Decision tree-based data mining approach for the evaluation of survival in primary malignant bone tumors: A surveillance,epidemiology and end results database study

Abstract

Purpose

This study aimed to conduct a large-scale population-based study to understand the epidemiological characteristics of Primary Malignant Bone Tumors (PMBTs) and determine the prognostic factors by concurrently using the classical statistical method and data mining methods.

Methods

Patients included in this study were extracted from the National Cancer Institute’s Surveillance, Epidemiology and End Results (SEER) database: “Incidence-SEER Research Data, 18 Registries, Nov 2020 Sub”. Patients with unclassified and incomplete information were excluded. This search algorithm resulted in a dataset comprising 6234 cases. Survival analyses were performed with Kaplan-Meier curves and the Log-rank test. Multivariate Cox regression analysis determined the independent prognostic factors of PMBT. A decision tree-based data mining technique was used in this study to confirm the prognostic factors.

Results

5-years survival rate was 63.6% and 10-years survival rate was 55.3% in the patients with PMBT. Sex, age, median household income, histology, primary site, grade, stage, metastasis, and the total number of malignant tumors were determined as independent risk factors associated with overall survival (OS) in the multivariate COX regression analysis. The prognostic factors resulting in five terminal nodes in the decision tree (DT) included stage, age, and grade. The stage was the most important determining factor for vital status. The terminal node with the shortest number of surviving patients included 801 (72.3%) deaths in 1102 patients with distant stage, and hazard ratio was calculated as 5.4 (95% CI: 4.9–5.9; p < .001). These patients had a median survival of only 17 months.

Conclusions

Rules extracted from DTs provide information about risk factors in specific patient groups and can be used by clinicians making decisions on individual patients. We recommend using DTs in combination with COX regression analysis to determine risk factors and the effect of these factors on survival.

Keywords

primary malignant bone tumors SEER database decision tree data mining

Introduction

Metastatic tumors that have spread to bones from other parts of the body are the most common type of bone malignancies. Primary malignant bone tumors (PMBTs) are rare and often aggressive tumors, that constitute less than 1% of all cancers in adults.^1,2 Osteosarcoma, chondrosarcoma, and Ewing’s sarcoma are the most frequently diagnosed PMBTs.³ The management of PMBTs can be complex in terms of diagnosis, follow-up, and treatment. These patients should be managed in tertiary centers with orthopaedics, radiology, pathology, oncology, radiation oncology and nuclear medicine departments. The management of PMBT has recently improved dramatically. Nowadays, most patients undergo limb-sparing procedures, and survival rates have improved significantly.

Data mining techniques, including neural networks, decision trees (DTs), artificial intelligence, machine learning, and genetic algorithms, are used to uncover previously unknown relationships and patterns between variables in large volumes of data, which may not be apparent through classical statistical methods that are primarily used to test specific hypotheses.^4–6 Decision trees create a flow chart-like model, starting from a root node and proceeding through multiple internal nodes until it reaches the leaf nodes, and can handle both discrete and continuous variables, missing values, and skewed continuous data by dividing it into ranges.⁷ Additionally, Decision trees are easy to understand, unambiguous, and can be compared with traditional statistical techniques in medicine.^5,6 Using decision tree-based data mining, we can identify critical features in a dataset that significantly impact PMBT prognosis.

This study aimed to conduct a large-scale population-based study to understand the epidemiological characteristics of PMBTs and determine the prognostic factors by concurrently using the classical statistical method and decision tree-based data mining methods.

Material and methods

National Cancer Institute’s Surveillance, Epidemiology and End Results (SEER) Program [http://seer.cancer.gov/], is a cancer registry covering 47.9% of the USA population.⁸ providing a limited-use data set for further analyses by researchers and the public.⁹ SEER reports patients’ demographic data and cancer incidence as well as chances of survival at a case-by-case level. Histological tumor type and tumor localization are currently classified in version 3 of the ICD code for Oncology (ICD-O-3).^10,11 The SEER research contract was signed online to access the database. All patient identifiers and all personal data in this study were removed from the SEER database. As a result, studies using the SEER database are exempt from institutional review board approval.¹² No ethics committee approval nor patient approval forms are necessary to use data on the SEER platform. SEER*stat software (SEER*Stat Version 8.3.9.2) is used to retrieve data from SEER, the software can be downloaded from http://seer.cancer.gov/seerstat/website.

