Phase I oncology trials, including rapid dose escalation, single-patient cohorts, intrapatient dose escalation, dose-limiting toxicity, moderate toxicity, transition to conventional escalation, dose-toxicity modeling, and a complete worked example.

Advanced 18 min read

What You'll Learn

  • Why accelerated titration designs were developed for Phase I oncology trials
  • How the rapid single-patient escalation phase works
  • How intrapatient dose escalation can reduce treatment at very low doses
  • How DLT and moderate toxicity determine when acceleration stops
  • How a dose-toxicity model can be used to estimate the recommended dose
  • How accelerated titration differs from conventional 3+3 escalation

Introduction

The central objective of a Phase I oncology trial is to learn how toxicity changes with dose while identifying a dose that can reasonably be taken forward into later development.

Traditional dose-escalation designs can expose substantial numbers of patients to doses that are far below biologically active or clinically useful levels. This can make the trial longer and can reduce the amount of useful information obtained from each patient.

The accelerated titration design was developed to address this problem by allowing substantially faster escalation during the early portion of a Phase I study.

The original accelerated titration framework introduced by Simon and colleagues combined a rapid initial escalation phase, single-patient cohorts, intrapatient dose escalation, and a transition to more conventional escalation once toxicity information became concerning. [1]

Key idea: An accelerated titration design does not simply "increase the dose faster." It uses a prespecified set of toxicity rules to determine when rapid escalation is appropriate and when the trial should revert to a more conservative dose-escalation strategy.

The Problem with Slow Dose Escalation

Consider a conventional Phase I trial in which patients are treated in cohorts of three and the dose increases by approximately 40% between dose levels.

If the starting dose is far below the dose associated with meaningful pharmacologic activity, several cohorts may be exposed to doses that produce little toxicity and potentially little therapeutic effect.

For example, suppose the planned dose sequence is:

$$ 10,\quad14,\quad20,\quad28,\quad40,\quad56,\quad78,\quad109,\ldots $$

A patient entering the study at 10 mg may be many dose levels away from the eventual recommended dose.

This creates an important ethical and scientific problem: patients may spend valuable treatment time at doses that are unlikely to provide meaningful therapeutic exposure.

Accelerated titration designs attempt to move through these low-dose levels more rapidly while retaining predefined toxicity safeguards.

What Is an Accelerated Titration Design?

An accelerated titration design is a Phase I dose-escalation design in which the early part of the trial uses substantially faster escalation than the later part.

The defining characteristics of the original framework include:

  • Rapid dose escalation during the initial phase.
  • One patient per dose level during the accelerated phase.
  • Intrapatient dose escalation when prior-course toxicity is sufficiently low.
  • Prespecified rules for identifying dose-limiting and moderate toxicity.
  • Transition to conventional cohort-based escalation when toxicity becomes concerning.
  • Optional use of a dose-toxicity model to characterize the relationship between dose and toxicity.

The original paper evaluated several accelerated designs using simulations based on data from 20 Phase I trials involving nine drugs. [2]

Accelerated Titration Is a Family of Designs

It is important not to think of "the accelerated titration design" as one single immutable algorithm.

Simon and colleagues evaluated several designs that differed in how quickly the dose was increased and how toxicity was monitored.

The original designs included a conventional control design and three accelerated alternatives.

Feature Conventional Control Accelerated Designs
Patients per early cohort 3–6 1
Rapid early escalation No Yes
Intrapatient escalation No Yes
Early escalation step Approximately 40% 40% or 100%, depending on design
Transition after concerning toxicity Not applicable Yes
Conventional escalation after transition Yes Yes

The original simulations found that the accelerated designs could reduce the number of patients treated at very low toxicity levels and reduce the average number of patients required compared with the conventional control design. [3]

Design 4: The Most Aggressive Accelerated Framework

The accelerated design most commonly discussed in reviews of the original framework is Design 4.

Its initial escalation phase uses:

  • One patient per dose level.
  • 100% dose increases during the accelerated phase.
  • Intrapatient escalation when toxicity is sufficiently low.

The accelerated phase continues until predefined toxicity criteria indicate that escalation should become more conservative.

The original framework then transitions to standard cohorts of three to six patients with smaller dose increments. [4]

Important: The 100% escalation rule is specific to particular accelerated titration designs. It should not be interpreted as a general recommendation to double doses in Phase I trials. The exact escalation rule must be selected and prespecified as part of the protocol.

Basic Notation

Let:

  • \(d_j\) = dose level \(j\)
  • \(T_{ij}\) = toxicity observed for patient \(i\) at dose level \(j\)
  • \(D_i\) = dose received by patient \(i\)
  • \(Y_i\) = indicator of a dose-limiting toxicity
  • \(M_i\) = indicator of moderate toxicity
  • \(n_j\) = number of patients treated at dose level \(j\)
  • \(p(d)\) = probability of DLT at dose \(d\)

The dose-toxicity relationship can be represented abstractly as:

$$ p(d)=P(\text{DLT}\mid D=d) $$

The practical goal is not necessarily to find the largest dose that can be administered.

Instead, the goal is generally to identify a dose or dose range with an acceptable balance between toxicity, pharmacologic activity, and clinical development objectives.

What Is Dose-Limiting Toxicity?

A dose-limiting toxicity, or DLT, is a protocol-defined toxicity severe enough to limit further dose escalation.

The exact definition depends on the disease, drug, treatment schedule, observation window, and protocol.

For example, a protocol might define certain severe hematologic or nonhematologic toxicities as DLTs during the first treatment cycle.

The important statistical point is that the dose-escalation algorithm operates on a prespecified toxicity classification.

Do not assume that "DLT" has one universal definition. The exact DLT criteria must come from the protocol. An accelerated titration calculation cannot be interpreted correctly without knowing how toxicity was classified.

What Is Moderate Toxicity?

The accelerated titration framework also uses a category of toxicity that is less severe than DLT but sufficiently concerning to indicate that the rapid escalation phase should end.

This is often referred to as moderate toxicity.

The precise definition is protocol-specific.

