Penel and Kramar BMC Medical Research Methodology 2012, 12:103

http://www.biomedcentral.com/1471-2288/12/103

RESEARCH ARTICLE Open Access

What does a modified-Fibonacci dose-escalation

actually correspond to?
Nicolas Penel1,2* and Andrew Kramar1,2

Abstract
Background: In most phase I oncology trials, it is often stated that the dose increments follow a “modified-
Fibonacci sequence”. This term, however, is vague.
Methods: To better characterize this sequence, we reviewed 81 phase I trials based on this concept.
Results: Out of 198 phase I oncology trials, 81 (41%) are based on modified-Fibonacci series. Actual incremental
ratios varied in a large range from 0.80 to 2.08. The median of actual increments was about 2.00, 1.50, 1.33, 1.33,
1.33, 1.33, 1.30, 1.35. . .. The “modified Fibonacci-sequence” gathers heterogeneous variation of the genuine
sequence, which does not tend to a constant number at higher dose-levels.
Conclusion: This confusing term should be avoided.
Keywords: Dose escalation, Dose increments, Fibonacci sequence, Phase I trials, Methodology

Background methods such as the different versions of the continual

Dose-finding phase I trials seek to determine an optimal reassessment method [1-3]. The guidance for determin-
recommended dose for a new compound for further ing which dose levels are to be explored is usually based
testing in phase II trials [1,2]. For cytotoxic drugs, this on pre-clinical and pharmacokinetic data. Once a start-
dose corresponds to the highest dose associated with an ing dose level has been determined, a Fibonacci se-
acceptable level of toxicity, referred to as the maximal quence, or its modified version, is one of the most
tolerated dose (MTD). Such trials consist in a design frequently used methods for determining dose
where successive cohorts of patients are treated with in- increments.
creasing doses of the drug. The phase II recommended The genuine and the modified Fibonacci sequence de-
dose is usually taken as the dose level just below the termine dose steps (increments). The genuine Fibonacci
MTD [1,2]. Endpoints other than the MTD have been sequence is defined by the linear recurrence equation
considered for molecular targeted therapies such as opti- Fn ¼ Fn1 þ Fn2 , which goes like this: 1, 2, 3, 5, 8, 13,
mal biological dose (OBD). Nevertheless, two different 21, 34, 55, 89. . . In other words, the next number in the
notions are largely confounded: guidance for determin- sequence is equal to the sum of its two predecessors.
ing the actual dose levels to be explored in the trial and The ratio of successive Fibonacci numbers Fn/Fn-1 (in-
guidance for the dose-escalation strategy between dose cremental ratio) tends rapidly to a constant ((1 + √5)/
levels, that is, the determination of the number of 2)  1.61) as n approaches infinity. For the purpose of a
patients enrolled at each dose level. Basically there are dose-seeking phase I trial, this sequence offers very
two approaches for guiding the number of patients en- appealing guidance for the increment set up; because
rolled at each dose level: algorithm-based methods such the absolute doses and the absolute dose change grow
as the classical 3 + 3 and more recent model-based larger and larger with constant incremental ratios
(1.61; +61%) [1,4,5]. The genuine Fibonacci sequence
* Correspondence: [email protected]
in itself is not commonly used as guidance for establish-
1
Methodology and Biostatistics Unit, Centre Oscar Lambret, 3 rue Frederic ing dose increments in the trial. Phase I methodologists
Combemale, 59020, Lille cedex, France
2
refer commonly to the “so-called modified Fibonacci se-
Research Unit EA2694, Medical School, Lille-Nord-de-France University, 1
place de Verdun, 59000, Lille, France
quence” [5], which has the followng incremental ratios:

© 2012 Penel and Kramar; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the
Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
Penel and Kramar BMC Medical Research Methodology 2012, 12:103 Page 2 of 5
http://www.biomedcentral.com/1471-2288/12/103