Selection of the study cohort and data collection

Patients included in this study were extracted from the SEER*Stat Database: Incidence—SEER Research Data, 18 Registries, Nov 2020 Sub (patients diagnosed between 2000 and 2018). To identify patients with PMBT, “Bones and Joints” in the “Site and Morphology: Site recode ICD-O-3/WHO 2008” category was selected (15,327 cases). To reach patients with PMBT, “Malignant Bone Tumors” in the “Site and Morphology: ICCC (International Classification of Childhood Cancer) site recode extended ICD-O-3/WHO 2008” category was chosen (14,604 cases). Later we obtained all available cases (14,604 cases) of primary malignant bone tumors identified with ICD-O-3 histology coding. Patients with “Not classified by ICCC or in situ” were excluded (Table 1). Also, cases with deficient information (Blanks, N/A, and Unknown) about the grade, stage, metastases, and survival months were excluded. This search algorithm resulted in a dataset of 6234 cases (Figure 1). For each case, we used sex, age at diagnosis, year of diagnosis, Median household income, Rural-Urban Continuum, primary site, ICCC histology, grade, laterality, stage, metastasis, vital status, cause of death, cause-specific death classification, the total number of malignant tumors, and survival months according to ICD-O-3.

Table 1.

Overview of 14,604 primary malignant bone tumors from the 2000–2018 SEER data (2020 release).

ICCC site recode extended ICD-O-3/WHO 2008	N (%)	ICD-O-3 histologic behavior	N (%)
Osteosarcomas, total	4937 (33.8)	9180/3: osteosarcoma, NOS	3347 (67.8)
		9181/3: chondroblastic osteosarcoma	707 (14.3)
		9182/3: chondroblastic osteosarcoma	211 (4.3)
		9183/3: telangiectatic osteosarcoma	151 (3.1)
		9184/3: osteosarcoma in paget disease of bone	51 (1)
		9185/3: small cell osteosarcoma	44 (0.9)
		9186/3: central osteosarcoma	140 (2.8)
		9187/3: intraosseous well differentiated osteosarcoma	9 (0.2)
		9192/3: parosteal osteosarcoma	198 (4)
		9193/3: periosteal osteosarcoma	55 (1.1)
		9194/3: high grade surface osteosarcoma	24 (0.5)
Chondrosarcomas	3978 (27.2)	9220/3: chondrosarcoma, NOS	3392 (85.3)
		9221/3: juxtacortical chondrosarcoma	44 (1.1)
		9230/3: chondroblastoma, malignant	38 (1)
		9240/3: mesenchymal chondrosarcoma	71 (1.8)
		9242/3: clear cell chondrosarcoma	73 (1.8)
		9243/3: dedifferentiated chondrosarcoma	360 (9)
Ewing tumor and related sarcomas of bone, total	2015 (13.8)
VIIIc1 ewing tumor and askin tumor of bone	1906 (13.1)	9260/3: ewing sarcoma	1905 (99.9)
VIIIc1 ewing tumor and askin tumor of bone	1906 (13.1)	9365/3: askin tumor	1 (0.1)
IIIc2 peripheral neuroectodermal tumor (pPNET) of bone	109 (0.7)	9364/3: peripheral neuroectodermal tumor	109 (100)
Other specified malignant bone tumors, total	2620 (17.9)
Malignant fibrous neoplasms of bone	265 (1.8)	8810/3: fibrosarxoma, NOS	76 (28.7)
		8811/3: fibromyxosarcoma	18 (6.8)
		8812/3: periosteal fibrosarcoma	5 (1.9)
		8830/3: malignant fibrous histiocytoma	166 (62.6)
Malignant chordomas	1700 (11.6)	9370/3: chordoma, NOS	1582 (93.1)
		9371/3: chondroid chordoma	101 (5.9)
		9372/3: dedifferentiated chordoma	17 (1)
Odontogenic malignant tumors	288 (2)	9270/3: odontogenic tumor, malignant	117 (40.6)
		9282/3: complex odontosarcoma	1 (0.3)
		9310/3: ameloblastoma	159 (55.2)
		9312/3: squamos odontogenic tumor	1 (0.3)
		9321/3: central odontogenic fibrosarcoma	1 (0.3)
		9330/3: ameloblastic fibrosarcoma	7 (2.4)
		9340/3: calcifying epithelial odontogenic tumor	1 (0.3)
		9342/3: odontogenic carcinosarcoma	1 (0.3)
Miscellaneous malignant bone tumors	367 (2.5)	9261/3: adamantinoma of long bones	69 (18.8)
Miscellaneous malignant bone tumors	367 (2.5)	9250/3: giant cell tumor of bone	298 (81.2)
Unspecified malignant bone tumors	1054 (7.2)
Total	14604 (100)

ICD-O-3: international classification of diseases for oncology (ICD-O), 3rd ed.

