WCRJ 2023; 10: e2634

DOI: 10.32113/wcrj_20237_2634

RANDOM SURVIVAL FOREST

IN DETERMINATION OF IMPORTANT
RISK FACTORS ON OVERALL SURVIVAL
AND DISEASE-FREE SURVIVAL
IN GASTRIC CANCER PATIENTS
M. SAFARI1, M. RAHIMI2, J. FARADMAL3, B. GHADERI4,
M.R ANARI5, G. ROSHANAEI3

Department of Biostatistics and Epidemiology, Arak University of Medical Sciences, Arak, Iran
1

2
Department of Biostatistics, School of Public Health, Hamadan University of Medical Sciences, Hamadan, Iran
3
Department of Biostatistics, School of Public Health, Modeling of Noncommunicable Diseases Research center,
Hamadan University of Medical Sciences, Hamadan, Iran
4
Department of Medicine, School of Medicine, Kurdistan University of Medical Sciences, Kurdistan, Iran
5
Department of Statistics, Shahid Beheshti University, Tehran, Iran

Corresponding Author
Ghodratollah Roshanaei, PhD; e-mail: [email protected]

ABSTRACT – Objective: Although the incidence of stomach cancer is decreasing in the world, its incidence
is still high in Iran. Despite different treatments for cancer, disease recurrence, and death may occur in some pa-
tients. Various factors affect survival and recurrence after treatment. This study aims to identify factors affecting
overall survival (OS) and disease-free survival (DFS) in patients with gastric cancer (GC) using a random survival
forest (RSF).
Patients and Methods: In this retrospective study, 553 patients with GC, diagnosed between 2010 and 2018 in
Kurdistan province in the west of Iran, were assessed. Important factors of OS and DFS were identified using the
COX model and RSF. Analysis of data was implemented by R free software version 3.5.3.
Results: The mean (Standard Deviation(SD)) age of patients was 66.99 (13.3) years. The median of OS and DFS
was 18 and 37.5 months, respectively. Using RSF, the important affected factors on OS were tumor grade, stage,
age, recurrence, surgery, and metastasis, respectively. Also according to the RSF model, stage, tumor grade, ra-
diotherapy, tumor site, surgery, and age were the important risk factors for DFS. Based on the prediction error
criterion, the random survival forest performed well in predicting disease-free survival. meanwhile, both RSF and
Cox models had the same performance in predicting overall survival.
Conclusions: Due to the relationship between tumor grade, disease stage and age, the random survival forest
identified these variables as important variables in predicting both outcomes, although the Cox model was not
able to detect these factors, which indicates better performance of RSF.

KEYWORDS: Gastric cancer, Cox model, Random survival forest, Overall survival, Disease free survival.

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License

1
2 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

INTRODUCTION

Gastric cancer (GC) is the third cause of cancer-related death worldwide, which is also higher among
Asians1,2. Gastric carcinoma is a fatal disease with low overall survival in the world, and new cases of this
disease mostly occur in Asian and South American countries1-4. The overall survival of patients is low as
the disease is diagnosed in advanced stages, which is associated with metastasis2,3. However, diagnosis
in the early stages of the disease significantly improves the survival of these patients4-6.
GC was the fifth cause of cancer and the fourth cause of death in the world in 2020 and one of the
most prevalent and deadly cancers on the globe5,7. The incidence of GC is high in Asia. A 5-year survival
rate of 55-66% is reported for this disease, and the main cause of death after curative surgery of GC is
recurrence because most patients experience this outcome. Although the recurrence rate is very low
in the early stages of cancer after curative resection (CR), advanced GC cases show a high rate of recur-
rence after CR8,9.
Recurrence is one of the key factors affecting the survival of GC patients, and post-CR recurrence of
GC usually has destructive effects on survival. Therefore, recurrence patterns should be identified after
the CR of GC by determining recurrence timing to provide information about the postoperative fol-
low-up to find recurrence in time10,11. The identification of factors affecting recurrence and death allows
the classification of patients based on risk prognosis to better manage treatment protocols, which will
improve survival and reduce recurrence rates12.
In many studies on survival, risk factors are identified using the Cox proportional hazards (CPH) mod-
el based on the time until the occurrence of the event13. The CPH model is the widely used semi-para-
metric model for modeling factors affecting survival and recurrence. Nonetheless, the presence of lim-
ited assumptions, such as the proportionality of hazards, the linear relationships between variables with
the hazard, and the limitation of the number of variables in the model, makes this model inefficient in
some applications of survival analysis14. In the presence of the mentioned limitations, the non-para-
metric RSF method is a powerful technique for risk prediction in right-censored data, which can be a
suitable alternative to the semi-parametric CPH model. The main feature of this method is its proper
performance in measuring the importance of each variable in predicting the time to the event. The RSF
is a non-parametric method that considers no specific assumptions and is more efficient than the clas-
sical methods of survival analysis, particularly when there are many predictor variables with collinearity
or the covariates have nonlinear and complex interactions15,16.
Since the identification of patients’ OS and DFS patterns and the affecting factors can help doctors
to determine the appropriate treatment to improve survival, this study aims to identify the important
variables affecting OS and DFS in GC patients applying the RSF model.