Conceptually, the design distinguishes among:

Toxicity Category Meaning for Escalation
None or mild Rapid escalation may continue
Moderate Accelerated phase may terminate
DLT Escalation becomes substantially more conservative

This intermediate toxicity category is important because waiting for a DLT before slowing escalation may expose patients to unnecessarily high doses.

The Three Core Ideas

The accelerated titration framework can be understood through three central ideas.

1
Rapid escalation: move quickly through dose levels while toxicity remains minimal.
2
Intrapatient escalation: allow an individual patient with low toxicity to receive higher doses in subsequent treatment courses.
3
Conservative escalation: once meaningful toxicity appears, transition to a conventional cohort-based approach.

Why Intrapatient Dose Escalation Matters

Traditional Phase I studies generally assign a patient to a dose level and keep that patient at that dose.

That means a patient who tolerates the starting dose extremely well may remain at a dose that is substantially below the eventual recommended dose.

Accelerated titration designs can instead allow that patient to move upward in dose during subsequent treatment courses when prior toxicity is sufficiently low.

For example, a patient might receive:

$$ 20\rightarrow40\rightarrow80\rightarrow160\text{ mg} $$

provided that the protocol's toxicity criteria permit those escalations.

Clinical interpretation: Intrapatient escalation gives some patients the opportunity to receive doses closer to the biologically active range rather than remaining at a low initial dose for the duration of the trial.

A Simple Dose Escalation Sequence

Suppose the starting dose is 10 mg and the accelerated phase uses 100% dose increments.

The sequence could be:

Dose Level Dose Increase
1 10 mg —
2 20 mg 100%
3 40 mg 100%
4 80 mg 100%
5 160 mg 100%
6 320 mg 100%

This is intentionally much faster than a conventional 40% escalation sequence.

If the first patient at 10 mg experiences no concerning toxicity, the next patient may enter at 20 mg.

If the 20 mg patient also experiences minimal toxicity, escalation can continue.

However, once the predefined toxicity criteria are met, the accelerated phase ends.

Why 100% Escalation Is Possible in the Accelerated Phase

The rationale is not that the investigators believe the dose-toxicity curve is perfectly known.

The rationale is that very low doses are expected to be far below the toxicity region, so spending many patients at each low dose may provide relatively little information about the clinically relevant portion of the dose-toxicity curve.

The design therefore concentrates more of the trial's information near the region where toxicity begins to emerge.

The safety rules are what determine when the trial must stop moving rapidly.

When Does the Accelerated Phase Stop?

The exact transition rule depends on the selected accelerated titration design.

In the original designs, the accelerated phase was terminated when toxicity information became sufficiently concerning—for example, after a patient experienced a DLT or after the prespecified number of patients experienced moderate toxicity. [5]

Once that occurs, escalation reverts to a more conventional approach.

$$ \text{Rapid escalation} \quad\longrightarrow\quad \text{toxicity signal} \quad\longrightarrow\quad \text{conventional escalation} $$

The Complete Algorithm

1
Treat one patient at the current dose.
2
Evaluate the prespecified toxicity window.
3
If toxicity remains below the acceleration threshold, escalate rapidly.
4
If the acceleration threshold is crossed, terminate the accelerated phase.
5
Move to a conventional cohort-based escalation strategy.
6
Continue dose escalation or de-escalation according to the protocol.
7
Estimate the recommended dose using the prespecified statistical and clinical criteria.

A Complete Worked Example

Consider a hypothetical first-in-human oncology study evaluating an oral targeted therapy.

The study begins at 10 mg once daily.

The accelerated phase will use one patient per dose level and 100% dose increments.

For illustration, assume the protocol defines:

  • A DLT as a protocol-defined severe treatment-related toxicity during the DLT observation period.
  • Moderate toxicity as a protocol-defined toxicity below the DLT threshold but sufficient to terminate accelerated escalation.
  • One patient per dose level during the accelerated phase.
  • Rapid escalation while no predefined toxicity trigger has occurred.
  • Transition to three-patient cohorts once the accelerated phase ends.
  • 40% dose increments during the conventional phase.
Important: The toxicity definitions and exact stopping rules below are illustrative. They demonstrate how the design works mathematically and operationally; they are not a universal clinical dosing rule.

Step 1: Define the Dose Levels

The planned accelerated dose sequence is:

$$ 10,\quad20,\quad40,\quad80,\quad160,\quad320,\quad640\text{ mg} $$

Each successive dose represents a 100% increase.

The first patient therefore receives:

$$ D_1=10\text{ mg} $$

If the toxicity criteria permit escalation, the next patient receives:

$$ D_2=20\text{ mg} $$

and so forth.

Step 2: Patient 1 at 10 mg

Suppose Patient 1 receives 10 mg and experiences no treatment-related toxicity above the mild category.

The acceleration rule therefore permits escalation.

$$ 10\rightarrow20\text{ mg} $$

Patient 1 may also be eligible for intrapatient escalation at the next treatment course if the protocol permits it.

Step 3: Patient 2 at 20 mg

Suppose Patient 2 receives 20 mg and again experiences no meaningful toxicity.

The next new patient can therefore enter at:

$$ D_3=40\text{ mg} $$

The study has now moved through two dose levels without requiring three-patient cohorts at each level.

Step 4: Patient 3 at 40 mg

Suppose Patient 3 experiences Grade 1 treatment-related fatigue.

Under the hypothetical toxicity classification, this remains below the moderate-toxicity threshold.

Therefore:

$$ 40\rightarrow80\text{ mg} $$

The accelerated phase continues.

Step 5: Patient 4 at 80 mg

Suppose Patient 4 experiences a moderate treatment-related toxicity.

The accelerated-phase stopping criterion has now been reached.

Therefore, the study does not proceed directly from 80 mg to 160 mg using the accelerated rule.

Instead:

$$ \text{Accelerated phase ends at 80 mg} $$

and the trial transitions to conventional escalation.

Step 6: Transition to Conventional Escalation

Suppose the protocol specifies three-patient cohorts with approximately 40% dose increments after the accelerated phase.