Figure 1.

Flow chart of patient selection from the SEER database.

Statistical analysis

The classical statistical methods

Statistical analysis was performed by SPSS 22.0 (IBM Corp, Armonk, NY, USA) package. Categorical variables were presented as numbers and percentages, and continuous variables were reported as mean ± standard deviation (SD) and median (interquartile range: IQR). Chi-square tests were used to compare categorical variables between groups. The normal distribution of continuous variables was assessed visually (histogram and probability graphs) and analytically (Kolmogorov-Smirnov/Shapiro-Wilk tests). The Kruskal–Wallis test was used to compare non-normally distributed variables. Survival analyses were performed with Kaplan-Meier curves and the Log-rank test. Variables with p < .05 determined by the log-rank test were entered into Multivariate Cox regression analysis to identify the independent prognostics factor for OS. The Hazard Ratios (HRs) with 95% CIs were used to show the effect of factors on OS. p < .05 was considered to be statistically significant.

Decision tree analysis: Recursive partitioning analysis (RPA)

A decision tree-based data mining technique was used in this study. We used the Rpart library function in R version 4.1.2 [http://www.r-project.org/] to create the recursive DT creation with the split criteria of p < .01 in the log-rank test. We used the classification and regression trees (CART) method in the analysis of RPA. The log-rank test was used as the splitting criterion for the regression tree during recursive partitioning analysis. The dichotomic prognostic parameter that results in the largest separation of survival at each node with a p < .05 was selected as the criterion for this node. This technique is a nonparametric methodology that creates a DT with respect to prognostic factors and their interactions which are found to be important in determining the outcome. A parent node would split into child nodes that are as homogenous as possible to dependent variables. The split also followed the rules that the corresponding cut-off points with the minimal p-value, provided the minimum p-value was ≤.0001 and that the number of patients within the child node was set to be at least 50. Then a tree-based model composed of nodes was created by recursively partitioning the study cohort. During the RPA process of this study, a set of variables were evaluated as prognostic factors (Table 2).

Table 2.

Variables for recursive partitioning analysis.

Age	<15 years, 15–44 years, 45–54 years, 55–64 years, 65–74 years, 75 years
Gender	Male versus female
Year of diagnosis	2000–2009, 2010–2014, 2015 and later
Median household income	≤$49,000, $50,000–$74,999, ≥$75,000
Histology	Osteosarcomas, chondrosarcomas, ewing sarcomas, other malignant bone tumors
Primary site	Upper limb, lower limb, skull and associated bones, vertebral column, thoracic cage, pelvic ring and sacrum
Grade	Grade 1, grade 2, grade 3, and grade 4
Laterality	Right, left, not a paired site
Stage	Localized, regional, and distant
Metastasis	No, lung metastasis, extrapulmonary metastasis/lymph node metastasis
Total number of malignant tumors	One, two or over

Results

Characteristics of the study population

In this study, the median age of the patients was 42 (19–61) years, and 55.8% of cases were male. The most commonly observed histopathological types in PMBT were osteosarcoma (41%) and chondrosarcoma (40.2%). Based on Figure 2, which illustrates the distribution of histopathological types among different age groups, Osteosarcoma was the most frequently observed type in individuals under the age of 15, accounting for 77.4% of cases under the age of 15. Whereas, the most frequently observed histopathological types in individuals aged 15-44 were osteosarcoma (50.7%) and chondrosarcoma (32.4%). The most common PMBT in groups aged 45 and above was chondrosarcoma (Figure 2).

Figure 2.

Distribution of histological types by age groups.

The highest localization was in the lower extremities with 44.9%. When the distribution according to grade was examined, 36.8% of them were grade 4 tumors. While no metastasis was observed in 85.3% of the patients, 7.1% had lung metastases. Also, 38.6% of the patients were not alive and the cause of death in 29.5% was related to the primary bone tumor (Table 3).

Table 3.

Clinical characteristics and survival outcomes of patients.

