Development of a model with which to predict the life expectancy of patients with spinal epidural metastasis




The surgical treatment of spinal epidural metastasis is evolving. To be a surgical candidate, a patient should have a life expectancy of at least 3 months. Estimation of survival by experienced specialists has proven to be unreliable.


The Cox proportional hazards model was used to make a prediction model. To validate the model, Efron optimism correction by bootstrapping was performed. Retrospective data of patients treated for a spinal metastasis were used. Possible predictive factors were defined based on clinical experience and the literature. Statistical methods and clinical knowledge were also used to reveal an optimal set of predictors of survival. Data from patients treated at the Department of Radiation Oncology for spinal metastasis between 1998 and 2005 were evaluated.


The case notes of 219 patients form the base of this study. In the final model, only 5 variables were required to predict the survival of a patient with spinal metastasis: sex, location of the primary lesion, intentional curative treatment of the primary tumor, cervical location of the spinal metastasis, and Karnofsky performance score. Examples with different predictors are given. The Rmath image index of Nagelkerke was 0.36 (95% confidence interval [95% CI], 0.28–0.48) and the c-index 0.72 (95% CI, 0.68–0.77).


A reliable and simple model with which to predict the survival of a patient with spinal epidural metastasis is presented. Without the need for extensive investigations, survival can be predicted and only 5 easily obtainable parameters are required. Cancer 2007. © 2007 American Cancer Society.

Spinal metastasis is frequently encountered, occurring in up to 70% of the patients with terminal cancer.1 In the last decennium, the management of spinal epidural metastasis has changed. At first, surgery was proven to fail in the management of spinal metastasis and, compared with radiation therapy, surgery had more complications, whereas neurologic recovery and survival rates did not improve. At that time, surgery was abandoned for the treatment of spinal metastases. It should be stated that surgery was in fact laminectomy.2 Gradually, surgical techniques, implants, and expertise evolved and instead of laminectomy, 360° decompression and stabilization was possible and became more usual in managing metastatic spine disease.3 A recent, randomized, controlled trial proved the effectiveness of surgery in spinal metastatic disease with regard to the improvement of neurologic function.4 It did not contribute to a longer survival, but to an improvement in the quality of remaining life. One of the most important selection criteria for surgery is life expectancy. Estimation of the life expectancy includes numerous factors such as nature and extension of the primary tumor, clinical performance of the patient, presence of metastases, etc.

Several score systems have been developed. Those designed by Tokuhashi et al.5 and Tomita et al.6 were developed to determine which surgical procedure was needed. The first system was recently revised.7 The estimation of survival was only a very rough indication. However, the goal of the study was essentially the same as the first. Furthermore, the influence of race was not discussed, questioning the validity regarding the European or American populations. Recently, a system was also developed to predict the survival of patients with spinal metastasis.8 This system is of limited validity because malignant melanoma, multiple myeloma, and renal cell carcinoma were not included. A cervical location was a reason for exclusion. Furthermore, it only applies to patients with little or no neurologic impairment and without evidence of vertebral collapse or instability.9 It is obvious that these are not the patients that will be referred for surgical treatment at the current time. It is generally accepted that 1 of the criteria to be a surgical candidate is that a patient should have a life expectancy of at least 3 months.10 Estimation of survival is difficult, demanding a multidisciplinary approach and good clinical judgment. Nevertheless, a more accurate estimation of survival is required, necessitating the development of a prediction model.


Data Collection

In the database of the Department of Radiation Oncology from the Radboud University Nijmegen Medical Centre, Nijmegen, The Netherlands, a search was performed to find all patients (aged ≥18 years) who underwent radiation therapy for spinal epidural metastases. None of the patients underwent nonconventional therapies, such as stereotactic radiosurgery. The adherence population of the Department of Radiation Oncology is comprised of approximately 500,000 persons, the largest percentage of which are Caucasian. The diagnosis of spinal epidural metastasis was based on findings from plain X-rays, computed tomography (CT), and/or magnetic resonance imaging (MRI) scans in patients with a tissue-proven primary lesion that was already known for up to a maximum of 3.0 years. Otherwise, a biopsy of the vertebral lesion was performed. Ideally, the diagnosis was proven histologically. The patients presented with neurologic deficits due to spinal cord compression and/or severe pain.