A subsequent sequence might therefore be:

$$ 80 \rightarrow 112 \rightarrow 157 \rightarrow 220 \rightarrow\cdots $$

The exact dose levels would normally be prespecified or determined according to the protocol's dose-level construction rules.

The key point is that the trial has shifted from:

$$ \text{one patient + rapid escalation} $$

to:

$$ \text{multiple patients + conservative escalation} $$

The Patient-Level Dataset

A simple representation of the hypothetical trial is:

Patient Initial Dose Worst Toxicity Accelerated Phase? Decision
1 10 mg None Yes Escalate
2 20 mg None Yes Escalate
3 40 mg Mild Yes Escalate
4 80 mg Moderate Yes Stop acceleration
5–7 80 mg Under evaluation No Conventional cohort

The important feature is that only one patient was required at each dose level during the initial accelerated phase.

Intrapatient Dose Escalation

Now suppose Patient 1 has completed the first treatment course at 10 mg with no clinically meaningful toxicity.

The protocol may permit Patient 1 to receive a higher dose during the next course.

For example:

$$ 10\rightarrow20\rightarrow40\text{ mg} $$

This means the patient is not necessarily locked into the initial dose.

The dose received by the patient can therefore evolve over time:

$$ D_{i,1}

provided that each escalation is allowed by the protocol.

Why Intrapatient Escalation Can Be Efficient

Suppose the eventual recommended dose is 160 mg.

Under a fixed-dose design, a patient who begins at 10 mg might receive several cycles at a dose far below the eventual recommended dose.

Under intrapatient escalation, that same patient may be able to receive:

$$ 10\rightarrow20\rightarrow40\rightarrow80\rightarrow160\text{ mg} $$

if toxicity remains acceptably low.

This can reduce the amount of treatment delivered at very low doses.

But there is a tradeoff: Intrapatient escalation can make interpretation of cumulative toxicity more complicated because toxicity observed later in treatment may reflect both dose and duration of exposure.

Cumulative Toxicity

One of the important challenges in Phase I oncology trials is that toxicity may not be completely determined by the current dose.

A patient's toxicity may depend on:

  • Current dose
  • Previous doses
  • Number of treatment cycles
  • Cumulative exposure
  • Patient-specific susceptibility

A conceptual representation is:

$$ P(\text{DLT}) = f(\text{current dose}, \text{cumulative exposure}, \text{patient characteristics}) $$

This is one reason why dose-toxicity modeling can be useful when analyzing accelerated titration studies.

Interpatient and Intrapatient Variability

Two patients receiving the same dose may not have the same probability of toxicity.

Let:

$$ p_i(d)=P(Y_i=1\mid D_i=d) $$

Then generally:

$$ p_i(d)\ne p_j(d) $$

for two different patients \(i\) and \(j\).

There can also be variation within a patient across treatment courses.

This distinction is important because the observed toxicity profile represents a combination of dose effects and patient-level variability.

Why a Dose-Toxicity Model Can Help

Rather than evaluating each dose level in isolation, investigators can model the probability of toxicity as a continuous function of dose.

A simple logistic model is:

$$ \operatorname{logit}\{p(d)\} = \alpha+\beta\log(d) $$

where:

  • \(\alpha\) controls the baseline toxicity level.
  • \(\beta\) describes how toxicity changes with dose.

The predicted probability of DLT is then:

$$ p(d) = \frac{\exp(\alpha+\beta\log d)} {1+\exp(\alpha+\beta\log d)} $$

This allows the entire dose-toxicity curve to be estimated rather than simply listing the observed proportion of DLTs at each dose.

Defining a Target Toxicity Probability

Suppose the investigator defines a target DLT probability:

$$ \phi=0.25 $$

The target dose can then be defined conceptually as the dose satisfying:

$$ p(d^*)=\phi $$

Under the logistic model:

$$ \frac{\exp(\alpha+\beta\log d^*)} {1+\exp(\alpha+\beta\log d^*)} = \phi $$

Taking the logit transformation gives:

$$ \alpha+\beta\log d^* = \log\left(\frac{\phi}{1-\phi}\right) $$

Therefore:

$$ d^* = \exp\left[ \frac{ \log\left(\frac{\phi}{1-\phi}\right)-\alpha }{\beta} \right] $$

This is a model-based illustration of how a dose associated with a target toxicity probability can be obtained.

Important distinction: The original accelerated titration framework was accompanied by a dose-toxicity model incorporating interpatient variation, intrapatient variation, and cumulative toxicity. A simple logistic model such as the one above is useful for teaching the core concept, but it is not a substitute for the full protocol-specific analysis model.

Worked Dose-Toxicity Example

Suppose a hypothetical analysis produces the following estimated toxicity curve:

Dose Estimated DLT Probability
20 mg 0.03
40 mg 0.07
80 mg 0.15
160 mg 0.27
320 mg 0.44
640 mg 0.66

If the target DLT probability is:

$$ \phi=0.25 $$

then 160 mg is close to the target because:

$$ p(160)\approx0.27 $$

while 320 mg appears substantially more toxic:

$$ p(320)\approx0.44 $$

A dose around 160 mg might therefore warrant consideration as a recommended dose candidate, subject to the full clinical, pharmacologic, safety, and efficacy assessment.

The Recommended Dose Is Not Automatically the MTD

A critical modern interpretation is that the dose taken into Phase II does not necessarily have to be the absolute maximum tolerated dose.

For cytotoxic therapies, MTD historically played a central role in dose selection.

For targeted therapies and other modern agents, the clinically useful dose may instead be informed by:

  • Pharmacokinetics
  • Pharmacodynamics
  • Target engagement
  • Biomarker response
  • Exposure-response relationships
  • Chronic toxicity
  • Preliminary antitumor activity

Therefore, the dose-toxicity curve is one component of dose selection rather than necessarily the entire decision.

Accelerated Titration vs. 3+3

The accelerated titration design is often compared with the traditional 3+3 design.

Feature 3+3 Accelerated Titration
Initial cohort size Usually 3 1 during accelerated phase
Early dose increments Relatively conservative Can be substantially larger
Intrapatient escalation Typically no Yes in the accelerated framework
Moderate-toxicity trigger Not a defining feature Yes
Transition to conventional escalation Not applicable Yes
Dose-toxicity model Not inherent Can accompany design

The accelerated titration design is therefore not simply a larger step size applied to a 3+3 trial.

