Original Article

Using machine learning models to predict synchronous

genitourinary cancers among gastrointestinal stromal tumor
Downloaded from [Link] by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AW

patients
nYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE= on 01/07/2025

Mohammad Alghafees, Raouf M Seyam, Turki Al‑Hussain1, Tarek Mahmoud Amin2, Waleed Altaweel,
Belal Nedal Sabbah3, Ahmad Nedal Sabbah3, Razan Almesned, Laila Alessa
Departments of Urology and 1Pathology and Laboratory Medicine, King Faisal Specialist Hospital and Research Center, 2Department of
Surgical Oncology, Oncology Center, King Faisal Specialist Hospital and Research Center, 3College of Medicine, Alfaisal University, Riyadh,
Saudi Arabia

Abstract Objectives: Gastrointestinal stromal tumors (GISTs) can occur synchronously with other neoplasms, including
the genitourinary (GU) system. Machine learning (ML) may be a valuable tool in predicting synchronous GU
tumors in GIST patients, and thus improving prognosis. This study aims to evaluate the use of ML algorithms
to predict synchronous GU tumors among GIST patients in a specialist research center in Saudi Arabia.
Materials and Methods: We analyzed data from all patients with histopathologically confirmed GIST at our
facility from 2003 to 2020. Patient files were reviewed for the presence of renal cell carcinoma, adrenal tumors,
or other GU cancers. Three supervised ML algorithms were used: logistic regression, XGBoost Regressor,
and random forests (RFs). A set of variables, including independent attributes, was entered into the models.
Results: A total of 170 patients were included in the study, with 58.8% (n = 100) being male. The median age
was 57 (range: 9–91) years. The majority of GISTs were gastric (60%, n = 102) with a spindle cell histology.
The most common stage at diagnosis was T2 (27.6%, n = 47) and N0 (20%, n = 34). Six patients (3.5%) had
synchronous GU tumors. The RF model achieved the highest accuracy with 97.1%.
Conclusion: Our study suggests that the RF model is an effective tool for predicting synchronous GU tumors
in GIST patients. Larger multicenter studies, utilizing more powerful algorithms such as deep learning and
other artificial intelligence subsets, are necessary to further refine and improve these predictions.

Keywords: Artificial intelligence, gastrointestinal oncology, gastrointestinal stromal tumors, genitourinary

oncology, urologic oncology, urology

Address for correspondence: Dr. Belal Nedal Sabbah, Takhassusi Street, Alfaisal University, Riyadh 11533, Saudi Arabia.
E‑mail: bsabbah@[Link]
Received: 07.03.2023, Revised: 24.08.2023, Accepted: 13.10.2023, Published: 25.01.2024.

INTRODUCTION frequently occurring mesenchymal tumor in this anatomical

region. [1] GISTs are known to have several distinct
Gastrointestinal stromal tumors (GISTs) are a rare molecular subtypes, including those with mutations in KIT
type of mesenchymal tumor that commonly develop in or PDGFRa. Detecting these molecular alterations at an
the gastrointestinal tract. In fact, GISTs are the most
This is an open access journal, and articles are distributed under the terms of the Creative
Commons Attribution‑NonCommercial‑ShareAlike 4.0 License, which allows others to
Access this article online remix, tweak, and build upon the work non‑commercially, as long as appropriate credit
is given and the new creations are licensed under the identical terms.
Quick Response Code:
Website:
[Link] For reprints contact: WKHLRPMedknow_reprints@[Link]

How to cite this article: Alghafees M, Seyam RM, Al‑Hussain T, Amin TM,
DOI: Altaweel W, Sabbah BN, et al. Using machine learning models to predict
10.4103/ua.ua_32_23 synchronous genitourinary cancers among gastrointestinal stromal tumor
patients. Urol Ann 2024;16:94-7.