Characteristics	Total N = 6234
Age, year
Mean ± SD	41.6 ± 23.1
Median (IQR)	42 (19–61)
Sex, n (%)
Female	2758 (44.2)
Male	3476 (55.8)
Median household income inflation adj to 2019
≤$49,000	806 (12.9)
$50,000–$74,999	3501 (56.2)
≥$75,000	1927 (30.9)
Rural urban continuum, n (%)
Counties in metropolitan areas 1 million pop	3968 (63.7)
Counties in metropolitan areas of 250,000 to 1 million pop	1222 (19.6)
Counties in metropolitan areas of lt 250 thousand pop/nonmetropolitan	1044 (16.7)
Year of diagnosis, n (%)
2004–2009	2544 (40.8)
2010–2014	2284 (36.6)
≥2015–2018	1406 (22.6)
Histology, n (%)
Osteosarcomas	2554 (41)
Chondrosarcomas	2507 (40.2)
Ewing sarcomas	312 (5)
Other malignant bone tumors	861 (13.8)
Primary site, n (%)
Upper limb	916 (14.7)
Lower limb	2801 (44.9)
Skull and associated bones	739 (11.9)
Vertebral column	273 (4.4)
Thoracic cage	569 (9.1)
Pelvic ring and sacrum	936 (15)
Grade, n (%)
Well differentiated, grade 1	1115 (17.9)
Moderately differentiated, grade 2	1413 (22.7)
Poorly differentiated, grade 3	1411 (22.6)
Undifferentiated; anaplastic, grade 4	2295 (36.8)
Laterality, n (%)
Right	2512 (40.3)
Left	2495 (40)
Not a paired site/others	1227 (19.7)
Stage, n (%)
Localized	2727 (43.7)
Regional	2405 (38.6)
Distant	1102 (17.7)
Metastasis, n (%)
No	5317 (85.3)
Lung metastasis	445 (7.1)
Extrapulmonary metastasis/LYMPH NODE metastasis	472 (7.6)
Total number of malignant tumors for patient, n (%)
One	5125 (82.2)
Two or over	1109 (17.8)
Vital status, n (%)
Alive	3829 (61.4)
Death	2405 (38.6)
Cause of death, n (%)
Alive	3829 (61.4)
Bone and joint related death	1838 (29.5)
Other causes related death	567 (9.1)
SEER cause specific death classification, n (%)
Alive or dead of other cause	4396 (70.5)
Dead (attributable to this BONE cancer dx)	1838 (29.5)

SD: standard deviation, IQR: interquartile range (25P–75P).

Survival analysis results using classical statistical method

Factors associated with overall survival were evaluated with Log-rank tests, and survival rates are presented in Table 4. The 1-year survival rate is 85.5% and 5-years survival rate is 63.6% (Figure 3(a)). Survival time was found to be significantly lower in males (p < .001). Survival times are significantly shorter in age groups 65 years and older than those under 65 years old (p < .001). The survival time of those with a household income of less than 50 thousand dollars is shorter than other income groups (p = .009). Among the histological groups, the survival time of Chondrosarcomas was higher than the Ewing and Osteosarcoma groups, especially in the group consisting of tumors other than these three tumors, and the survival time was found to be considerably shorter than the patients with these three tumors (p < .001) (Figure 3(b)). The survival rates of tumors located in the pelvic and vertebral regions were found to be significantly lower than those located in other locations (p < .001). Variables with a log-rank test result of p < .05 (Table 4) were evaluated with Multivariate COX regression analysis and independent factors associated with overall survival are presented in Table 5. According to the multivariate COX regression analysis model, sex, age, median household income, histology, primary site, grade, stage, metastasis and the total number of malignant tumors were determined as independent risk factors associated with overall survival (Table 5).

Table 4.

Overall survival rates according to factors.