It changes the structure of patient assignment, the use of intrapatient escalation, and the conditions under which the trial transitions to a more conservative strategy.

Why the Comparison Is More Subtle Than "Which Is Better?"

It is tempting to conclude that faster escalation is automatically better.

That is not the appropriate interpretation.

The design objective is to obtain useful dose-toxicity information while protecting patients from unacceptable exposure.

Accelerated titration can reduce exposure at very low doses, but it does so by accepting larger early dose increments.

The safety rules are therefore critical.

Design principle: The accelerated phase is deliberately aggressive only while the observed toxicity information supports that strategy. Once the toxicity signal becomes meaningful, the design becomes more conservative.

How the Accelerated Phase Saves Patients

Suppose the eventual recommended dose is approximately 160 mg.

A conventional escalation scheme might require patients to populate several intermediate dose levels.

An accelerated design can instead move:

$$ 10\rightarrow20\rightarrow40\rightarrow80\rightarrow160 $$

using single-patient cohorts during the early phase.

This reduces the number of patients required to characterize the very-low-dose portion of the dose range.

The original simulation study found that the average number of patients required was lower for the accelerated designs than for the conventional control design. [6]

A Patient-by-Patient Worked Example

Consider the following hypothetical sequence.

Patient Dose Toxicity Decision
1 10 mg None Escalate
2 20 mg None Escalate
3 40 mg Mild Escalate
4 80 mg None Escalate
5 160 mg Moderate End acceleration
6–8 160 mg Under evaluation Conventional cohort
9–11 224 mg Under evaluation Conventional cohort

Notice what happened.

The study reached 160 mg after only five initial patients.

Once a moderate toxicity signal appeared, the design stopped using 100% dose increments.

The trial then moved to a more conservative escalation strategy.

What If a DLT Occurs?

Suppose instead that the patient at 160 mg experiences a DLT.

The accelerated phase would terminate.

The study would then follow the protocol-defined de-escalation and cohort expansion rules.

For example, the protocol might require evaluation of additional patients at 160 mg or de-escalation to the previous dose.

The exact decision is protocol-specific.

There is no universal "DLT = do X" rule. The accelerated titration framework provides a design philosophy and prespecified escalation structure. The actual dose decision after a DLT must be defined in the study protocol.

What If Two Patients Experience Moderate Toxicity?

Suppose the accelerated phase permits only one patient with moderate toxicity before the rapid escalation phase must stop.

If a second patient experiences moderate toxicity, the acceleration criterion has clearly been exceeded.

The trial should therefore already have transitioned to the conventional escalation framework.

This illustrates why moderate toxicity is useful as an early warning signal.

The Dose Assignment Function

Conceptually, the dose assignment can be represented as a function:

$$ D_{i+1} = g(D_i,T_i,\text{phase}) $$

where:

  • \(D_i\) is the current dose.
  • \(T_i\) represents the observed toxicity information.
  • phase indicates whether the trial is in accelerated or conventional escalation.

During the accelerated phase, the function may produce a large dose increase:

$$ D_{i+1}=2D_i $$

whereas during the conventional phase the increase might be approximately:

$$ D_{i+1}=1.4D_i $$

The transition between these rules is controlled by toxicity.

A Generic R Implementation

A simplified representation of the accelerated escalation logic can be written in R. 

dose <- 10

dose_levels <- c(
  10,
  20,
  40,
  80,
  160,
  320,
  640
)

accelerated <- TRUE

for (i in seq_along(dose_levels)) {

  current_dose <- dose_levels[i]

  print(current_dose)

  # In an actual trial, toxicity would be
  # observed and classified here.

  # Example:
  # if (DLT || moderate_toxicity_trigger) {
  #   accelerated <- FALSE
  # }
}

This is not a clinical dose-assignment program.

It illustrates the computational structure of the design:

$$ \text{assign dose} \rightarrow \text{observe toxicity} \rightarrow \text{classify toxicity} \rightarrow \text{select next dose} $$

Implementing the Escalation Rule

A simple educational function can represent the transition from accelerated to conventional escalation. 

next_dose <- function(
  current_dose,
  moderate_trigger = FALSE,
  dlt = FALSE
) {

  if (dlt || moderate_trigger) {

    return(
      current_dose * 1.40
    )

  } else {

    return(
      current_dose * 2.00
    )
  }
}

next_dose(40)
next_dose(40, moderate_trigger = TRUE)
next_dose(40, dlt = TRUE)

The resulting conceptual behavior is:

Current Dose Toxicity Signal Illustrative Next Dose
40 mg None 80 mg
40 mg Moderate toxicity 56 mg
40 mg DLT Protocol-defined conservative action

Again, the numerical rules in this code are educational examples rather than a recommendation for clinical use.

Modeling the Probability of DLT

Suppose we observe \(Y_i=1\) if patient \(i\) experiences a DLT and \(Y_i=0\) otherwise.

A simple logistic dose-toxicity model is:

$$ Y_i\sim\operatorname{Bernoulli}(p_i) $$

with:

$$ \operatorname{logit}(p_i) = \alpha+\beta\log(D_i) $$

The likelihood contribution from patient \(i\) is:

$$ L_i(\alpha,\beta) = p_i^{Y_i} (1-p_i)^{1-Y_i} $$

The full likelihood is:

$$ L(\alpha,\beta) = \prod_{i=1}^{n} p_i^{Y_i} (1-p_i)^{1-Y_i} $$

Maximum likelihood estimation can then be used to estimate the dose-toxicity relationship.