94 © 2024 Urology Annals | Published by Wolters Kluwer - Medknow

 Alghafees, et al.: Machine learning in predicting GU involvement among GIST patients

early stage is critical as it can significantly impact the choice States). Continuous data are presented as averages with
of adjuvant and metastatic treatments.[2] Existing literature standard deviations, while categorical data are depicted
suggests that GISTs have a nearly equal distribution using absolute numbers and percentages.
between genders, with a higher incidence among individuals
over the age of 60. Furthermore, GISTs tend to present Four types of AI algorithms were employed in this study to
with symptoms, indicating a symptomatic nature of the predict the presence of GU cancer in the presence of GIST.
These include RF, XGBoost classifier, CatBoost classifier,
Downloaded from [Link] by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AW

disease.[3] Studies conducted in Saudi Arabia have shown

that GISTs are predominantly located in the stomach and and support vector machine. After running a baseline
have a higher incidence in males over the age of 40 years.[4] prediction model, some variables were dropped because
they were not significant to the prediction of the model.
nYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE= on 01/07/2025

Although GISTs primarily occur in the stomach and The ML models were fitted using scikit‑learn 0.18 modules
intestine, some patients may experience lower urinary tract of Python throughout this study. The data set was randomly
symptoms that suggest synchronous genitourinary (GU) divided into 80% of the training set and 20% of the test set
tumors. In addition, extragastrointestinal stromal tumors at 8:2 (136: 34). The target variable was encoded in a binary
of the urinary bladder wall have been observed in rare format with 1 (presence of GU cancer) and 0 (absence of
cases.[5] Currently, an accurate diagnosis of GISTs requires GU cancer). The RF model is a decision tree‑based ML
extensive imaging studies, pathological examination, model. Each node of the decision tree divides the data into
and immunohistochemical analysis.[6] Early diagnosis is two groups using a cutoff value inside one of the features.
imperative to achieve high rates of disease‑free survival, By building an ensemble of randomized decision trees,
yet the extensive testing required for a diagnosis takes each of which overfits the data and averages the results to
substantial time.[7] Therefore, implementing technology that obtain a better classification, the RF technique can reduce
predicts a concomitant tumor of other organs among GIST the effect of the overfitting problem.
patients could significantly impact the overall prognosis of
this condition. This retrospective chart review study involving human
participants followed the standards of the 1964 Helsinki
Recent research suggests that using artificial intelligence (AI) Declaration and its later amendments. This study is a
and deep learning algorithms may provide more accurate secondary analysis of datasets from an already approved
confirmation of the malignant potential of GISTs.[8] The study by the Human Investigation Committee (IRB) and
implementation of machine learning (ML) techniques, Research Ethics Committee of King Faisal Specialist
including supervised learning algorithms, has shown Hospital and Research Center.
promising results in improving the accuracy of predictions
for various medical conditions. RESULTS

In this study, we aim to utilize ML to predict GU cancer in A total of 170 GIST patients were detected. As shown
GIST patients, with a particular focus on the Saudi Arabian in Table 1, most of the patients (58.8%; n = 100) were
population. By utilizing a large dataset of patients males. The median age was 57 (9–91) years. The majority
diagnosed with GIST from our specialist research center of the GISTs were gastric (60%; n = 102) with spindle
between 2003 and 2020, we aim to determine the accuracy cell histology. The most common stage at diagnosis
and effectiveness of three supervised ML algorithms: is T2 (27.6%; n = 47) and N0 (20%; n = 34). Six
logistic regression, XGBoost Regressor, and random patients (3.5%) had synchronous GU tumors. Of them,
forests (RF). The identification of predictive variables and three patients had renal cell carcinomas (RCC). Two were
the accuracy of these models will provide valuable insight histologically diagnosed to have clear cell RCC and one with
into the potential for AI and ML to improve the diagnosis only a radiological diagnosis of RCC. Three other patients
and management of GIST patients, particularly in the had adrenal tumors (one adrenal carcinoma, one isolated
context of having a synchronous GU neoplasm. adrenal GIST, and one pheochromocytoma).

MATERIALS AND METHODS After all modes of hyperparameter tuning were done to the
model, the RF model achieved the highest accuracy with
This retrospective study included all patients with a 97.1%. It predicted that based on the input variables and
histopathological diagnosis of GIST at King Faisal patient characteristics, 97.1% still did not have associated
Specialist Hospital and Research Centre between 2003 and GU cancer and that only 2.9% of those who had GIST
2020. Any concomitant GU cancer was identified. Data had associated GU cancer. On more analysis to ascertain
were analyzed using SPSS v26 (IBM, New York, United the specificity of the model, Figure 1 shows the confusion
Urology Annals | Volume 16 | Issue 1 | January‑March 2024 95
 Alghafees, et al.: Machine learning in predicting GU involvement among GIST patients

matrix for the RF models which explains the specificity index 5, 3, and 6 contributed more in the prediction. These
of the model in terms of how true the predicted values variables were associated with cancer taking the highest,
are accurate to the original values. It showed that out gender, and site of GIST, respectively. Therefore, even
of a random 34 number of patients, the model predicts with the presence of GIST‑associated cancer, there is a
32 patients to be GU cancer‑free even in the presence of rare correlation between GIST and GU cancer.
GIST and only 1 patient to have associated GU cancer in
DISCUSSION
Downloaded from [Link] by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AW

the presence of GIST.

