(662c) A Simulation-Based Optimization Approach to Develop Personalized Colorectal Cancer Screening Strategies | AIChE

(662c) A Simulation-Based Optimization Approach to Develop Personalized Colorectal Cancer Screening Strategies


Young, D. - Presenter, Auburn University
Cremaschi, S., Auburn University

A Simulation-based Optimization
Approach to Develop Personalized Colorectal Cancer Screening Strategies

David Young, Selen

Department of Chemical Engineering, Auburn University, Auburn, AL, USA


cancer (CRC) is the third leading cause of cancer and death among cancers. Colorectal
cancer survival is closely linked to the stage of
cancer at diagnosis, decreasing from 92% for early stage diagnosis to 11% for
advanced stages[1]. Early
detection by screening is an effective way to reduce the risk of CRC. Currently,
there are three main methods for CRC screening, endoscopy, stool testing, or
computed tomography (CT) scans. There are numerous guidelines on CRC screening,
e.g., by American College of Gastroenterology[2]
and by American Cancer Society[3].
However, these strategies were developed using what-if analysis, i.e., they
specify a set of screening strategies, e.g., colonoscopy every 10 years
starting at the age of 50, and compare these strategies in terms of their
cost-effectiveness and health outcomes with each other and with no screening,
e.g., in  [4].
A screening strategy includes: (1) appropriate screening method(s), (2)
frequency of screening for each recommended method, and (3) start and
potentially end of screening window. These attributes define a large number of
potential screening strategies, which cannot be explored efficiently using
what-if analysis. Also, the screening frequency may
need to be adjusted with the changes in the risk factors and age of the
patient, which are not currently considered in the existing analysis.

this paper, we present a simulation-based optimization methodology to generate
the optimum personalized screening strategy for CRC, given defined risk factors.
We embed a microsimulation model of the CRC progression and screening in an
optimization framework to identify the best screening strategy. The
optimization framework employs stochastic programming (SP) to capture the uncertain
nature of CRC development and progression, i.e., CRC natural history can be
analyzed using probabilities but may not be predicted
precisely. The objective of the stochastic programming model is to minimize the
expected total CRC related cost per quality adjusted life years (QALY) gained,
where the cost is defined per person for services related to CRC, and the
quality adjusted life years is a weighted metric based on years lived and CRC
stage of each simulated individual as defined in4.

microsimulation model consists of two sections, a natural history section, and
a screening section. The natural history portion is a probabilistic, continuous
time model to simulate the progression of colorectal cancer. The model is based
on the microsimulation model, CRC-SPIN[5],
and currently only incorporates sex and age as risk factors. The initial stage for
the individual is lesion free, or healthy, followed by the development and
growth of lesion(s), and then from those lesions are pathways to preclinical cancer
and then to clinical cancer. The individual may develop multiple lesions in
their lifetime, and of those lesions, some may never transition into a
cancerous state. Once a lesion passes into a preclinical cancerous state, it is
then stochastically assigned a time to become symptomatic leading to clinically
detected cancer, where mortality due to cancer is then considered. At the beginning
of the model, each individual is assigned an age of death, which accounts for
death due to natural causes or reasons other than CRC. The screening portion of
the model includes colonoscopy, sigmoidoscopy, fecal immunochemical test, with
their respective sensitivities, and specificities. If screening successfully
detects a lesion in a pre-cancerous state, the lesion is
removed halting that lesion’s ability to progress to a cancerous state.
If screening detects a lesion in its cancerous state, the individual is then shifted into a clinically detected cancer state. Our
model does not explicitly incorporate CRC treatment currently; it is assumed the patient receives the appropriate care based
on the CRC stage at diagnosis.

optimization portion of the framework wraps around the microsimulation using
screening strategies as an input to the microsimulation and acquiring expected
total cost per QALY gained as an output from the simulation. The optimization
section then uses the microsimulation output as an objective value attempting
to minimize it using a derivative free optimization algorithm, treating the
microsimulation as a black-box function. There are numerous derivative free
optimization algorithms available for use[6].
We employed genetic algorithm(GA) and particle swarm
optimization(PSO)[7], NOMAD[8],
and DFL[9]
as derivative-free optimization methods for our analysis, and compared their
performances for equal numbers of microsimulation evaluations.