R Code for a Simple Dose-Toxicity Model

dose <- c(
  20, 40, 80, 160, 320, 320,
  80, 160, 160, 320, 640
)

dlt <- c(
  0, 0, 0, 1, 1, 0,
  0, 0, 1, 1, 1
)

fit <- glm(
  dlt ~ log(dose),
  family = binomial()
)

summary(fit)

The fitted model can then be used to estimate the DLT probability at each candidate dose. 

new_data <- data.frame(
  dose = c(40, 80, 160, 320, 640)
)

new_data$predicted_dlt <- predict(
  fit,
  newdata = new_data,
  type = "response"
)

new_data

Finding a Target Dose from the Model

Suppose the target DLT probability is:

$$ \phi=0.25 $$

One simple approach is to evaluate the predicted DLT probability over a grid of candidate doses and select the dose closest to the target. 

candidate_doses <- c(
  40, 80, 160, 320, 640
)

predicted <- predict(
  fit,
  newdata = data.frame(
    dose = candidate_doses
  ),
  type = "response"
)

target <- 0.25

result <- data.frame(
  dose = candidate_doses,
  predicted_dlt = predicted
)

result$distance <- abs(
  result$predicted_dlt - target
)

result[
  which.min(result$distance),
]

This provides a simple computational illustration of selecting a dose near a target toxicity probability.

Do not confuse this with the original accelerated titration algorithm. The model-based calculation is an analysis component. The actual dose assignment and escalation rules must remain consistent with the prespecified trial design.

Maximum Tolerated Dose

Historically, Phase I oncology studies often focused on identifying the maximum tolerated dose, or MTD.

Conceptually, the MTD can be represented as a dose associated with a specified toxicity probability:

$$ P(\text{DLT}\mid D=\text{MTD})=\phi $$

where \(\phi\) is the target toxicity probability.

The precise target depends on the development setting and protocol.

Accelerated titration designs were developed in an era when identifying an MTD was a central Phase I objective.

Modern dose-escalation studies may instead seek a recommended Phase II dose using a broader body of evidence.

Recommended Phase II Dose

A recommended Phase II dose can incorporate information beyond toxicity.

For example:

$$ \text{RP2D} = f( \text{toxicity}, \text{PK}, \text{PD}, \text{exposure}, \text{biomarkers}, \text{activity} ) $$

This is particularly important for targeted agents where biological activity may occur at doses substantially below the dose producing the maximum tolerated toxicity.

Accelerated Titration and Pharmacokinetics

Pharmacokinetic data can provide additional information during dose escalation.

For example, investigators may characterize exposure using:

  • AUC
  • Cmax
  • Ctrough
  • Half-life
  • Accumulation ratio

Suppose exposure is approximately proportional to dose:

$$ AUC\propto D $$

Then rapid dose escalation may also produce rapid increases in exposure.

If instead:

$$ AUC\propto D^\gamma $$

with:

$$ \gamma>1 $$

then exposure increases more rapidly than dose.

This can be particularly important when evaluating large dose increments during an accelerated phase.

Exposure-Based Dose Escalation

A pharmacologically guided strategy may use exposure information to supplement toxicity-based escalation.

For example, suppose a preclinical exposure target is:

$$ AUC_{\text{target}}=500 $$

and the observed human exposure at 80 mg is:

$$ AUC(80)=120 $$

while:

$$ AUC(160)=245 $$

and:

$$ AUC(320)=510 $$

Then 320 mg may be much more informative pharmacologically than 80 mg even if both doses are tolerated.

This illustrates why modern Phase I programs frequently integrate PK and PD information with the toxicity-based escalation framework.

Operating Characteristics of Accelerated Titration

Unlike a simple fixed-sample hypothesis test, an accelerated titration design does not have a single type I error and power pair that completely describes its performance.

Useful operating characteristics include:

  • Number of patients treated at each dose level
  • Probability of reaching each dose level
  • Probability of experiencing DLT
  • Probability of overdosing
  • Probability of selecting a dose near the target
  • Number of patients treated below the eventual recommended dose
  • Total number of patients enrolled
  • Trial duration
  • Number of dose levels explored

These quantities are generally evaluated through simulation.

Why Simulation Is Important

The dose-escalation process is sequential and depends on the toxicity history observed during the trial.

For example:

$$ D_{i+1} = g(D_i,T_1,\ldots,T_i) $$

Therefore, the distribution of the next dose depends on the entire toxicity history.

Simulation allows investigators to repeatedly generate hypothetical trials and evaluate how the design behaves under different true dose-toxicity relationships.

A Simple Simulation Framework

Suppose the true DLT probability at dose \(d\) is known.

For example:

Dose True DLT Probability
10 mg 0.01
20 mg 0.02
40 mg 0.05
80 mg 0.10
160 mg 0.25
320 mg 0.45
640 mg 0.70

For each simulated patient:

$$ Y_i\sim\operatorname{Bernoulli}\{p(D_i)\} $$

The simulated toxicity result is then fed back into the dose-assignment algorithm.

R Simulation Skeleton

set.seed(123)

dose_levels <- c(
  10, 20, 40, 80, 160, 320, 640
)

true_dlt <- c(
  0.01, 0.02, 0.05,
  0.10, 0.25, 0.45, 0.70
)

n_sim <- 1000

results <- vector(
  "list",
  n_sim
)

for (sim in seq_len(n_sim)) {

  selected_doses <- numeric()
  observed_dlt <- numeric()

  # Trial-specific dose assignment
  # would be performed here.

  results[[sim]] <- data.frame(
    dose = selected_doses,
    dlt = observed_dlt
  )
}

A complete simulation would implement the exact protocol-specific escalation rules, toxicity windows, cohort rules, intrapatient escalation rules, and stopping criteria.

Evaluating the Number of Patients at Low Doses

One of the key motivations for accelerated titration is to reduce the number of patients treated at doses far below the useful range.

Define:

$$ N_{\text{low}} = \sum_{i=1}^{N} I(D_i

where \(d_{\text{target}}\) represents a dose range considered clinically relevant.

Across simulated trials, investigators can estimate:

$$ E(N_{\text{low}}) $$

and compare this quantity with alternative dose-escalation designs.

Evaluating Overdose

A design must also be evaluated for the possibility that patients are exposed to excessively toxic doses.

Define an overdose threshold \(d_{\text{OD}}\).

Then:

$$ N_{\text{OD}} = \sum_{i=1}^{N} I(D_i>d_{\text{OD}}) $$

A useful operating characteristic is:

$$ P(N_{\text{OD}}>0) $$

which represents the probability that at least one patient receives a dose above the predefined overdose region.