PATIENTS AND METHODS

Patients

This retrospective cohort study was conducted on 553 GC patients referring to Tohid Hospital in Kurd-
istan province, during 2012-2018. The collected data, including demographic, clinical, and pathological
variables, specifically age at the diagnosis, gender, tumor grade, tumor site, disease stage, surgery, che-
motherapy, radiotherapy, local recurrence, distant metastasis, the number of chemotherapy courses,
history of smoking, and family history of cancer, were extracted from patients’ records. The patients’
survival status was monitored through periodic visits and telephone calls. The overall survival was de-
fined from diagnosis to death or censoring in months. DFS of patients with surgery was calculated from
the time of surgery to the occurrence of local recurrence or metastasis.

Methods

The CPH model and the RSF model (a non-parametric method) were applied in this study. Three spiliting
rules of log-rank, log-rank score, and random were used in this research17,18.
The models were compared by the Integrated Brier Score (IBS) index in which values ​​close to zero in-
dicate better performance of the model. The efficiency and comparison of the models were examined
using the prediction error index, which ranges between 0 and 1, with a value of zero meaning accurate
prediction or better efficiency19,20.
3 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Statistical Analysis

Analyses were performed using “random-ForestSRC” and “Survival”, a freely available package from
the Comprehensive R software (CRAN) in version 3.5.3. A p-value < 0.05 was defined as statistically
significant.

RESULTS

In this study, 412 (74.5%) out of 553 patients were males, with a mean (SD) age at the diagnosis of 64.1
(13.2) years in the range of 19-94 years. Entirely, 375 (67.8%) patients were dead by the end of the study.
Table 1 represents the demographic and clinical characteristics of the patients. The mean and median
follow-up period of the patients were 28.6 and 18 months, respectively.

Table 1. Demographic and clinical characteristics of gastric cancer patients.

Variable Subgroup N (%) Median OS p-value

   
Sex Female 141 (25.5) 23
0.06
Male 412 (74.5) 18
Age (year) ≤ 55 120 (21.7) 40
56-70 233 (42.1) 20 <0.001
> 70 200 (36.2) 13
Number of 1-5 211(42.4) 16
chemo cycle 6-10 229(46) 18 0.015
>11 58(11.6) 32
Tumor grade Well 70(12.7) 49
Moderate 90(16.3) 16
<0.001
Poor 108(19.5) 10
Unknown 51.5 20
Stage II 55(9.9) 61
III 97(17.5) 17
<0.001
IV 188(34) 12
Missing 213(38.5) 22
Surgery No 360 (65.1) 16
<0.001
Yes 193 (34.9) 27
Radiotherapy No 367(66.4) 17
0.03
Yes 186(33.6) 25
Chemotherapy No 55(9.9) 18
0.76
Yes 498(90.1) 19
Site of Tumor Antrum 120(21.7) 24
Body 57(10.3) 17
Cardia 277(50.1) 17 0.39
Fundus 53(9.6) 20
Unknown 46(8.3) 24
Distance metastasis No 367(68) 60
< 0.001
Yes 177(32) 9
Local recurrence No 471(85.2) -
0.004
Yes 82(14.8) 18
Smoking No 286(51.7) 35
0.85
yes 267(48.3) 34
Family history Yes 67(12.1) 30
0.34
of cancer No 486(87.9) 35
4 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

The mean and median survival periods of all patients were 45.6 and 43 months, respectively, and the
median survival periods in men and women were 43 and 41 months, respectively. OS rates at 1, 3, and
5 years were 86%, 62%, and 31%, respectively. The important factors affecting OS were identified by the
RSF method with all three splitting rules (Table 2). The OS of gastric cancer patients is shown in Figure 1.

