METHODS
The study included 125 rectal cancer cases who underwent neoadjuvant radiotherapy (RT)±CT between
2010 and 2020, and the cases with a good response (grade 0-1) according to the Modified Ryan classification
were predicted using machine learning. A total of 26 variables were evaluated. After determining
key variables, the dataset was divided into training/test sets at 80%/20%. Logistic regression, artificial
neural network-multilayer perceptron classifier, XGBoost, support vector classification, random forest,
and Gaussian Naive Bayes algorithms used to establish a prediction model. In the prediction of the
group with a good response, 173 cases were created and evaluated with the synthetic minority oversampling
technique method.
RESULTS
Of the 125 cases, 15 had a complete response and 33 had a good response (Modified Ryan grades 0 and
1). Six algorithms were tested in terms of their ability to predict a good response. Key variables for this
prediction were found to be tumor localization, RT break time, age, gender, Karnofsky Performance
Scale score, body mass index, pre- and post-treatment carcinoembryonic antigen levels, pre-treatment
hemoglobin and neutrophil-to-lymphocyte ratio and platelet-to-lymphocyte ratio, radiological T and N
stages, perineural and lymphatic invasion, tumor grade, radiological metastatic lymph node region, RT
dose and technique, and presence and scheme of concurrent CT. The algorithm that showed the best
performance was determined as logistic regression with an accuracy rate of 84% (CI: 0.69-0.98), sensitivity
of 83%, and specificity of 85%.
CONCLUSION
It is very important to predict the cases with a good response and plan treatment accordingly to ensure
effective personalized treatment. Machine learning can successfully predict these cases.
Keywords: Artificial intelligence; machine learning; neoadjuvant chemoradiotherapy; prediction of treatment response; rectal cancer
intelligence (AI) is a branch of computer
science that tries to imitate human-like intelligence
in machines using computer software and algorithms
without direct human stimuli to perform certain tasks.
[
In locally advanced rectal cancer, neoadjuvant
CRT (n-CRT) improves local control, disease-free survival,
and sphincter preservation rates.[
This study aimed to predict response to n-CRT
among 125 cases who underwent this treatment at the
Department of Radiation Oncology of Eskisehir Osmangazi
University Faculty of Medicine between 2010
and 2020.
Treatment Characteristics
Radiotherapy (RT) was applied to all cases as neoadjuvant.
Considering the patient"s KPS score, age, and
comorbidities, an evaluation for concomitant CT was
made and the CT scheme was determined. As concomitant
CT, continuous 5-FU (5-Fluorouracil 225
mg/m2) or capecitabine (825 mg/m2 5 days a week for
5 weeks during RT) was used. During the treatment,
the cases were evaluated at least once a week in the
outpatient clinic based on complete blood count and
blood biochemistry and examination findings. A close
follow-up of toxicity and weight was undertaken.
All the patients were immobilized in the supine position, with their arms up. Computed tomography was performed with the Somatom Definition AS® Device with a 5 mm slice interval. Pelvic MRI fusion was used to contour gross tumor volume (GTVt). In the presence of pathological lymph nodes, MRI and positron emission tomography-computed tomography fusion were used for GTVn. For the clinical target volume (CTV)high, the area of margin was selected as 2 cm for GTVt and 0.5 cm for GTVn, and mesorectum and presacral areas were included. CTVstandard was obtained by adding elective nodal areas to CTVhigh. Regional lymphatics included obturator, presacral, internal iliac, and external iliac lymph nodes (for T4 tumors only). The target volume (PTV) margin planned according to the RT technique was determined as 0.5-1 cm. The bladder, small intestines, and femoral heads were contoured as organs at risk (OARs).
The RT dose was planned as 45-54 (1.8-2 Gy/day). RT was applied with the Varian Trilogy®/TrueBeam®/ Elekta Precise® device accompanied by three-dimensional (3D) conformal RT and volumetric modulated arc therapy (VMAT). 3D conformal RT was applied to 73 cases and VMAT technique was applied to 52 cases.
Evaluation of Treatment Response
ML
MLP is a feed-forward class of ANN. The term MLP
is sometimes loosely used to refer to any feed-forward
ANN and can also refer to specific networks consisting
of more than one sensor layer (with threshold activation).
Multilayer sensors are called "vanilla" NNs in the
spoken language, especially when they have a single hidden
layer.[
XGBoost is an optimized distributed gradient
boosting library designed as a highly efficient, flexible
and portable tool. XGB provides parallel tree enhancement
(also known as GBDT and GBM) that quickly
and accurately solves many data science problems.