Total N = 6234	Log rank test p	Overall survival, months Median^a (95% CI)	1-year survival rate, %	5-years survival rate, %	10-years survival rate, %
All patients		112.1 ± 1.1	85.5	63.6	55.3
Sex	<.001
Female		119 ± 1.5	86.9	66.6	60.5
Male		106.3 ± 1.4	84.5	61.2	51.3
Age	<.001
<15 years		127.4 ± 2.5	93.8	72	64.6
15–44 years		128.8 ± 1.6	92.3	72	65.4
45–54 years		119.3 ± 2.7	88.3	66.2	60
55–64 years		107.4 ± 2.9	83.1	61.3	52.5
65–74 years		65 (48.1–81.9)	73.8	51.7	40
≥75 years		19 (14.9–23)	58.7	28.5	13.7
Year of diagnosis	.030
2000–2009		114.6 ± 1.5	86.6	65	56.9
2010–2014		72.8 ± 0.9	85.1	63.2	55
2015 and over		36.3 ± 0.5	84.4	—	—
Median household income inflation adj to 2019	.009
≤$49,000		106.1 ± 3	82.8	60.8	51.4
$50,000–$74,999		111.4 ± 1.4	85.3	62.9	54.7
≥$75,000		115.6 ± 1.1	87.2	66	58
Rural urban continuum	.490
Counties in metropolitan areas 1 million pop		112.5 ± 1.3	85.6	63.6	55.7
Counties in metropolitan areas of 250,000 to 1 million pop		112.5 ± 2.3	87.2	65	55.1
Counties in metropolitan areas of lt 250 thousand pop/nonmetropolitan		109.9 ± 2.5	83.4	62.1	54.3
Histology	<.001
Osteosarcomas		106.6 ± 1.6	85.6	58.5	52.1
Chondrosarcomas		127.4 ± 1.5	89.4	74.3	65
Ewing sarcomas		114.2 ± 4.6	88.3	65.5	53.6
Other malignant bone tumors		50 (38.3–61.7)	73	47.1	37.6
Primary site	<.001
Upper limb		123.6 ± 2.6	88.5	69.6	62.3
Lower limb		116.1 ± 1.5	87.3	65.6	57.8
Skull and associated bones		111.5 ± 3	84.9	65.2	54.5
Vertebral column		61 (17.6–104.5)	74.7	50.5	44.1
Thoracic cage		127.5 ± 3.2	92.6	73.9	64.6
Pelvic ring and sacrum		53 (39.6–66.4)	76.9	48	39.2
Grade	<.001
Well differentiated, grade 1		149.2 ± 1.9	95.6	86.9	79.2
Moderately differentiated, grade 2		131.3 ± 1.9	93.2	76.7	66.8
Poorly differentiated, grade 3		68 (52.8–83.1)	79.1	51.2	43.9
Undifferentiated; anaplastic, grade 4		68 (56.5–79.4)	80	51.7	43.6
Laterality	<.001
Right		112.9 ± 1.6	86.7	63.6	56
Left		116.1 ± 1.6	86	65.7	57.5
Not a paired site/others		102.5 ± 2.3	82.1	59.4	49.8
Stage	<.001
Localized		138.1 ± 1.4	93.6	79.4	71.2
Regional		110.4 ± 1.6	88.4	63.2	53.7
Distant		52.2 ± 2.2	59.2	26.1	20.8
Metastasis	<.001
No		123.3 ± 1.1	90.9	70.7	61.7
Lung metastasis		51.7 ± 3.3	63.2	24.9	21.1
Extrapulmonary metastasis/LYMPH NODE metastasis		38.6 ± 2.9	46.2	18.5	14.2
Total number of malignant tumors for patient, n (%)	<.001
One		117.9 ± 1.1	87.1	66.2	59.5
Two or over		87.1 ± 2.3	78.6	52.2	58.2
Decision tree, nodes	<.001
Node 1		135 ± 1.1	95	77	69.6
Node 2		139 (123.3–154.7)	92.8	77.1	60.9
Node 3		61 (44.6–77.4)	79.6	50	23.1
Node 4		20 (15.9-24)	62.5	30.5	20.2
Node 5		17 (15.1–18.9)	59.2	26.1	20.8

NA: not available.

^aSince the median survival time could not be reached, the mean ± standard error is presented.

Figure 3.

(a) Kaplan-Meier curves for overall survival, (b) Kaplan-Meier curves for overall survival according to histology.

Table 5.

Multivariate COX regression analysis for overall survival (OS) for patients identified in the SEER program database.

	Multivariate COX regression analysis model for OS
	Adjusted HR (95% CI)	p
Sex
Female	Reference
Male	1.15 (1.06–1.25)	<.001

Age
<15 years	Reference
15–44 years	1.55 (1.34–1.79)	<.001
45–54 years	2.76 (2.31–3.29)	<.001
55–64 years	3.39 (2.85–4.03)	<.001
65–74 years	4.81 (4.03–5.73)	<.001
≥75 years	8.49 (7.14–10.11)	<.001
Year of diagnosis
2004–2009	Reference
2010–2014	1.07 (0.97–1.17)	.168
2015 and over	1.08 (0.96–1.23)	.204
Median household income inflation adj to 2019
<=$49,000	1.25 (1.09–1.42)
$50,000–$74,999	1.09 (1.01–1.20)	.001
>=$75,000	Reference	.045
Histology, n (%)
Chondrosarcomas	Reference
Osteosarcomas	0.92 (0.74–1.15)	.470
Ewing sarcomas	1.32 (1.17–1.50)	<.001
Other malignant bone tumors	1.28 (1.12–1.45)	<.001
Primary site
Upper limb	Reference
Lower limb	0.96 (0.84–1.09)	.541
Skull and associated bones	1.01 (0.83–1.23)	.878
Vertebral column	1.56 (1.22–1.98)	<.001
Thoracic cage	0.77 (0.64–0.94)	.012
Pelvic ring and sacrum	1.43 (1.23–1.67)	<.001
Grade
Well differentiated, grade 1	Reference
Moderately differentiated, grade 2	1.39 (1.17–1.66)	<.001
Poorly differentiated, grade 3	2.96 (2.49–3.51)	<.001
Undifferentiated; anaplastic, grade 4	3.17 (2.67–3.78)	<.001
Laterality
Right	Reference
Left	0.95 (0.87–1.05)	.327
Not a paired site/others	1.07 (0.93–1.24)	.343
Stage
Localized	Reference
Regional	1.52 (1.37–1.69)	<.001
Distant	2.48 (2.07–2.97)	<.001
Metastasis
No	Reference
Lung metastasis	1.91 (1.56–2.33)	<.001
Extrapulmonary metastasis	2.63 (2.19–3.17)	<.001
Total number of malignant tumors for patient
One	Reference
Two or over	1.14 (1.03–1.27)	.007