Table 2. Evaluation indicators of the Cox and RSF models.

OS

Model Error rate IBS [0,time=71]

Cox 27.2 0.109

RSF (log.rank.score) 27.1 0.110
RSF (random) 29.3 0.115
RSF (log.rank) 27.8 0.113

Figure 1. OS of gastric cancer patients.

Table 2 displays the evaluation indices of the goodness of fit of the model. Based on the IBS index and
the prediction error rate, the log-rank score splitting method was chosen as the appropriate model. Fig-
ure 2 depicts the goodness of fit results of the RSF model with the log-rank score splitting rule to identify
the key OS-affecting variables. As shown in Figure 2, tumor grade, disease stage, age at the diagnosis,
local recurrence, surgery, and distant metastasis are the important OS-affecting variables. The 5-year
survival for the important variables identified using the RSF model is depicted in Figure 3. The 5-year
survival probabilities are adjusted for the other variables. The results show that the predicted 5-year
survival decreases with increasing age, and the probability of 5-year survival decreases with increasing
the disease grade and stage. Also Figure 3 illustrates the estimated 5-year survival probabilities for the
levels of the other important variables identified based on the RSF method.
Figure 4 compares the estimated prediction error for CPH and RSF models with different splitting
rules. The lowest prediction error was obtained for the RSF model with the log-rank score-splitting rule.
The effect of factors on patients’ OS was determined using the multivariate Cox model. The results (Ta-
ble 3) of this model revealed that the age at the diagnosis, local recurrence, disease stage, and surgery
influenced the survival of patients.
5 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Figure 2. Out-of-Bag variable importance and Error rate of RSF for Log-Rank score Splitting Rule.

Figure 3. Partial 5-year predicted survival for six most influential variables on survival in gastric cancer
data. Values on the vertical axis represent the predicted survival probability for a given predictor, after
adjusting for all other predictors.
6 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Figure 4. Prediction Error Curves

for Cox model (Red), RSF with ran-
dom splitting rule (black), RSF with
log-rank score splitting rule (Green),
and RSF with log-rank splitting rule
(blue).

Table 3. Multivariable Cox Regression model of Prognostic Factors on OS.

Variable level HR 95% CI for HR

lower upper

Age_At_Diagnosis 1.03 1.02 1.04

Tumor Grade Well 1
Moderate 1.73 0.88 3.39
Poor 2.28 0.72 4.24
Family_History_of_Cancer No
Yes 1.14 0.81 1.61
Smoking No 1
Yes 0.90 0.72 1.13
Local_Recurrence No 1
Yes 1.84 1.33 2.55
Distance_Metastasis No
Yes 1.41 0.97 2.05
Number_of_Chemotherapy_Course 1-5 1
6-10 0.85 0.59 1.22
>11 0.59 0.37 1.12
Tumor Site Antrum 1
Body 1.17 0.79 1.74
Cardia 0.97 0.73 1.28
Fundus 0.95 0.63 1.44
Previous_History_of_Cancer No 1
Yes 0.78 0.46 1.32
Sex Male 1
Female 0.9 0.7 1.2
Stage II
III 2.21 1.14 4.69
IV 3.86 1.83 8.14
Surgery_Treatment Yes
No 1.52 1.04 2.22
Chemotherapy_Treatment Yes 1
No 0.59 0.37 1.04
Radiotherapy_Treatment Yes 1
No 0.87 0.60 1.28
7 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Disease-free survival (DFS)

In this study, metastasis was observed in 88 (45.6%) out of 193 patients who underwent surgery (distant
metastasis in 38 patients, local recurrence in 34 patients, and both local and distant recurrence in 16
patients). The most common site of metastasis was liver in 59.1% of metastatic cancer patients. Other
sites of metastasis in patients and over survival of them are presented in Table 4. The mean (SD) age
of the operated patients was 61.5 (14.1) years. Table 5 shows the characteristics of operated patients.

Table 4. Metastatic sites of patients with gastric cancer.