Figure 2 shows the feature importance of each variable The study’s findings demonstrate the potential of AI
column used for the RF model which is the one with the technology to accurately predict having synchronous GU
cancer among GIST patients, as evidenced by the RF
nYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE= on 01/07/2025

best prediction accuracy. It is evident that variables in the

model’s 97.1% accuracy. The patient population analyzed
Table 1: Demographic and tumor‑related characteristics of was mostly male. Only a small portion of patients had an
patients (n=170) accompanying GU cancer, at <5%. The diagnoses for these
Continuous variables n Median (range) patients included RCC, adrenal carcinoma, adrenal GIST,
Age at diagnosis (years) 170 57 (9–91) and pheochromocytoma.
GIST size (cm) 161 6 (0.3–36)
Categorical variables n (%)
Our study’s findings are consistent with existing literature
Gender
Male 100 (58.8) regarding patient demographics and disease characteristics,
Female 70 (41.2) showing that GISTs are predominantly located in the
GIST primary site stomach (61%). The reported age of onset varies across
Gastric 102 (60)
Small intestine 47 (27.6) studies, with median diagnosis age ranging from 50 to
Omentum/peritoneum/mesenteric 12 (7.1) 60 years.[9,10] However, a study conducted in Saudi Arabia
Other 9 (5.3)
reported a lower mean age at diagnosis of 40 years, which
GIST TNM stage
T1 25 (14.7) is substantially lower than the median age reported in other
T2 47 (27.6) studies.[4] Thus, our study’s results indicate that the age
T3 44 (25.9)
T4 45 (26.5)
of onset of GIST in our cohort is higher than what has
N0 34 (20.0) been reported in other studies conducted in Saudi Arabia.
N1 5 (2.9) This difference may be due to various factors, including
M0 13 (7.6)
M1 25 (14.7)
differences in sample sizes, selection criteria, and genetic
Histopathological subtype and environmental factors. However, further studies are
Spindle cell 85 (50.0) required to confirm this observation.
Epithelioid type 16 (9.4)
Mixed epithelioid and spindle 10 (5.9)
Other 2 (1.2) This study represents an initial attempt to utilize ML
GIST: Gastrointestinal stromal tumor, TNM: Tumor, node, and algorithms to predict the presence of GU tumors in
metastasis GIST patients. However, ML models have recently been
the subject of numerous research studies across various
cancer types, including ovarian, thyroid, and breast
cancer.[11‑13] These studies demonstrate the potential of ML

Figure 1: Confusion Matrix Figure 2: Most Effective Parameters

96 Urology Annals | Volume 16 | Issue 1 | January‑March 2024

 Alghafees, et al.: Machine learning in predicting GU involvement among GIST patients

in predicting disease outcomes and identifying biomarkers Financial support and sponsorship
for early diagnosis. Toth et al. demonstrated the successful Nil.
use of the RF model in clinical practice for the detection
of biomarkers for prostate cancer progression. Their Conflicts of interest
study utilized an RF‑based classification model to predict There are no conflicts of interest.
the aggressive behavior of prostate cancer, achieving
REFERENCES
Downloaded from [Link] by BhDMf5ePHKav1zEoum1tQfN4a+kJLhEZgbsIHo4XMi0hCywCX1AW

an accuracy of 95%. The application of the RF model

in their study allowed for the identification of a set of 1. Rubin BP, Heinrich MC, Corless CL. Gastrointestinal stromal tumour.
biomarkers that could predict the likelihood of disease Lancet 2007;369:1731‑41.
2. Kelly CM, Gutierrez Sainz L, Chi P. The management of metastatic
progression and guide clinical decision‑making.[14] The high
nYQp/IlQrHD3i3D0OdRyi7TvSFl4Cf3VC4/OAVpDDa8K2+Ya6H515kE= on 01/07/2025