The same code runs in a large distributed environment
(e.g., Hadoop, Sun Grid Engine, and Message Passing
Interface) and can solve problems beyond billions of
examples. The most important features of the algorithm
are its high predictive power, its ability to prevent overlearning
and manage empty data, and perform these
tasks in a quick manner.[
SVC involves clustering the data set according to
some criteria to organize the data in a more meaningful
way. There are many ways to achieve this goal. Clustering
can proceed by performing grouping according to a certain
parametric model or based on a measure of distance
or similarity, as in hierarchical clustering. A natural way
of setting cluster boundaries is the use of "valleys" in regions
where there is very little data in the data area; i.e.,
in the probability distribution of the data.[
RFC method has multiple estimation trees and
combines each tree to depend on the randomly selected
vector value equally distributed among all the trees in
the forest. Thus, in RFC, a random ?k vector independent
of the previous random vectors and distributed
across all trees is selected, and each tree is grown using
a training set and a random θk vector, resulting in an
ensemble of trees.[
GNB classifier is one of the top 10 algorithms in
data mining. GNB is a useful classifier widely used in
many applications, such as text categorization, spam
filtering, and data flow classification. Bayesian classifiers
operate based on the Bayes rule and probability
theorems.[
Logistic regression model simply uses a logistic
function to model a binary dependent variable, but it
is fundamentally much more complex. In regression analysis, logistic (or logit) regression estimates the parameters
of a logistic model. This is a method of classifying
the relationship between multiple independent
variables and dependent variables. In logistic regression
analysis, the probability of a dependent variable is
estimated with two values. In addition, the variables in
the model are continuous, which makes this technique
favorable for use in classifying observations.[
A total of 26 variables were evaluated: Age, gender,
KPS score, history of comorbidities, body mass index
(BMI), tumor grade, radiological T, N, and TNM
stages, radiological metastatic lymph node region, tumor
localization (cm), tumor localization (lower/middle/
upper), pre-treatment carcinoembryonic antigen
(CEA) level, post-treatment CEA level, pre-treatment
hemoglobin, neutrophil-to-lymphocyte ratio (NLR)
and platelet-to-lymphocyte ratio (PLR), RT dose (Gy),
RT technique, RT break time, presence and scheme of
concomitant CT, time from RT to surgery, presence of
lymphatic, vascular, and perineural invasion. Key variables
were selected by the permutation-based feature
selection method, which is based on the evaluation of
the significance of each feature separately. This technique
measures changes in prediction quality (based
on the coefficient of decrease in determination score)
after processing in a single feature vector. The rate of
decrease in the coefficient of determination shows how
important a feature is.[
The dataset was divided into training and test sets
at 80% and 20%, respectively. Models were created using
the training set and verified using the test set. The
optimal model was selected according to the receiver
operating characteristic curves. Cross validation is
a model validation technique that tests what result a
statistical analysis will yield in an independent dataset.
The main use of this technique is to predict the accuracy
of a predictive system in practice. In a prediction
problem, the model is usually trained with a "known
dataset" (training set) and tested with an unknown
dataset (verification or test set), which is also known as
supervised learning. The purpose of this test is to measure
the generalizability of the trained model to new
data and identify problems of overfitting or selection
bias.[
The dataset contained 15 cases evaluated to have a
complete response and 33 cases with a good response
(modified Ryan classification grades 0 and 1). In an
unbalanced dataset, the model predicts in favor of the
group with a higher number of samples, which results
in overfitting. In statistics, overfitting refers to a produced analysis aligning too closely to a certain dataset
(memorization), which leads to the inability to adapt
to new data that are not included in this dataset, and
to eliminate the possibility of this problem, a balanced
dataset should be used.[
Statistical analysis was performed and ML algorithms
were run using Python software (Python Software
Foundation. Python Language Reference, version
3.5. Available from http://www.python.org) and Scikit-
Learn library. All analyses and operations were undertaken
on a computer with the specifications of Intel
Core i7-9750 CPU 2.6 GHz 12MB Cache and 16GB
2666 MHz DDR4 RAM memory running 64-bit Windows
10 operating system.