Sites of Metastases Frequency Percent OS (month)

   
Liver 52 59.1 15
Lung 9 10.2 17
Bone 4 4.5 23
Intestine 5 5.7 24
Bladder 1 1.1 -
Liver and Intestine 5 5.7 14
Liver and Lung 4 4.6 15
Liver, Lung and Intestine 2 2.3 -
Bladder and Intestine 5 5.7 13
Lung and Intestine 1 1.1 -
Total 88 100 22

Figure 5 shows the probability of DFS in operated GC patients. The mean and median DFS values were 31
and 37.5 months, respectively, and the median DFS values in men and women were 30 and 48 months,
respectively. DFS rates at 1, 3, and 5 years were 74.5%, 45.5%, and 13.5%, respectively. The important
risk factors for DFS were determined using the RSF method. Table 6 shows the goodness of fit indices
of the model. According to the IBS index and the prediction error rate, the RSF model with the log-rank
splitting method was selected as the appropriate model. Figure 6 illustrates the error rate and variable
importance of the RSF model with the log-rank splitting rule to identify the important variables in the
prediction of the DFS. As shown in Figure 6, disease stage, tumor grade, radiotherapy, tumor site, sur-
gery, and age at the diagnosis are the major DFS-affecting variables. Figure 7 displays the prediction
error values for CPH and RSF models with different splitting rules. The 5-year survival rates for the
important variables identified using the RSF model are presented in Figure 8. The adjusted 5-year DFS
probabilities in the presence of other variables reveal that the predicted 5-year survival rates decrease
with increasing age. Also, the increased grade and stage of the disease reduce the 5-year DFS probabil-
ities. The 5-year DFS probabilities for the levels of the other identified important variables are shown
in Figure 8. Table 7 presents the results obtained for the effects of factors on DFS determined using the
CPH model. The results of the CPH model revealed that the DFS was significantly affected by the age at
the diagnosis, chemotherapy, tumor stage, and tumor site. The risk of disease recurrence increases with
increasing the age of the diagnosis and the disease stage. The risk of recurrence decreased in patients
who received chemotherapy, also patients with a tumor site in the upper part of the stomach showed an
increased risk of recurrence. The proportional hazard (PH) assumption for OS wasn’t satisfied (p<0.001)
while the PH assumption for DFS was held.

DISCUSSION

In this study, the important factors affecting OS and DFS in GC patients were determined using the RSF
and COX models. The results indicated that the RSF model performed better than the CPH model in
determining the important variables, and the RSF model is advantageous as it does not require limited
assumptions. The major OS predictors in the RSF method were determined according to tumor grade,
disease stage, age at the diagnosis, local recurrence, surgery, and distant metastasis. Disease stage, tu-
mor grade, radiotherapy, tumor site, surgery, and age at the diagnosis were the main predictors of DFS.
8 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Table 5. Demographic and clinical characteristics of gastric cancer patients.

Variable Subgroup N (%) Median DFS p-value

   
Sex Female 47 (24.4) 30
0.46
Male 146 (75.6) 48
Age ≤ 55 56 (29) 31
56-70 74 (39.4) 48 0.06
> 70 61 (31.6) 29
Number of 1-5 75 (46) 23
chemo course 6-10 64 (39.3) 36 0.26
>11 24 (14.7) 25
Tumor grade Well 36 (18.7) 48
Moderate 33 (17.1) 36
0.001
Poor 28 (14.5) 10
Unknown 96 (49.7) 36
Stage II 26 (13.5) 51
III 33 (17.1) 47
<0.001
IV 61 (31.6) 16
Unknown 73 (37.8) 48
Radiotherapy No 32 (16.6) --
0.08
Yes 161 (83.4) 31
Chemotherapy No 30 (15.5) --
0.02
Yes 163 (84.5) 30
Site of Tumor Antrum 51 (26.4) 47
Body 11 (5.7) 17
Cardia 100 (51.8) 29 0.27
Fundus 14 (7.3) 27
Unknown 17 (8.8) 40
Distance metastasis No 155 (80.3) 51
alone < 0.001
Yes 38 (19.7) 12
Local recurrence No 159 (82.4) 56
alone < 0.001
Yes 34 (17.6) 19
Smoking No 109 (56.5) 36
0.15
Yes 84 (43.5) 31
Family history Yes 26 (13.5) 27
of cancer 0.34
No 167 (86.5) 36

Figure 5. DFS of gastric

cancer patients undergo-
ing surgery
9 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Table 6. Goodness of fit indices of the models.

Model Error rate IBS[0,time=63]

   
Cox 22.7 0.242
RSF(log.rank.score) 21.9 0.219
RSF(random) 22.6 0.242
RSF(log.rank) 21.8 0.212

Figure 6. Out-of-Bag variable importance and Error rate of RSF for Log-Rank Splitting Rule.