Efficiency cannot be evaluated without safety. A design that reaches high doses quickly is not automatically efficient if it substantially increases the probability of excessive toxicity.

Accelerated Titration Is Not a Model-Based Design

This distinction is important.

Accelerated titration designs are generally classified as rule-based designs because dose assignment follows prespecified rules based on observed toxicity.

A dose-toxicity model can accompany the design and can be used to analyze the trial, but the dose assignment itself need not be determined by continuously updated model predictions.

This differs from methods such as:

  • Continual reassessment methods
  • Bayesian logistic models
  • Bayesian optimal interval designs
  • Escalation with overdose control

where model-based posterior or likelihood information directly determines dose assignment.

Accelerated Titration vs. CRM

Feature Accelerated Titration CRM
General classification Rule-based Model-based
Early single-patient escalation Yes Possible, depending on design
Intrapatient escalation Yes Not inherent
Explicit dose-toxicity model for assignment Not required Yes
Rapid initial escalation Defining feature Not defining

The important distinction is whether the dose assignment is primarily generated by a fixed toxicity algorithm or by continual updating of a statistical model.

Accelerated Titration vs. BOIN

The Bayesian optimal interval design, or BOIN, is another modern Phase I dose-escalation method.

BOIN uses prespecified probability intervals around a target toxicity probability to guide escalation and de-escalation.

Accelerated titration instead emphasizes rapid early dose movement and intrapatient escalation.

Feature Accelerated Titration BOIN
Primary emphasis Rapid early escalation Efficient target-toxicity allocation
Single-patient early cohorts Yes Design-dependent
Intrapatient escalation Yes Not inherent
Statistical model Optional accompanying model Bayesian decision framework
Target toxicity Can be used in analysis Central to design

Common Mistakes

  1. Thinking accelerated titration simply means doubling every dose. Only particular accelerated designs use 100% dose increments, and escalation must stop when prespecified toxicity criteria are met.
  2. Ignoring moderate toxicity. Moderate toxicity is an important signal that the accelerated phase should become more conservative.
  3. Assuming one patient means one observation forever. Intrapatient escalation means that the same patient may receive multiple dose levels over successive courses.
  4. Assuming the MTD is automatically the RP2D. Modern dose selection may incorporate PK, PD, biomarkers, exposure-response, and preliminary activity.
  5. Using an unvalidated toxicity model. A dose-toxicity model should be prespecified, justified, and assessed for adequacy.
  6. Ignoring cumulative toxicity. Later-course toxicity may depend on cumulative exposure and treatment duration.
  7. Confusing accelerated titration with CRM. Accelerated titration is fundamentally a rule-based approach even when a dose-toxicity model accompanies the analysis.
  8. Changing escalation rules after observing toxicity. The operating characteristics depend on the prespecified decision rules.
  9. Using large dose jumps without adequate safety justification. The accelerated phase is only appropriate within a carefully controlled protocol.

Advantages of Accelerated Titration

The major potential advantages include:

  • Fewer patients treated at very low doses.
  • Faster exploration of the dose range.
  • Potentially shorter Phase I trials.
  • More opportunities for intrapatient dose escalation.
  • More information near clinically relevant doses.
  • Potentially more efficient use of patient participation.

The original simulation study found substantial reductions in average patient numbers and the number of patients experiencing only minimal toxicity compared with the conventional control design.

Limitations of Accelerated Titration

Accelerated titration is not universally superior.

Important limitations include:

  • Large early dose increments can increase risk if the starting dose or dose-toxicity assumptions are poor.
  • Intrapatient escalation complicates interpretation of cumulative toxicity.
  • Single-patient cohorts provide limited information at each early dose level.
  • Delayed toxicity can undermine rapid escalation.
  • The design can be operationally complex.
  • Modern targeted therapies may not have a toxicity-defined optimal dose.
  • Model assumptions can affect dose-toxicity interpretation.

Delayed Toxicity Is Especially Important

The accelerated framework works best when toxicity information becomes available quickly enough to inform subsequent dose assignments.

Suppose a severe toxicity occurs only after several treatment cycles.

Then a patient may already have received multiple escalating doses before the toxicity becomes apparent.

This creates a fundamental timing problem:

$$ \text{dose decision} \rightarrow \text{exposure} \rightarrow \text{delayed toxicity} \rightarrow \text{information arrives too late} $$

Therefore, delayed and cumulative toxicity should be incorporated into the design whenever clinically relevant.

Protocol Considerations

An accelerated titration protocol should clearly specify:

  • Starting dose
  • Candidate dose levels
  • Initial escalation increment
  • Number of patients per early cohort
  • Definition of DLT
  • Definition of moderate toxicity
  • Observation window
  • Rules for intrapatient escalation
  • Criteria for ending the accelerated phase
  • Rules after a DLT
  • Rules after moderate toxicity
  • Conventional escalation strategy after transition
  • Rules for de-escalation
  • Rules for cohort expansion
  • Rules for replacing unevaluable patients
  • Criteria for declaring the recommended dose

Intrapatient Escalation Requires Careful Tracking

When patients receive multiple dose levels, the database must retain the complete treatment history.

For patient \(i\), the dose history may be:

$$ D_{i1},D_{i2},\ldots,D_{iK_i} $$

and the corresponding toxicity history is:

$$ T_{i1},T_{i2},\ldots,T_{iK_i} $$

This allows investigators to evaluate both current-dose and cumulative exposure effects.

Handling Unevaluable Patients

As with any Phase I design, patients may become unevaluable because of:

  • Early withdrawal
  • Disease progression
  • Non-treatment-related death
  • Protocol violations
  • Missing toxicity assessments
  • Inadequate treatment exposure

The protocol must specify how such patients affect dose-escalation decisions.

Do not improvise replacements. If unevaluable patients are replaced, the replacement rules should be prespecified and incorporated into the simulation or operating-characteristic assessment.

Safety Monitoring Is Separate from Dose Escalation

The dose-escalation algorithm does not replace overall safety monitoring.