Demographic and Clinical Factors

Based on the literature5, 11–16 and clinical practice, possible predictive factors were identified. These factors were extracted from the patients' case notes. Sex was noted, as was the result of the pathologic examination of the primary tumor, the grade of differentiation, the location of the primary lesion, the TNM classification, and whether the lesion was treated with a curative intent. The patients' age when they first presented with the primary lesion also was noted. With regard to the spinal metastasis, its location and the number of affected spinal levels was determined. The time between diagnosis of the primary tumor and of the spinal metastasis was calculated. The presence of metastases elsewhere within the body and the possibility of an additional systemic treatment for the spinal metastasis were also noted. Finally, data regarding the patients' clinical condition were extracted: Karnofsky performance score (KPS), pain in the spine, sensibility loss, loss of strength, and the presence of micturation disturbances. If >1 primary tumor was present, the patient was considered as actually suffering from the tumor that was confirmed by a pathologic specimen of the vertebral lesion or other active pathologic lesion. If a biopsy was not performed, the pathologic examination was considered unknown.


Date of death was the only primary endpoint. Survival time was defined as the time between first visit to the radiotherapist for spinal metastasis and death. In 9 of the patients, the date of death was not known or the patients were still alive at the end of the study. Their survival times are right-censored.

Statistical Analysis

Model building

A prediction model was built using Cox proportional hazards survival analysis. Using clinical knowledge, the items were reduced by simplification of scores and the clustering of data. Variables missing > 50% of their values were excluded from analysis.

The proportional hazards assumption was checked both graphically and by hypothesis testing as described in Collett.17 Graphic examination was performed in advance of model fitting using log-cumulative hazard plots. Hypothesis tests were performed after generating scaled Schoenfeld residuals and by adding time-dependent variables.

Log-rank tests and univariate Cox regressions were used to detect a set of factors each being of significant importance in predicting survival. Again using clinical knowledge and correlation analysis, this set of potential univariate predictors was reduced. Forward and backward selection procedures, combined with clinical judgment, were performed to reveal an optimal set of predictors for survival.

Missing data

Variables with ≥50% missing values were excluded from analysis.


To assess the accuracy of the final model we calculated Rmath image and the concordance index, c. The generalized Rmath image index is a measure of predictive ability for a Cox proportional hazards model that ranges from 0 (poor) to 1 (perfect). The c-index is a generalization of the area under the receiver operating characteristic (ROC) curve used in logistic regression modeling and can be regarded as an index of discrimination for survival models. The c-index is the probability of concordance between predicted and observed survival, with c = 0.5 for random prediction and c = 1 for perfect discrimination.

We used Efron optimism correction by bootstrapping to correct these indices for overfitting.

With the same procedure we estimated a linear shrinkage factor γ to quantify the amount of overfitting and to shrink the estimated hazards ratios toward unity.18, 19 External validation is the subject of a separate study and will be presented in the future.

Sample size

Because a model has to be simple, clinically useful, and plausible, the use of interaction terms should be minimized.19 To avoid overfitting the model, 10 to 15 patients are needed for each possible predictive factor or interaction term.18 Therefore, nearly 200 patients are needed to incorporate 15 predictors. To obtain at least 200 patients, the database was searched from January 1998 until December 2005.

SAS statistical software (version 8.1; SAS Institute, Inc, Cary, NC) was used for the statistical analysis. Significance was set at P < .05. Data are presented as the mean ± standard deviation or, in the case of survival, as the median ± standard deviation.


Within the investigation period 253 patients appeared to be eligible. However, after studying the case notes, 34 patients were excluded. The reasons were: younger than the minimum age requirement of 18 years, primary bone tumor, leptomeningeal metastasis, tumor of the spinal cord, or the absence of any pathologic examination. The data for the remaining 219 patients was evaluable. Thirty-four patients presented solely with severe pain, whereas the remaining patients were found to have some neurologic deficit. The mean age at death was 63.4 ± 12.1 years (range, 20.4–88.5 years). Males outnumbered females: 128 versus 91. The mean age at the time of the first presentation of the spinal metastatic lesion was 62.7 ± 12.5 years (range, 20.3–88.0 years). Therefore, the median time between the presentation of the spinal lesion and death or the survival time was 0.25 years or 3 months (range, 0.0–6.2 years). The median time between the first presentation of the malignancy and presentation of a spinal metastasis was 1.7 years (range, 0.0–22.9 years).