At 4 to 6 weeks after treatment, response to n-CRT was
evaluated with pelvic MRI, surgery was planned 6 to 12
weeks later. The modified Ryan classification was used
for response evaluation after n-CRT.[
In this study, for the evaluation of response to n-CRT,
logistic regression, artificial NN (ANN)-multilayer
perceptron (MLP) classifier, XGBoost, support vector classification (SVC), random forest (RF), and gaussian
naive bayes (GNB) algorithms were used.
Treatment Response
The modified Ryan classification was used for response
prediction, and the number of cases with grades 0, 1, 2,
and 3 was 15 (12.0%), 18 (14.4%), 80 (64.0%), and 12
(9.6%), respectively.
Results of ML
Of the 26 variables, ten were determined as important
using the permutation-based feature selection method:
Tumor localization, RT break time, age, gender, KPS score, BMI, pre- and post-treatment CEA levels, pretreatment
hemoglobin, NLR and PLR values, radiological
T and N stages, perineural and lymphatic invasion,
tumor grade, radiological metastatic lymph node area,
RT dose and technique, and presence and scheme of
concomitant CT. The feature importance graph is given
in Figure
CEA: Carcinoembryonic antigen; NLR: Neutrophil-tolymphocyte
ratio; PLR: Platelet-to-lymphocyte ratio; BMI:
Body mass index; RT: Radiotherapy; VMAT: Volumetric
modulated arc therapy; KPS: Karnofsky performance scale;
r: Radiological; CT: Chemotherapy; ROC: Receiver operating
characteristic; ANN: Artificial neural network; SVC:
Support vector classification; AUC: Area under curve.
With the early detection and prediction of their
response to treatment and following a personalized
treatment approach, patients can be divided into different
prognostic groups. Of the patients with locally
advanced rectal cancer that has undergone n-CRT and
surgery, 45% will require permanent colostomy. Identifying
those with a complete or good clinical response
and good response before surgery will allow for the
optimization of the surgical approach with "organ-sparing"
procedures, resulting in a reduction in surgical
morbidity. In addition, among patients diagnosed
with locally advanced rectal cancer, the early detection
of those with a poor response to n-CRT will offer the
opportunity to directly move on to surgery, thus avoiding
morbidities associated with n-CRT or intensified
treatment regimen.[
In a study conducted with 696 patients with Stage
I-III rectal cancer, Cai et al.[
In a study conducted with 248 patients diagnosed
with locally advanced rectal cancer who underwent
n-CRT, Huang et al.[
Imaging methods have also been used for pCR
prediction.[
In the literature, there are very few studies that predict
good response to n-CRT based on patient, tumor
and treatment characteristics, and such prediction
evaluations have mostly been undertaken using imaging
methods and radiomics. However, response rates
are also related to patient and treatment characteristics.
In the current study, a prediction model was created
using not only patient and treatment but also tumor
characteristics. An accurate classification of cases
with a good response could help determine less invasive
therapeutic strategies, such as sphincter-sparing
surgery, mucosectomy, or the wait-watch approach. In
addition, the prediction of cases that do not respond to
n-CRT would allow for these patients to be referred to
more effective treatments and thus significantly reduce
unnecessary health expenses.
The limitations of the study are the small number
of cases, the inclusion of metastatic (single liver metastasis)
cases and being a single-center study. Prediction
software obtained in such studies has not yet entered
into routine treatment use, and it is not clear which
health authorities can give their ethical approval. The
strengths of the study are the inclusion of patient, tumor,
and treatment characteristics in the algorithm. In
addition, this study is important in terms of forming
the basis for the decisions to be taken about the patient
in the next oncology councils.
Peer-review: Externally peer-reviewed.
Conflict of Interest: All authors declared no conflict of interest.
Ethics Committee Approval: The study was approved by the Eskisehir Osmangazi University Non-Invasive Clinical Research Ethics Committee (No: 25, Date: 30/03/2021).
Financial Support: None declared.
Authorship contributions: Concept - D.E., M.Y., Ö.Ç., B.B.; Design - D.E., M.Y., Ö.Ç., B.B.; Supervision - D.E., M.Y., A.Ö., B.B., E.Y.; Funding - D.E., M.Y., A.Ö., B.B., E.Y.; Materials - D.E., M.Y., B.B., D.K., E.Y.; Data collection and/ or processing - D.E., M.Y., Ö.Ç., D.K., E.Y.; Data analysis and/or interpretation - D.E., M.Y., Ö.Ç., D.K.; Literature search - D.E., M.Y., B.B., D.K.; Writing - D.E., M.Y.; Critical review - D.E., M.Y., Ö.Ç., A.Ö., E.Y.