A Phase I program may require separate monitoring of:

  • Serious adverse events
  • Deaths
  • Laboratory abnormalities
  • Organ-specific toxicities
  • Infusion reactions
  • Cardiac events
  • Neurologic events
  • Other protocol-defined safety signals

The escalation algorithm should operate within the broader safety governance structure of the trial.

Simulation-Based Design Evaluation

Before implementation, candidate accelerated titration designs should be evaluated under multiple plausible dose-toxicity scenarios.

For example, investigators might consider:

  • Low toxicity across the entire dose range.
  • A gradual increase in toxicity with dose.
  • A steep toxicity curve.
  • Large interpatient variability.
  • Large cumulative toxicity.
  • Delayed toxicity.
  • A plateau in toxicity.

For each scenario, thousands of simulated trials can be generated.

The design can then be evaluated according to:

$$ E(N) $$

the expected number of patients, and:

$$ P(\text{select dose near target}) $$

the probability of selecting an appropriate dose.

A Generic Simulation Algorithm

1
Specify the true dose-toxicity relationship.
2
Generate the toxicity outcome for the first patient.
3
Apply the accelerated escalation rule.
4
Generate the next patient's toxicity outcome.
5
Continue until the accelerated-phase stopping criterion is reached.
6
Transition to conventional escalation.
7
Continue until the dose-selection stopping rule is reached.
8
Record the selected dose and patient exposure history.
9
Repeat the simulation many times.
10
Summarize the operating characteristics.

Simulation Outputs

A useful simulation report might include:

Operating Characteristic Question
Mean number of patients How large is the typical trial?
Mean number of dose levels How much of the dose range is explored?
Probability of selecting target dose How often is the correct dose selected?
Probability of overdosing How often are excessively toxic doses assigned?
Patients at low doses How much under-treatment occurs?
Probability of DLT How frequently does serious toxicity occur?
Trial duration How quickly can the study complete?

Why the Original Simulations Were Important

The original accelerated titration paper did not simply propose rapid dose escalation.

The investigators compared alternative designs through stochastic simulation using parameters estimated from actual Phase I trials.

The objective was to evaluate not only patient numbers but also the number of patients treated at low doses, severe toxicity, trial duration, and the amount of information obtained.

This is an important lesson in Phase I design: a dose-escalation method should be evaluated by simulation before being treated as an operational clinical algorithm.

Interpreting the Original Results

In the original simulation study, the conventional control design required an average of approximately 39.9 patients.

The corresponding averages for the three accelerated designs were approximately:

$$ 24.4,\quad20.7,\quad21.2 $$

respectively.

The study also found fewer patients experiencing only Grade 0–1 toxicity as their worst toxicity under the accelerated designs, reflecting the reduced time spent at very low doses.

Important interpretation: These numbers are simulation results from the original study, not universal expectations for every accelerated titration trial.

A Practical Design Workflow

1
Define the clinical objective of the Phase I study.
2
Define the starting dose using the appropriate preclinical and clinical evidence.
3
Specify candidate dose levels.
4
Define DLT criteria.
5
Define moderate-toxicity criteria.
6
Specify the accelerated-phase cohort size and dose increments.
7
Specify intrapatient dose-escalation rules.
8
Specify the criteria for ending accelerated escalation.
9
Specify the conventional escalation strategy after transition.
10
Define the dose-selection objective.
11
Specify the dose-toxicity model, if applicable.
12
Simulate the complete design under plausible toxicity scenarios.
13
Evaluate safety, efficiency, dose selection, and trial duration.
14
Prespecify the final algorithm in the protocol and statistical analysis plan.

What Should Be Reported in a Statistical Analysis Plan?

For an accelerated titration study, the statistical documentation should make the dose-assignment process reproducible.

At minimum, specify:

  • Starting dose
  • Dose levels
  • Accelerated dose increment
  • Number of patients per accelerated dose level
  • Definition of DLT
  • Definition of moderate toxicity
  • DLT observation window
  • Rules for intrapatient escalation
  • Rules for ending acceleration
  • Rules for conventional escalation
  • Rules for de-escalation
  • Rules for cohort expansion
  • Rules for unevaluable patients
  • Dose-selection criterion
  • Statistical model, if applicable
  • Simulation assumptions
  • Simulation operating characteristics

Accelerated Titration and Modern Dose Selection

The original accelerated titration framework was developed primarily in the context of oncology drug development where the MTD was an important target.

Modern oncology development can involve substantially more information.

For example:

$$ \text{Dose selection} = f( \text{toxicity}, \text{PK}, \text{PD}, \text{target engagement}, \text{biomarker}, \text{activity} ) $$

This means that a contemporary Phase I study may identify several candidate doses rather than one obvious MTD.

For example:

Dose DLT Risk PK Exposure PD Activity Potential Interpretation
80 mg Low Low Weak Possibly underexposed
160 mg Moderate Therapeutic Strong Potential RP2D candidate
320 mg High Higher Similar Limited incremental benefit

In such a situation, simply selecting the highest tolerated dose may not be the best development decision.

Relationship to PK/PD Modeling

Accelerated titration is especially interesting when combined with PK/PD modeling.

Suppose exposure follows:

$$ AUC(D)=\theta D^\gamma $$

and pharmacodynamic response follows an Emax relationship:

$$ E(D) = E_0+ \frac{E_{\max}D} {EC_{50}+D} $$

while toxicity follows a separate dose-response relationship:

$$ P(\text{DLT}\mid D) = \frac{ \exp(\alpha+\beta\log D) }{ 1+\exp(\alpha+\beta\log D) } $$

The investigator can then examine the region where pharmacologic activity is high while unacceptable toxicity remains low.

$$ \text{Useful dose region} = \left\{ D: \text{adequate PD activity} \land \text{acceptable toxicity} \right\} $$

This illustrates how accelerated titration can form part of a larger model-informed drug development strategy.

Common Misinterpretation: "Faster Is Better"

The purpose of accelerated titration is not simply to make dose escalation faster.

The actual objective is to improve the information gained per patient while maintaining acceptable safety.