GIST: Current standard and investigational therapeutics. J Hematol

accuracy of the RF model in predicting prostate cancer Oncol 2021;14:2.
behavior suggests its potential for use in other cancer 3. Søreide K, Sandvik OM, Søreide JA, Giljaca V, Jureckova A, Bulusu VR.
types, including the prediction of GU cancers in GIST Global epidemiology of gastrointestinal stromal tumours (GIST):
A systematic review of population‑based cohort studies. Cancer
patients as demonstrated in our study. These findings Epidemiol 2016;40:39‑46.
support AI as an externally valid classification model to 4. Bokhary RY, Al‑Maghrabi JA. Gastrointestinal stromal tumors in
support the clinical management of prostate cancer.[14] Western Saudi Arabia. Saudi Med J 2010;31:437‑41.
5. Mekni A, Chelly I, Azzouz H, Ben Ghorbel I, Bellil S, Haouet S, et al.
Another study by Xiao et al. reported on similar outcomes Extragastrointestinal stromal tumor of the urinary wall bladder: Case
predicting the occurrence of prostate cancer using the report and review of the literature. Pathologica 2008;100:173‑5.
RF algorithm. Here, transrectal ultrasound findings, age, 6. Yadav SC, Menon S, Bakshi G, Katdare A, Ramadwar M, Desai SB.
Gastrointestinal stromal tumor presenting with lower urinary tract
and serum levels of prostate‑specific antigen were taken
symptoms – A series of five cases with unusual clinical presentation.
into account, yielding a predictive accuracy of 83.10%. Indian J Urol 2021;37:357‑60.
The results of this study permitted the statement that the 7. Uzunoglu H, Tosun Y, Akinci O, Baris B. Gastrointestinal stromal
adoption of an RF model and AI technology demonstrates tumors of the stomach: A 10‑year experience of a single‑center. Niger
J Clin Pract 2021;24:1785‑92.
superior diagnostic performance than individual diagnostic 8. Seven G, Silahtaroglu G, Kochan K, Ince AT, Arici DS, Senturk H.
indicators alone.[15] This is supported by the findings of Use of artificial intelligence in the prediction of malignant potential
the present study. of gastric gastrointestinal stromal tumors. Dig Dis Sci 2022;67:273‑81.
9. Joensuu H, Hohenberger P, Corless CL. Gastrointestinal stromal
tumour. Lancet 2013;382:973‑83.
Limitations 10. Vij M, Agrawal V, Kumar A, Pandey R. Gastrointestinal stromal tumors:
There are several limitations worth noting in this study. A clinicopathological and immunohistochemical study of 121 cases.
First, we did not include all potential predictive factors for Indian J Gastroenterol 2010;29:231‑6.
11. Park S, Yi G. Development of gene expression‑based random
having GU cancer in GIST patients, such as family history
forest model for predicting neoadjuvant chemotherapy response in
of malignancy and exposure to risk factors. Second, this triple‑negative breast cancer. Cancers (Basel) 2022;14:881.
study was conducted at a single center, which may limit 12. Liu YH, Jin J, Liu YJ. Machine learning‑based random forest for
the generalizability of our results to other populations. predicting decreased quality of life in thyroid cancer patients after
thyroidectomy. Support Care Cancer 2022;30:2507‑13.
Third, there are currently no other studies in the literature 13. Cheng L, Li L, Wang L, Li X, Xing H, Zhou J. A random forest classifier
that explore the use of ML to predict synchronous GU predicts recurrence risk in patients with ovarian cancer. Mol Med Rep
tumors and GISTs, which makes it difficult to compare 2018;18:3289‑97.
14. To t h R , S c h i f f m a n n H , H u b e ‑ M a g g C, B ü s c h e c k F,
and validate our findings. Fourth, our model accuracy Höflmayer D, Weidemann S, et al. Random forest‑based modelling to
needs to be tested and validated by an external series in detect biomarkers for prostate cancer progression. Clin Epigenetics
the future. Future research should aim to address these 2019;11:148.
limitations by exploring whether incorporating additional 15. Xiao LH, Chen PR, Gou ZP, Li YZ, Li M, Xiang LC, et al. Prostate
cancer prediction using the random forest algorithm that takes into
predictive factors into the RF model can improve its account transrectal ultrasound findings, age, and serum levels of
accuracy. prostate‑specific antigen. Asian J Androl 2017;19:586‑90.

Urology Annals | Volume 16 | Issue 1 | January‑March 2024 97