Survival analysis results using decision-tree model

In the DT analysis, we identified the variables that play important roles in explaining vital status (Table 5). The stage was the most important determining factor for vital status. This first-level split produced the two initial branches of the classification tree: the localized and regional stage versus the distant stage. DT identified the following five nodes (groups) having different levels of survival possibility (Figure 4): (Node 1) localized/regional stage and < 65 ages; (Node 2) localized/regional stage, grade ≤ 2, and 65–74 ages; (Node 3) localized/regional stage, grade ≤ 2, and 75 ages and upper; (Node 4) localized/regional stage, 65 ages and upper and grade ≥ 3; (Node 5) distant stage. Thus, three prognostic factors were identified (Stage, age, and grade) for survival, resulting in five terminal nodes (Figure 4). Seventy-five percent of patients in Node 1 were alive and 72.3% of patients in Node 5 died.

Figure 4.

Decision tree constructed by recursive partitioning analysis from 6234 patients with malignant bone tumors. Proportions of patients with alive (blue) and dead patients (green) are represented at each node.

It was found that the survival time of the patients in Nodes 3, 2, and 1 was longer than the patients in Nodes 4 and 5 (p < .001) (Figure 5). The longest surviving terminal node (node 1) included only 1031 (25%) events (death) in 4120 patients. For these patients, the mean survival time was 135 ± 1.1 months. The ten-year survival rate was 69.6% (Table 4). A second terminal node (node 2) with a relatively long survival included 101 (31.2%) events in 324 patients with localized/regional stage, grade <2, and age of diagnosis 65–74 years old. These patients had a median survival of 139 months and the ten-year survival rate was 60.9% (Table 4). The shortest surviving terminal node (node 5) included 801 (72.3%) events in 1102 patients with the distant stage (Figure 4). These patients had a median survival of only 17 months. The ten-year survival rate was 20.8% (Table 4). In the COX regression analysis model in which 5 terminal nodes were included, the HRs were calculated (reference: node 1): HR = 1.3 (95% CI: 1.1–1.6; p = .005) for node 2, HR = 3.2 (95% CI: 2.7–3.9; p < .001) for node 3, HR = 5.1 (95% CI: 4.5–5.7; p < .001) for node 4 and HR = 5.4 (95% CI: 4.9–5.9; p < .001) for node 5.

Figure 5.

Kaplan-Meier curves for overall survival according to decision tree nodes.

Discussion

In this study, we applied DT analysis alongside traditional analysis methods in the survival problem of primary malignant bone tumors extracted from SEER records between 2000 and 2018. We used the SEER database in this study as it allows access to actual patient data with a large sample. The study is one of the largest series of studies on primary malignant bone tumors with 6234 patients. In the DT model, stage, grade, and age were variables that seem to explain the prognosis. This study solely focused on overall survival. Among the histological types, chondrosarcoma has the longest survival time. Although the survival times of osteosarcoma and Ewing sarcomas are shorter than chondrosarcoma, it was observed that there was no significant difference between the survival times. The other PMBTs had a median survival time of 50 (38.3–61.7) months, and this group had a shorter survival time than the other three common PMBTs.