A useful conceptual objective is:

$$ \max \frac{ \text{useful dose-toxicity information} }{ \text{patients exposed} } $$

subject to clinical safety constraints.

This is fundamentally different from maximizing the number of dose levels explored.

When Accelerated Titration May Be Attractive

An accelerated titration strategy may be attractive when:

  • The starting dose is expected to be substantially below the biologically active range.
  • The therapeutic index permits meaningful dose exploration.
  • Toxicity can be assessed promptly.
  • Intrapatient escalation is clinically appropriate.
  • Rapid dose exploration is scientifically valuable.
  • The dose-toxicity relationship is expected to be reasonably monotonic.
  • A substantial number of patients would otherwise be treated at very low doses.

When Another Design May Be Preferable

Accelerated titration may be less attractive when:

  • Toxicities are highly delayed.
  • Cumulative toxicity is substantial and difficult to assess quickly.
  • The starting dose is already close to the expected therapeutic range.
  • The therapeutic window is extremely narrow.
  • Intrapatient dose escalation is not clinically appropriate.
  • The primary objective requires a highly model-based dose-selection framework.
  • The drug's pharmacology makes large dose increments inappropriate.

Accelerated Titration Does Not Eliminate Uncertainty

The early phase of a first-in-human study contains substantial uncertainty.

The true dose-toxicity curve is unknown.

The trial therefore learns about the curve sequentially.

$$ \text{Prior knowledge} \rightarrow \text{patient 1} \rightarrow \text{patient 2} \rightarrow \cdots \rightarrow \text{dose-toxicity information} $$

Accelerated titration changes the speed at which the design moves through the dose range; it does not eliminate the underlying uncertainty.

The Most Important Safety Principle

The most important principle is: the accelerated phase must have a clearly defined exit.

The trial should not continue rapidly escalating simply because the first few patients have tolerated their doses.

The protocol should specify exactly what toxicity signal causes the trial to transition to the more conservative phase.

Bottom line: The value of accelerated titration comes from combining rapid escalation with prespecified safety triggers. The design is not "aggressive dosing without rules"; it is rapid escalation within a carefully defined statistical and clinical framework.

Worked Example Summary

The hypothetical example can be summarized as follows.

Component Illustrative Value
Study type Phase I oncology dose escalation
Starting dose 10 mg
Accelerated cohort size 1 patient
Accelerated dose increment 100%
Dose sequence 10 → 20 → 40 → 80 → 160 mg
Acceleration stopping signal Prespecified moderate toxicity or DLT
Intrapatient escalation Permitted when protocol toxicity criteria are satisfied
Post-acceleration cohort size Illustratively 3 patients
Post-acceleration increment Illustratively 40%
Dose-toxicity model Illustrative logistic model
Target DLT probability Illustratively 25%

The Complete Concept in One Diagram

A
Start low: begin at the prespecified first-in-human dose.
B
Escalate rapidly: use one-patient cohorts and large dose increments while toxicity remains low.
C
Escalate within patients: permit patients with acceptable toxicity to receive higher doses in later courses when allowed.
D
Detect toxicity: monitor for moderate toxicity and DLT.
E
Slow down: end the accelerated phase once the prespecified toxicity trigger is reached.
F
Refine the dose: use conventional escalation, de-escalation, and cohort expansion to characterize the relevant dose range.
G
Select the dose: integrate toxicity with PK, PD, biomarkers, exposure, and other relevant clinical evidence.

The Most Important Concept

The most important conceptual point is that an accelerated titration design is not simply a more aggressive version of a conventional dose-escalation trial.

It is a structured strategy for moving quickly through doses that appear unlikely to produce clinically important toxicity while deliberately slowing the escalation process as evidence of toxicity emerges.

The defining components are:

  • Rapid early dose escalation.
  • One-patient cohorts during the accelerated phase.
  • Intrapatient dose escalation when appropriate.
  • Explicit toxicity triggers.
  • Transition to conventional escalation.
  • Potential use of dose-toxicity modeling.
  • Simulation-based evaluation of safety and efficiency.

The design therefore attempts to solve a fundamental Phase I problem: how can investigators obtain meaningful information about the clinically relevant dose range without exposing large numbers of patients to doses that are unlikely to provide meaningful benefit?

Bottom line: Accelerated titration designs are Phase I dose-escalation strategies designed to reduce the number of patients treated at very low doses and accelerate movement toward clinically relevant exposure. The original framework combines rapid initial escalation, single-patient cohorts, intrapatient dose escalation, toxicity-based transition rules, and subsequent conventional escalation. Because the true dose-toxicity relationship is unknown, candidate designs should be evaluated through simulation, and modern dose selection may integrate toxicity with PK, PD, biomarkers, exposure-response, and preliminary activity rather than relying exclusively on the MTD.

References

Simon, R., Freidlin, B., Rubinstein, L., Arbuck, S.G., Collins, J. & Christian, M.C. (1997). Accelerated titration designs for Phase I clinical trials in oncology. Journal of the National Cancer Institute, 89(15), 1138–1147.
Le Tourneau, C., Lee, J.J. & Siu, L.L. (2009). Accelerated titration designs for Phase I clinical trials. Clinical Cancer Research.
NCI Investigational Drug Steering Committee Clinical Trial Design Task Force. Approaches to Phase 1 Clinical Trial Design Focused on Safety, Efficiency, and Selected Patient Populations. Clinical Cancer Research.
Ji, Y. & Wang, S.J. (2013). Modified toxicity probability interval design: a safer and more reliable method than the 3+3 design for Phase I dose-finding studies. Clinical Trials.
O'Quigley, J., Pepe, M. & Fisher, L. (1990). Continual reassessment method: a practical design for Phase 1 clinical trials in cancer. Biometrics, 46, 33–48.
National Cancer Institute. Accelerated Titration Design Software. NCI Center for Cancer Research.

Related Tutorials

Related Tools

  • → All Statistical Tools
  • → Accelerated Titration Calculator (coming soon)
  • → Dose-Toxicity Model Calculator (coming soon)
  • → Phase I Dose-Escalation Simulator (coming soon)
← Back to All Tutorials