2.00, 1.67, 1.50, and then 1.33 for all subsequent dose- standard derivation, the median and the range. We then
levels. In this form, the incremental ratio tends to a calculated the ratio of the median increments. For ex-
smaller constant number: 1.33. ample, the ratio of median increment at dose-level 3 was
Given that the so-called modified Fibonacci sequence the ratio of the median dose at level 3 over the median
remains the most popular dose increment determination dose at level 2 across the studies which included at least
scheme, but is also a non-standardized method, we car- these both dose-levels.
ried out a literature review over a ten-year period of Lastly we plotted the mean incremental ratio (±95%
published dose-finding phase I trials investigating one confidence interval) on the log scale for the trial data
drug to better characterize this series. compared to the incremental ratios of the genuine and
modified Fibonacci series.
Methods
The reports, considered for the present study, were Results
extracted from a large database of 327 phase I trials (311 Figure 1 describes the pattern of studies. The modified-
manuscripts) published between January 1997 and De- Fibonacci-sequence was the most common method of
cember 2008 in five major journals in the field of oncol- dose-escalation (92/197, 46%). The contemporary studies
ogy(Annals of Oncology, Clinical Cancer Research, still rarely used sophisticated designs such as Bayesian
European Journal of Cancer, Investigational New Drugs and pharmacologically guided dose escalation (Table 1).
and Journal of Clinical Oncology; see Additional file 1). The median number of dose-levels explored was 5
We have used Pubmed-Medline and the following key- (range, 2–12). Four studies actually investigated only 2
words: “dose-seeking”, “phase 1 trial” and “cancer”. From dose-levels (Table 2).
these 327 trials, 197 were single-agent dose-seeking The term modified-Fibonacci-sequence covered a vast
trials. From each report, we extracted the method used range of actual doses (Table 2). For example, the actual
for guiding the dose-escalation and the result of the trial. increment at the fifth dose-level varied from 1.20 to
“Failure of the trial” described a study failing to deter- 6.00. For every dose-level the observed dispersion of the
mine a phase-II-recommended-dose. The dose- actual increment was large (Table 2).
increment guidance was classified into five categories: The actual sequences used in the trials differed from
modified Fibonacci sequence, pharmacologically guided the genuine Fibonacci sequence and the modified one by
method [6], constant ratio incremental >+2.0 (aggressive several points. The median actual dose was far from the
increment, including most of accelerated titration genuine Fibonacci sequence, with values lower than
designs [7]), constant ratio incremental <1.33 (prudent expected starting from the fifth dose-level (5.32 instead
increment) and other methods (including bayesian of 8) and then huge differences in later dose-levels
approaches [8]). Among studies which used a modified (38.44 instead of 233 at the 12th dose-level, see Figure 2
Fibonacci sequence, we extracted the precise dose used and Table 2). The median of actual increments was
at each dose-level. By convention, we have considered about 2.00, 1.50, 1.33, 1.33, 1.33, 1.33, 1.30, 1.35, 1.32,
the actual dose by successive cohorts even if de- 1.24 and 1.25 (Table 3). This sequence did not tend to a
escalation occurred or the sponsor decided to explore an constant number (see 4th column of Table 3).
intermediate dose-level. De-escalation have been seen in The mean incremental ratios followed the modified Fi-
only two trials. We then calculated for every dose-level: bonacci series more closely than the genuine series and
the mean dose administered at each dose level, its they were statistically lower than the genuine series for

Figure 1 Pattern of 327 dose-seeking phase I trials.

Penel and Kramar BMC Medical Research Methodology 2012, 12:103 Page 3 of 5
http://www.biomedcentral.com/1471-2288/12/103

Table 1 Dose-escalation-methods and results of 198

single-agent dose-seeking phase I trials
Guidance for the number of
patients enrolled by dose-levels
Algorithm-based Model-based
methods (such as methods (such
3 + 3) n (%) as CRM) n (%)
Guidance for Modified-Fibonacci 92 (46.4) 0 (0)
the dose- sequence
escalation
Fixed increment: 28 (14.1) 0 (0)
<+33%
Fixed increment: 62 (1.3) 1 (0)
>100%
PK-guided 3 (1.5) 1 (0)
Other 7 (3.5) 4 (2)
Figure 2 Comparison between actual sequence, genuine
Fibonacci sequence and modified Fibonacci sequence.

dose levels 2 to 7 and significantly lower than the modi-

fied series for dose levels 2 to 5 (Figure 2). From dose dose-escalation is slower than planned by the genuine
levels 8 onwards, the mean incremental ratios were con- and modified Fibonacci sequence (Figure 2).
tained within the two series. The algorithm-based designs tend to be more complex
Moreover, in 13 studies (16%), at least one increment than the model-based-designs. Most of these new
was significantly higher (e.g. >50%) than the increment designs do not use the Fibonacci series. For example,
planned with the modified-Fibonacci series. Simon et al have proposed different designs called accel-
erated titration design, in which the dose doubling is
used until there is one dose-limiting toxicity or two
Discussion patients with at least two grade II toxicities at least pos-
Our analysis highlights the following facts. Most of con- sibly related to drug when a 40% increment is used [7].
temporary dose-seeking phase I trials remain guided by The Bayesian designs do not use the Fibonacci series
the “modified Fibonacci sequence” (about 50%, Table 1). [8,9]. For example, the Bayesian logistic regression
This term includes a vast variation in doses (from -20% model with escalation with overdose control (EWOC)
to +208.3%, Table 2). Moreover, the actual series does uses a model-based approach to determine the potential
not tend to a constant incremental ratio as expected unsafe doses. The maximum dose increment may be
from the modified Fibonacci sequence (Table 2) The capped at a doubling but a smaller dose increment can