Figure 7. Prediction Error

Curves for Cox model (Red),
RSF with log-rank splitting
rule (Black), RSF with ran-
dom splitting rule (Green),
and RSF with log-rank score
splitting rule (Blue).
10 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Figure 8. Partial 5-year predicted survival for six most influential variables on survival in colorectal can-
cer data. Values on the vertical axis represent the predicted survival probability for a given predictor,
after adjusting for all other predictors.

Also in the present study, the result of the CPH model showed that OS was influenced by the variables
of age at the diagnosis, local recurrence, disease stage, and surgery. Furthermore, age at the diagnosis,
tumor site, chemotherapy, and disease stage were among the main predictors of DFS.
In a study on GC patients by Toyokawa et al21, age at the diagnosis and chemotherapy were factors
affecting OS and DFS in patients who were in stage I of the disease, and tumor size and chemotherapy
were factors affecting OS and DFS in the patients at stage II of the disease21.
Yaprak et al22 reported median OS and DFS times of 51 and 35 months, respectively, in GC patients
without metastasis who were in stages 1-3 of the disease. In their study, tumor grade and the disease
stage significantly affected survival rates, and survival probabilities of 85%, 55%, and 45% were respec-
tively obtained for one, three, and five years, and the DFS probabilities for one, three, and five years
were 72%, 49%, and 38%, respectively22.
In a study on GC patients with metastases, Safari et al23 identified the type of surgery, metastasis site,
chemotherapy, age, tumor grade, and surgery, the number of involved lymphomas, gender, and radio-
therapy as the major OS-affecting variables. Adham et al24 introduced age, tumor size, and metastasis as
the key OS-affecting variables based on the RSF model.
In a study on GC patients after CR, Zhu et al25 estimated survival rates of 92.5%, 65.3%, and 46.8% for
one, three, and five years, respectively. In their study, OS was influenced by the variables of age, disease
stage, and tumor site, and DFS was significantly affected by the disease stage25. Itaimi et al26 estimated a
three-year OS rate of 58% in GC patients. In their study, OS was significantly affected by local recurrence
11 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

Table 7. Multivariable Cox Regression of Prognostic Factors on DFS.

Variable level HR 95% CI for HR

lower upper

Age_At_Diagnosis 1.03 1.02 1.04

Grade Well 1
Moderate 0.77 0.32 1.86
Poor 1.02 0.51 2.07
Family_History_of_Cancer No 1
Yes 0.86 0.57 1.31
Smoking No
Yes 0.75 0.55 1.01
Number of Chemo Course 1-5 1
6-10 1.55 0.90 2.66
>11 1.07 0.63 1.80
Tumor_Site Antrum 1
Body 2.25 1.28 3.93
Cardia 1.25 0.84 1.87
fundus 0.76 0.38 1.51
Previous_History_of_Cancer Yes 1
No 1.55 0.76 3.18
Sex female Male 1
Female 0.77 0.53 1.12
Stage II 1
III 1.65 0.63 4.30
IV 6.86 3.13 15.05
Surgery_Treatment Yes 1
No 0.64 0.34 1.22
Chemotherapy_Treatment Yes 1
No 0.40 0.19 0.86
Radiotherapy_Treatment Yes 1
No 1.42 0.74 2.70

and tumor stage, and the number of involved lymph nodes was one of the factors affecting DFS. Han et
al27 conducted a study on men with GC in stages 3-4 of the disease and observed that smoking was the
only factor affecting OS and DFS. In most of the reviewed studies, age at the diagnosis, tumor grade, and
stage, surgery, and chemotherapy were among the variables affecting OS, while DFS was influenced by
age, disease stage, and tumor site22-27. The difference between the previous studies in determining the
influential variables can be attributed to various characteristics of examined patients and variables in
such studies. In particular, the effect of some genes alongside demographic and clinical characteristics
was investigated in some studies. In most studies based on the CPH model, the non-significance of this
variable can be the reason for the collinearity and correlation between the variables. As such, variables
(e.g., the age of diagnosis), which are associated with the variables of disease stage and tumor size, may
not be recognized as significant variables. This seems reasonable in the CPH model, but all three vari-
ables are identified as important in the RSF model regardless of the correlation between the variables.
In the present study, the RSF model performed better than the CPH model for identifying the variables
affecting DFS. However, the performance of CPH and RSF models was almost the same in evaluating the
variables affecting OS, although this has not been confirmed in some studies on survival23-25. In this study,
the RSF model with the log-rank score division rule had the best performance in determining the key
variables affecting OS, and the coordination indices of this model and the CPH model were respectively
obtained at 73.1% and 72.9% in this study. With a coordination index of 70.3%, this model was also select-
ed as an appropriate model by Adham et al24. Ingrisch et al28 obtained coordination indices of 65.7% and
12 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