Decision trees are visualization and prediction tools that belong to the family of supervised learning algorithms used to solve regression and classification problems. One advantage of DTs is that they can be implemented quickly and are easy to interpret.¹³ The main purpose of our decision tree was to visualize the data and create a diagram that we can use for discussion. Stage, age, and grade were the most prognostic factors in both our DT model and the multivariate COX regression analysis. Also, the Multivariate Cox regression model is the most common tool for simultaneously investigating the influence of several factors on the survival time of patients.¹⁴ Other factors associated with poor prognosis in the multivariate COX regression model are male gender, low income, pelvic and vertebral location, metastasis, and having two or more tumors. In this model, the high grade was associated with a 3-fold decrease in survival, while being 75 years or older was associated with an 8-fold reduction in survival. Distant or metastatic tumors were also associated with a 2.5-fold reduction in survival. Wang et al.¹⁵ determined the stage, age, and grade as factors that are most associated with survival in the sample of patients with primary bone sarcoma in the hands or feet from the SEER database. In the study by Fukushima et al.,¹⁶ age and histological grade were found to be the two most prognostic markers for bone sarcomas. It has been shown that those aged 65 and over had 4.4 times more mortality and those with high histological grade had 3.8 times more mortality. Histological cancer grade is one of the most important prognostic markers, an indicator of the differentiation of tumor cells used to express the grade of cancer progression.^17,18

In our study, the 5-years survival rate was 63.6% and the 10-years survival rate was 55.3% in the PMBT patients. There has been no significant improvement in the 5-years survival of primary bone cancer over the past 25 years.^2,19 In the United States, the National Cancer Institute reported an overall 5-years survival of 66%.^2,20 In the univariate survival analysis, older adults (>65 years), males, patients with tumor sites at the vertebral column, pelvic ring and sacrum, osteosarcoma and Ewing sarcoma, patients with low income, location in non-metropolitan areas, high grade, distant stage, metastasis and two and over malignant tumors had a lower 5-years survival rate. Similarly, Hu et al.²¹ found a lower 5-years survival rate in men, the elderly older adults (>65 years), patients with tumor sites at the vertebral column, pelvic ring and sacrum, osteosarcoma, and Ewing sarcoma, low income, location in non-metropolitan areas.

In our study, similar to the literature, Osteosarcoma was the most frequently observed PMBT in the <15 years and 15–44 years age groups.^16,18,22 With the introduction of adjuvant chemotherapy, 10-years survival has increased from 30% to approximately 50%, but there has been no improvement in 10-years survival since the 1990s.^22,23 The 10-years survival rate for osteosarcoma in this study was 52.1%.

One of the main limitations of our study is that SEER records do not contain detailed information and variations in specific treatment strategies. SEER database does not include variables such as pathological fracture and surgical margin status, which are known as potential prognostic factors.

Conclusion

The decision tree analysis we utilized to evaluate mortality in PMBT showed five decision nodes among 11 prognostic factors by examining the stage, age, and grade variables concurrently. The implementation of decision trees may prove beneficial in the assessment of the prognosis of PMBT patients and could yield satisfactory outcomes. The extracted rules from decision trees offer insights into the risk factors in specific patient populations and can be utilized by clinicians in their decision-making process for individual patients. It is recommended to utilize decision trees in conjunction with Cox regression analysis to identify risk factors and their effects on survival.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Ethical approval

No ethics committee approval nor patient approval forms are necessary to use data on the SEER platform. The SEER research contract was signed online to access the database. All patient identifiers and all personal data in this study are removed from the SEER database. As a result, studies using the SEER database are exempt from institutional review board approval.

Informed consent

No informed consent nor patient approval forms are necessary to use data on the SEER platform.

English language editing

This manuscript was edited by Gazi University Academic Writing, Application and Research Center. The manuscript was edited for organizational flow and proper English grammar, vocabulary, punctuation, and spelling. (Certificate Number: 05.01.2023/002).

ORCID iDs

Aliekber Yapar

Ugur Bilge

Appendix

References

Kumar

Gupta

. Global incidence of primary malignant bone tumors. Curr Orthop Pract 2016; 27: 530–534. DOI: 10.1097/bco.0000000000000405

Pullan

Lotfollahzadeh

Sökmen

Ulucaköy

. Diagnostic and prognostic role of mean platelet volume and mean platelet volume/platelet ratio in patients with primary malignant bone tumor. Jt Dis Relat Surg 2021; 32: 198–203. DOI: 10.5606/ehc.2021.76357 (accessed 20 January 2021.

Decherchi

Pedrini

Mordenti

, et al. Opportunities and challenges for machine learning in rare diseases. Front Med 2021; 8: 1696.

Liu

C-Y

Cheng

C-Y

Yang

S-Y

, et al. Mortality evaluation and life expectancy prediction of patients with hepatocellular carcinoma with data mining. Healthcare 2023; 11: 925.