Table 2 Dose and incremental ratios in 81 modified-Fibonacci-sequence designed studies

Planned dose in Actual doses
Dose-level (n of studies Genuine Fibonacci Modified Fibonacci Median of Min. of actual Max. of actual Mean (± SD)
reaching the dose-level) sequence sequence actual dose dose dose of actual dose
1(81) 1 1.0 1 - - -
2(81) 2 2.0 2.00 0.80 10 1.82 (0.65)
3 (77) 3 3.3 3.00 0.80 3.33 4.08 (1.65)
4 (67) 5 5.1 4.00 0.97 10.00 9.23 (3.12)
5 (58) 8 6.60 5.32 1.20 33.33 17.61 (4.71)
6 (43) 13 8.80 7.80 0.90 56.25 34.49 (6.89)
7 (32) 21 11.8 11.93 1.86 80.00 67.96 (9.77)
8 (19) 34 15.6 15.78 6.67 128.00 35.64 (5.43)
9 (13) 55 20.8 19.50 13.33 256.00 78.11 (8.69)
10 (7) 89 27.8 23.00 18.00 512.00 133.38 (12.69)
11 (5) 144 36.8 28.89 23.50 1024.00 227.16 (17.85)
12 (3) 233 49.0 38.44 35.83 2048.00 707.43 (29.89)
13 (0) 377 65.2 - - - -
n: number of studied reaching the specified dose-level; SD: standard deviation.
Penel and Kramar BMC Medical Research Methodology 2012, 12:103 Page 4 of 5
http://www.biomedcentral.com/1471-2288/12/103

Table 3 Incremental ratios approaches or PK-guided dose-escalation. However, this

Dose-levels Incremental Incremental ratio Incremental method is very rudimentary. Several analyses have demon-
ratio in genuine in Modified ratio of median strated that Bayesian approaches, such as Continual Re-
Fibonacci Fibonacci dose in 81
sequence sequence phase I assessment Method or EWOC reduce the number of
patients treated at very low doses far from doses with po-
1 - - -
tential therapeutic effect [2,9,10]. Some authors suggest
2 2.000 2.000 2.000
than Fibonacci series is preferred when the dose-toxicity
3 1.500 1.670 1.500 curve in animal toxicology is steep [10].
4 1.667 1.500 1.333 The main limitation of the present analysis is inherent
5 1.600 1.333 1.333 to its retrospective nature. This includes a publication
6 1.615 1.333 1.333 bias that implies a selection of manuscripts issued in
journals with higher impact factors and consequently
7 1.619 1.333 1.333
the trials with “positive” results.
8 1.617 1.333 1.300
9 1.618 1.333 1.350
Conclusion
10 1.618 1.333 1.320
Modified Fibonacci series remains largely used [1,11].
11 1.618 1.333 1.240 The actual dose escalation referring to the “modified Fi-
12 1.618 1.333 1.250 bonacci scheme” is a dose-escalation term close to and
13 1.618 1.333 - intermediate between the genuine and the “modified Fi-
bonacci” schemes (Figure 2). In practice, however, the
modified Fibonacci sequence, the most popular dose-
be used according to the specified overdose control or if escalation guidance scheme in dose-seeking phase I trials
there are other safety concerns. These approaches let the is a term which is too vague and should be avoided. The
emerging data determine the future dose increment sub- precise description of each dose-level in the study report
ject to a cap rather than a fixed algorithm [9]. Based on should be certainly better than the use of this ambiguous
the observed toxicity profiles experienced in the trial, term [5].
the continual reassessment method estimates the dose-
level closest to the target (eg. 33% of patients presenting
Additional file
with dose limiting toxicities) and recommends treating
the next patient at that dose level (without skipping dose
Additional file 1: List of the analyzed trials.
levels). This implies an a priori specification of the drug
toxicity profile characterized by a dose-response rela-
Competing interests
tionship for the drug and requires real-time data- The authors declare that they have no competing interests.
collection, data-management and data-analysis of the
toxicities used for the DLT definitions [8]. This process Authors’ contributions
may slow down the completion of the trial. PK-guided NP collected the data and participated to the statistical analysis. NP and AK
conceived the study and supervised the statistical analysis. All authors read
designs require a validated, sensitive and specific drug
and approved the final manuscript.
assay. PK-guided dose-escalation also implies a relation-
ship between plasma drug concentration and toxicity Authors’ information
profile. This process also implies permanent sample col- NP is medical oncologist and head of the General Oncology Department at
lection, shipping, biological analysis and data-analysis. the Oscar Lambret Cancer Center. His main fields of expertise are treatment
of rare tumors (including, adult sarcoma and cancer of unknown primary
This may also slow down the trial [6]. As previously site) and early phase trials. AK is methologist and biostatistician. His main
reported [1,2,10], about 50% of contemporary phase I fields of expertise are development and validation of prognostic/predictive
trials still use the “Fibonacci series”; in our study, PK- model and early phase trials methodology.