65.2% for the RSF model with the log-rank division rule and the CPH model, respectively28. In other stud-
ies conducted on other types of diseases, the RSF model can better identify influential variables in these
conditions unlike the CPH model because survival-affecting variables may have collinearity or complex
relationships in such situations. Similarly, the better performance of the RSF model than the CPH model
in survival prediction was confirmed in studies on patients with cardiac arrhythmia by Miao et al29, kidney
transplant patients by Roshanaei et al30, colorectal cancer by Myte et al31, acute liver failure by Zhang et
al32, time to recurrence in ovarian cancer patients by Deldar et al33, and head and neck patients by Datma
et al34. Therefore, the RSF model works at least the same as the CPH model in identifying survival-affecting
variables without limiting assumptions. Thus, the results of this model in identifying important variables
influencing survival can help doctors in diagnostic and preventive assessments.
In the present study, the analysis of three division methods in the RSF model revealed better per-
formance of log-rank score and log-rank division rules, which corresponds to most previous stud-
ies23,29,30,34,35. Regarding the benefits of this study, the main strength might be that GC patients were
monitored in the long-term, and this enabled us to assess affected risk factor on interesting outcomes
precisely. Moreover, the applied method is strongly recommended when the predictor variables are
correlated or there is a nonlinear relationship among the independent variables. Finally, the suggested
method requires no limiting assumptions for the analysis, which is another compelling benefit to use
RSF rather than the conventional methods of survival analysis.
The first strength of the current study is the long-term monitoring of GC patients. Secondly, this
method determines the effect of variables influencing the response prediction in order of importance.
This method also works well if the variables of interest are correlated or there is a nonlinear structure
and even interactions between the variables. Moreover, it requires no limiting assumptions for the anal-
ysis unlike the conventional methods of survival analysis.
As for limitation of the research, it is important to keep in mind that the study was conducted with a
single-center data. It is obvious that the results will be more accurate if more samples in multiple cen-
ters and more auxiliary variables would be available. Lastly, this is a retrospective study in which some
data were not fully recorded for some cases.

CONCLUSIONS

RSF complements the Cox model by providing the relative importance of model covariates, though the
Cox model gives a clinically understandable result on effect of each covariate on survival. Compared
with Cox models, the RSF model can effectively predict the survival of patients with better performance.

Ethics approval:
The study was approved by the Ethics Committee of the Hamadan University of Medical Sciences (Ethics Committee approval
code IR.UMSHA.REC.1397.103; Project number 97030110).

Informed consent:
In this study, informed consent was not necessary because of the use of anonymized patients’ records.

Availability of data and materials:

Data are available on reasonable request from the corresponding author.

Conflict of Interests:
The authors have no conflicts of interest.

Funding:
No specific funding was obtained for this study.

Author contributions:
MS, MR, and GH designed the study. MR and BG collected the data. MS, GR, and JF analyzed and interpreted the data. MS and
MRA drafted the manuscript. GR, JF and MS provided administrative, technical, or material support. All authors contributed
to the article and approved the submitted version.
13 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

ORCID ID:
Malihe Safari: https://orcid.org/0000-0002-8068-5921
Mohamadreza Anari: https://orcid.org/0009-0001-8930-9551
Ghodratollah Roshanaei: https://orcid.org/0000-0002-3547-9125
Mastore Rahimi: https://orcid.org/0000-0002-9742-7293
Bayazid Ghaderi: https://orcid.org/0000-0001-9174-829X
Javad Faradmal: https://orcid.org/0000-0001-5514-3584