applied our simulation-based optimization approach to identify screening
strategies for the 1975 US male population[10],
a population that had little to no screening, to study the impact on the CRC
related costs. To identify the optimal strategy, we simulated a male cohort of
one million individuals all born at the same point in time. The cohort
simulation was then used as the black box model to
generate the objective function values, i.e. expected total cost per QALY
gained within each of the optimization algorithms. The best screening strategy
identified by each optimization algorithm is then used
as the screening strategy for the male population of 1975 using our
microsimulation, applying the screening strategy to each individual with a 100
% compliance rate. It was found that the best-identified
screening strategy using the genetic algorithm and NOMAD would have resulted in
a 79 % reduction in total cases of CRC, and an 82 % reduction of CRC related
deaths as compared to not screening the male population at all. For PSO, the
strategy would have resulted in a 52 % reduction in CRC cases and a 57 %
reduction in CRC related deaths. For DFL, the strategy would have resulted in a
54 % reduction in CRC cases and a 59 % reduction in CRC related deaths and that
the best screening test to use was a colonoscopy, which is to be expected. All of the algorithms also identified that the
starting age for screening should be around 50 years. Frequency of screening,
however, differed for each of the algorithms. Both DFL and PSO recommended a
single colonoscopy around the age of 50 to be optimal for the average male. The
genetic algorithm identified the best strategy to be a colonoscopy every ten
years from age 50 to age 85, one of the current recommended strategies1,2. The NOMAD algorithm identified that screening with
a colonoscopy every 18 years starting at age 49 and ending at age 83 was the
best strategy. Coupled with the reductions observed in CRC cases and CRC related
deaths from the 1975 male population simulation, our results suggest that for an
average risk male, screening with one colonoscopy around age 50 may not be enough
to drastically reduce the risk of CRC. However, our results reveal that even screening
with a colonoscopy once around the age of 50 is able to lower the overall risk
of CRC.

[1] American Cancer
Society - https://www.cancer.org/cancer/colon-rectal-cancer/about/key-statistics.html, accessed
September 2017

[2] Rex, D. K. et
American College of Gastroenterology guidelines for colorectal cancer
screening 2009 [corrected]. Am. J. Gastroenterol.
104, 739–50 (2009).

[3] Robert A. Smith,
Kimberly Andrews et al. , R. C. W. Cancer Screening in
the United States , 2016 : A Review of Current American Cancer Society
Guidelines and Current Issues in Cancer Screening. A Cancer J. Clin. 66, 95–114

[4] Br J Surg. 2017
May 31. doi:

ADDIN Mendeley
Bibliography CSL_BIBLIOGRAPHY 1.Roberts, S., Wang, L., Klein, R., Ness, R. & Dittus, R.
Development of a simulation model of colorectal cancer. ACM Trans. Model.
Comput. Simul.
18, 1–30 (2007).

Rios, L. M. & Sahinidis, N. V. Derivative-free
optimization: A review of algorithms and comparison of software
implementations. J. Glob. Optim. 56,
1247–1293 (2013).

Fortin, F. De Rainville, M. Gardner, M. Parizeau, and
C. Gagné, 2012, “DEAP: Evolutionary Algorithms Made
Easy”, Journal of Machine Learning Research, 13, pp. 2171-2175.

[8] Audet, C., Custódio, A. L. & Dennis, J. E. Erratum: Mesh Adaptive
Direct Search Algorithms for Constrained Optimization. SIAM J. Optim. 18,
1501–1503 (2008).

[9] G.Liuzzi, S.Lucidi, F.Rinaldi. Derivative-Free Methods for Mixed-Integer
Constrained Optimization Problems, Journal of Optimization Theory and
Applications, 164(3):933-965 (2015). DOI: 10.1007/s10957-014-0617-4

National Cancer Institute. Cancer
Incidence – Surveillance, Epidemiology, an End Results (SEER) Registries
Research Data.
Accessed September 18, 2017.