Priyam

Abhijeeta

Rathee

, et al. Comparative analysis of decision tree classification algorithms. J. Eng. Technol 2013; 3: 334–337.

Song

Y-Y

Ying

. Decision tree methods: applications for classification and prediction. Shanghai Arch Psychiatry 2015; 27: 130.

National Cancer Institute’s (NCI’s) . Surveillance, epidemiology, and end results (SEER) Program. Number of Persons by Race and Hispanic Ethnicity for SEER Participants (2020 Census Data). Available at https://seer.cancer.gov/registries/data.html#a5

CDC. U.S. Cancer Statistics Home . Surveillance, epidemiology, and end results (SEER) program, CDC. U.S. Cancer Statistics Home. 2022. Available at https://www.cdc.gov/cancer/uscs/technical_notes/contributors/seer.htm#:∼:text=The_SEER_Program_issues_a,by_researchers_and_the_public.&text=A_public_use_interactive_website,incidence_rates_called_SEER*Explorer (accessed 14 October 2022).

10.

Benndorf

Neubauer

Langer

, et al. Bayesian pretest probability estimation for primary malignant bone tumors based on the surveillance, epidemiology and end results program (SEER) database. Int J Comput Assist Radiol Surg 2017; 12: 485–491.

11.

Canda

MŞ

Eroğlu

Hapa

. International classification of diseases for oncology (ICD-O) coding system, language for oncology implications and update at orthopaedic oncology. Turk Onkol. Derg 2021; 36: 242–246.

12.

Aytekin

Öztürk

Amer

, et al. Epidemiology, incidence, and survival of synovial sarcoma subtypes: SEER database analysis. Orthop Surg 2020; 28. DOI: 10.1177/2309499020936009

13.

Raitière

Berthelot

Fauvel

, et al. The dangerous and contradictory prognostic significance of PVR < 3WU when TAPSE < 16 mm in postcapillary pulmonary hypertension. ESC Heart Fail 2020; 7: 2398–2405. DOI: 10.1002/ehf2.12785 (accessed 20 July 2020).

14.

George

Seals

Aban

. Survival analysis and regression models. J Nucl Cardiol 2014; 21: 686–694. DOI: 10.1007/s12350-014-9908-2 (accessed 09 May 2014).

15.

Wang

, et al. Prognostic factors for survival among patients with primary bone sarcomas of small bones. Cancer Manag Res 2018; 10: 1191–1199. DOI: 10.2147/cmar.S163229 (accessed 26 May 2018).

16.

Fukushima

Ogura

Akiyama

, et al. Descriptive epidemiology and outcomes of bone sarcomas in adolescent and young adult patients in Japan. BMC Musculoskelet Disord 2018; 19: 297. DOI: 10.1186/s12891-018-2217-1 (accessed 20 August 2018).

17.

Edge

Compton

. The American joint committee on cancer: the 7th edition of the AJCC cancer staging manual and the future of TNM. Ann Surg Oncol 2010; 17: 1471–1474. DOI: 10.1245/s10434-010-0985-4 (accessed 25 February 2010).

18.

Jiang

Wang

Wei

. Construction and validation of nomograms for predicting the prognosis of juvenile osteosarcoma: A real-world analysis in the SEER database. Technol Cancer Res Treat 2020; 19. DOI: 10.1177/1533033820947718 (accessed 16 October 2020).

19.

Keegan

Ries

Barr

, et al. Comparison of cancer survival trends in the United States of adolescents and young adults with those in children and older adults. Cancer 2016; 122: 1009–1016. DOI: 10.1002/cncr.29869 (accessed 06 June 2016).

20.

Whelan

McTiernan

Cooper

, et al. Incidence and survival of malignant bone sarcomas in England 1979–2007. Int J Cancer 2012; 131: E508–E517. DOI: 10.1002/ijc.26426 (accessed 14 September 2011).

21.

Deng

, et al. Trends in tumor site-specific survival of bone sarcomas from 1980 to 2018: A surveillance, epidemiology and end results-based study. Cancers 2021; 13: 5381.

22.

Simpson

Brown

. Understanding osteosarcomas. Jaapa 2018; 31: 15–19. DOI: 10.1097/01.JAA.0000541477.24116.8d (accessed 07 July 2018).

23.

Berner

Johannesen

Berner

, et al. Time-trends on incidence and survival in a nationwide and unselected cohort of patients with skeletal osteosarcoma. Acta Oncol 2015; 54: 25–33. DOI: 10.3109/0284186x.2014.923934 (accessed 25 June 2014).