REFERENCES
  
   1. Prashanth R, Barsouk A. Epidemiology of gastric cancer: global trends, risk factors and prevention. Przeglad Gastroenterol
2019; 14: 26-38.
   2. Toni G, Panarese I, Di Francia R, Franco R. Molecular Classification of Gastric Cancer. WCRJ 2020; 7: e1472.
   3. Mazidimoradi A, Baghernezhad Hesary F, Gerayllo S, Banakar N, Allahqoli L, Salehiniya H. Global distribution of incidence,
mortality, and burden of stomach cancers and its relationship with the sociodemographic index. WCRJ 2023; 10: e2519.
   4. Yoshida N, Doyama H, Yano T, Horimatsu T, Uedo N, Yamamoto Y, Kakushima N, Kanzaki H, Hori S, Yao K, Oda I, Katada C, Yokoi
C, Ohata K, Yoshimura K, Ishikawa H, Muto M. Early gastric cancer detection in high-risk patients: a multicentre randomised
controlled trial on the effect of second-generation narrow band imaging. Gut 2021; 70: 67-75.
   5. Wu J, Wu XD, Gao Y, Gao Y. Correlation between preoperative systemic immune-inflammatory indexes and the prognosis of
gastric cancer patients. Eur Rev Med Pharmacol Sci 2023; 27: 5706-5720.
   6. Katai H, Ishikawa T, Akazawa K, Isobe Y, Miyashiro I, Oda I, Tsujitani S, Ono H, Tanabe S, Fukagawa T, Nunobe S, Kakeji Y,
Nashimoto A; Registration Committee of the Japanese Gastric Cancer Association. Five-year survival analysis of surgically
resected gastric cancer cases in Japan: a retrospective analysis of more than 100,000 patients from the nationwide registry
of the Japanese Gastric Cancer Association (2001-2007). Gastric Cancer 2018; 21: 144-154.
   7. .7Cuzzuol BR, Vieira ES, Araújo GRL, Apolonio JS, de Carvalho LS, da Silva Junior RT, Bittencourt de Brito B, Freire de Melo F.
Gastric Cancer: A Brief Review, from Risk Factors to Treatment. Arch Gastroenterol Res 2020; 1: 34-39.
   8. .8 Markar SR, Karthikesalingam A, Jackson D, Hanna GB. Long-term survival after gastrectomy for cancer in randomized,
controlled oncological trials: comparison between West and East. Ann Surg Oncol 2013; 20: 2328-2338.
   9. .9Lai JF, Xu WN, Noh SH, Lu WQ. Effect of World Health Organization (WHO) Histological Classification on Predicting Lymph
Node Metastasis and Recurrence in Early Gastric Cancer. Med Sci Monit 2016; 22: 3147-3153.
10. .10 Spolverato G, Ejaz A, Kim Y, Squires MH, Poultsides GA, Fields RC, Schmidt C, Weber SM, Votanopoulos K, Maithel SK,
Pawlik TM. Rates and patterns of recurrence after curative internt resection for gastric cancer: a United States multi-institu-
tional analysis. J Am Coll Surg 2014; 219: 664-75.
11. BY Zhu, SQ Yuan, RC Nie, SM Li, LR Yang, JL Duan, YB Chen, XS Zhang, Prognostic Factors and Recurrence Patterns in T4 Gastric
Cancer Patients after Curative Resection. J Cancer 2019; 10: 1181-1188.
12. Itaimi A, Baraket O, Triki W, Ayed K, Bouchoucha S. Prognostic factors affecting survival and recurrence in gastric carcinoma.
Cancer Rep Rev 2018; 2: 1-4.
13. Kleinbaum D, Klein M. Survival Analysis: A Self-Learning Text. Third ed. New York: Springer; 2012.
14. Ishwaran H, Kogalur UB, Gorodeski EZ, Minn AJ, Lauer MS. High Dimensional Variable Selection for Survival Data. J Am Stat
Ass 2010; 105: 205-17.
15. Breiman L. Random forests. Machine Learning 2001; 45: 5-32.
16. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. The Annals of Applied Statistics 2008; 2: 841–60.
17. Wang H, Shen L, Geng J, Wu Y, Xiao H, Zhang F, Si H. Prognostic value of cancer antigen -125 for lung adenocarcinoma patients
with brain metastasis: A random survival forest prognostic model. Sci Rep 2018; 8: 5670.
18. Nasejje JB, Mwambi H. Application of random survival forests in understanding the determinants of under-five child mortality
in Uganda in the presence of covariates that satisfy the proportional and non-proportional hazards assumption. BMC Res
Notes 2017; 10: 459.
19. Gerds TA, Schumacher M. Consistent Estimation of the Expected Brier Score in General Survival Models with Right-Censored
Event Times. Biometr J 2006; 48: 1029–1040.
20. Pencina MJ, D'Agostino RB. Overall C as a measure of discrimination in survival analysis: model specific population value and
confidence interval estimation. Statis Med 2004; 23: 2109–2123.
21. Toyokawa T, Ohira M, Sakurai K, Kubo N, Tanaka H, Muguruma K, Hirakawa K. The Role of Adjuvant Chemotherapy for Patients
with Stage IB Gastric Cancer. Anticancer Res 2015; 35: 4091-4097.
22. Yaprak G, Tataroglu D, Dogan B, Pekyurek M. Prognostic factors for survival in patients with gastric cancer: Singlecentre
experience. North Clin Istanb 2020; 7: 146–152.
23. Safari M, Abbasi M, Gohari Ensaf F, Berangi Z, Roshanaei G. Identification of Factors Affecting Metastatic Gastric Cancer
Patients’ Survival Using the Random Survival Forest and Comparison with Cox Regression Model. irje 2020; 15: 343-351.
24. Adham D, Abbasgholizadeh N, Abazari M. Prognostic Factors for Survival in Patients with Gastric Cancer using a Random
Survival Forest. Asian Pac J Cancer Prev 2017; 18(1):129-134.
25. Zhu BY, Yuan SQ, Nie RC, Li SM, Yang LR, Duan JL, Chen YB, Zhang XS. Prognostic Factors and Recurrence Patterns in T4 Gastric
Cancer Patients after Curative Resection. J Cancer 2019; 10: 1181-1188.
26. Itaimi A, Baraket O, Triki W, Ayed K, Bouchoucha S, Prognostic factors affecting survival and recurrence in gastric carcinoma
carcinoma. Cancer Rep Rev 2018; 2: 1-4.
27. Han MA, Kim YW, Choi IJ, Oh MG, Kim CG, Lee JY, Cho SJ, Eom BW, Yoon HM, Ryu KW. Association of smoking history with
cancer recurrence and survival in stage III-IV male gastric cancer patients. Cancer Epidemiol Biomarkers Prev 2013; 22:1805-
1812.
14 RANDOM SURVIVAL FOREST IN GASTRIC CANCER PATIENTS