guided dose-escalation or continual reassessement

Acknowledgements
method are used in 3 and 2% of the trials, respectively. To Séverine Marchant for the manuscript editing.
There are many reasons explaining that Fibonacci series To all the authors of the phase I trials analyzed in the present study (see
remains largely used. First, this is one of the first Additional file 1).

described methods. This is easy to understand for the Received: 19 January 2012 Accepted: 23 July 2012
patient, the investigator and the regulatory bodies (such Published: 23 July 2012
as Ethics Committee). The Fibonacci series is apparently
cautious with a very progressive dose-escalation. Lastly, References
1. Rogatko A, Schoeneck D, Jonas W, Tighiouart M, Khuri FR, Porter A:
the implementation of this method does not need exten- Translation of innovative designs into phase I trials. J Clin Oncol 2005,
sive statistical or biological work as compared to Bayesian 25:4982–6.
Penel and Kramar BMC Medical Research Methodology 2012, 12:103 Page 5 of 5
http://www.biomedcentral.com/1471-2288/12/103

2. Ratain MJ, Mick R, Schilsky RL, Siegler M: Statistical and ethical issues in
the design and conduct of phase I and II clinical trials of new anticancer
agents. J Natl Cancer Inst 1993, 85:1637–43.
3. Storer BE: Choosing a phase I design. In Handbook of statistics in Clinical
Oncology. Second Editionth edition. Edited by Crowley J, Hoering A, Ankerst
D. Taylor and Francis Group LLC, 600 Broken Parkway, NW, Suie 300, Boca
Raton, FL: Chapman & Hall/CRC Press; 2005:59. Third edition April 2012.
4. Omura GA: Modified Fibonacci search. J Clin Oncol 2003, 21:3177–79.
5. Omura GA: Phase 1 Dose-finding trials and Fibonacci. Clin Cancer Research
2006, 12:321.
6. Collins JM, Grieshaber CK, Chabner BA: Pharmacologically guided phase I
clinical trials based upon preclinical drug development. J Natl Cancer Inst
1990, 82:1321–26.
7. Simon R, Freidlin B, Rubinstein L, Arbuck SG, Collins J, Christian MC:
Accelerated Titration Designs for phase I clinical trials in oncology. J Natl
Cancer Inst 1997, 89:1138–47.
8. O’Quigley J, Sloan LR: Continual reassessment method: A likelihood
approach. Biometrics 1996, 52:673–84.
9. Babb J, Rogatko A, Zacks S: Cancer Phase I clinical trials: efficient dose
escalation with overdose control. Stat Med 1998, 17:1103–20.
10. Zelen M: Theory and practive of clinical trials. In Holland-Frei Cancer
Medicine. Edited by Kufe DW, Pollock RE, Weischselbaum RR, et al. Hamilton,
Ontario BC: Decker; 2003.
11. Dent SF, Eisenhauer EA: Phase I trial design: are new methodologies
being put into practice? Ann Oncol 1996, 7:561–66.

doi:10.1186/1471-2288-12-103
Cite this article as: Penel and Kramar: What does a modified-Fibonacci
dose-escalation actually correspond to?. BMC Medical Research
Methodology 2012 12:103.

Submit your next manuscript to BioMed Central

and take full advantage of:

• Convenient online submission

• Thorough peer review
• No space constraints or color figure charges
• Immediate publication on acceptance
• Inclusion in PubMed, CAS, Scopus and Google Scholar
• Research which is freely available for redistribution

Submit your manuscript at

www.biomedcentral.com/submit