28. Ingrisch M, Schöppe F, Paprottka K, Fabritius M, Strobl FF, De Toni EN, Ilhan H, Todica A, Michl M, Paprottka PM.: "Prediction
of 90Y Radioembolization Outcome from Pretherapeutic Factors with Random Survival Forests". Journal of Nuclear Medicine
2018; 59: 769-773.
29. Miao F, Cai YP, Zhang YX, Li Y, Zhang YT. Risk prediction of one-year mortality in patients with cardiac arrhythmias using
random survival forest. Comput Math Methods Med 2015; 2015: 303250.
30. Roshanaei G, Omidi T, Faradmal J, Safari M, Poorolajal J. Determining affected factors on survival of kidney transplant in living
donor patients using a random survival forest. Koomesh 1397; 20: 517-523.
31. Myte R. Covariate selection for colorectal cancer survival data: A Comparison case study between random survival forests
and the cox proportional-hazards model. Umeå: Umeå University; 2013.
32. Zhang, Zhi-Qiao; He, Gang; Luo, Zhao-Wen; Cheng, Can-Chang; Wang, Peng1; Li, Jing1; Zhu, Ming-Gu; Ming, Lang1; He, Ting-
Shan1; Ouyang, Yan-Ling1; Huang, Yi-Yan1; Wu, Xing-Liu; Ye, Yi-Nong. Individual mortality risk predictive system of patients
with acute-on-chronic liver failure based on a random survival forest model, Chinese Medical Journal 2021; 134: 1701-1708.
33. Deldar M, Anbiaee R, Sayehmiri K. Predicting Epithelial Ovarian Cancer first recurrence with Random Survival Forest: Com-
parison Parametric, Semi-Parametric, and Random Survival Forest Methods. JBE 2021; 6: 267-274.
34. Datema FR, Moya A, Krause P, Bäck T, Willmes L, Langeveld T, Baatenburg de Jong RJ, Blom HM. Novel head and neck cancer
survival analysis approach: random survival forests versus Cox proportional hazards regression. Head Neck 2012; 34: 50-8.
35. Roshanaei G, Safari M, Faradmal J, Abbasi M, Khazaei S. Factors affecting the survival of patients with colorectal cancer using
random survival forest. J Gastrointest Cancer 2022; 53: 64-71.