Considering the time between the diagnosis of the primary lesion and the moment the spinal metastasis was discovered, the proportional hazard assumption was not fulfilled when survival exceeded 10 months. Also for KPS, a cervical location of the spinal metastasis, curatively intended treatment of the primary lesion, and a breast cancer or cancer of the prostate, the proportional hazards assumption was not met. However, by censoring all patients who were still alive 10 months after first presentation of the spinal metastasis, all these variables were found to meet the proportional hazards assumption.20

Data are reduced by excluding variables with >50% missing values, clustering of data, and simplifying score systems using clinical and statistical knowledge. Only the data regarding the TNM classification and the grade of differentiation of ≥50% (112 patients and 116 patients, respectively, out of 219) are missing. Therefore, these were excluded from further analysis.

After univariate analysis, the following items were selected for further analysis: sex (P < .0001), location of the primary cancer (P < .0001), curatively intended treatment of the primary tumor (P < .0001), additional systemic treatment possible for spinal metastasis (P = .0043), cervical location of spinal metastases (P = .035), KPS (P ≪ .0001), time between diagnosis of the primary lesion and spinal metastasis (hazards rate 0991 per month; P ≪ .0001), and age at the time of the first diagnosis of the primary cancer (P = .0124).

The histologic examination of the primary tumor (Table 1) and its location (Table 2) were clustered. For example, carcinoma of the prostrate is generally an adenocarcinoma and a carcinoma of the breast a ductal carcinoma. The combination of both did not create a loss of information, whereas the investigated items were reduced from 34 variables to 5 (lung carcinoma, prostate carcinoma, breast carcinoma, renal clear cell carcinoma, and other). Because carcinoma of the breast was found to have the best survival rate, this is categorized as a reference for calculating the hazards rates of lung carcinoma, prostate carcinoma, renal cell carcinoma, and carcinoma at other sites. When the final model was defined, it appeared that the hazards ratio of prostate carcinoma compared with breast carcinoma was nearly 1. Therefore, these 2 were grouped into 1 reference group, thereby reducing the total of 5 groups to only 4 (Table 3).

Table 1. Histology of Primary Tumour or Spinal Metastasis
Histology of primary lesion or spinal metastasisNo.
Adenocystic carcinoma1
Alveolar rhabdomyosarcoma1
Burkitt lymphoma1
Ductal carcinoma44
Endometrial carcinoma3
Follicular thyroid carcinoma1
Renal clear cell carcinoma12
Nonsmall cell carcinoma2
Hepatocellular carcinoma3
Multiple myeloma11
Small cell carcinoma2
Lobular carcinoma4
Non-Hodgkin lymphoma2
Papillary thyroid carcinoma1
Squamous cell carcinoma21
Urothelial carcinoma3
Table 2. Location of Primary Tumor Known at the Moment of Presentation for Spinal Metastasis
Location of primary tumorNo.
Digestive tract18
Table 3. Variables With Their Respective Number of Items to Which an Answer Is Needed at the Beginning of the Study and After Various Methods of Data Reduction and at the Final Model
Item before building modelNo. of itemsData reductionItems after reductionFinal items after backward stepdown regression and bootstrapping
  1. KPS indicates Karnofsky performance score.

Primary lesion
Age at first diagnosis of primary lesion1 (continuous)None10
Time between primary lesion and spinal metastasis1 (continuous)None10
Histology of primary tumour20Clustering50
Grade of differentation?Exclusion (> 50% missing)00
Site of primary lesion14Clustering (with histology)54
TNM classification3Exclusion (> 50% missing)00
Primary tumor treated with curative intent1None11
Spinal metastasis
Location11Absent statistical association11
No. of affected vertebra4Clustering00
Metastasis elsewhere5Clustering00
Additional systemic treatment1None10
Neurologic deficit9Strong statistical association with KPS00
KPS8Simplifying score55
Total79 2012

The findings of neurologic examinations are statistically strongly associated with the KPS; therefore, these were also combined. Furthermore, KPS was grouped into 5 groups: Group 1: 10% to 20% (severely ill or moribund); Group 2: 30% to 40% (bedridden or severely dependent); Group 3: 50% to 70% (disabled, needing assistance in daily life); Group 4: 80% (minor deficit); and Group 5: 90% (nearly normal). None of the patients had a 100% KPS. To calculate hazards rates, the group with a KPS of 90% was taken as a reference group. In this way, a reduction from 17 items to 4 was established. A possible correlation between cervical location and KPS was assumed and investigated. However, it did not appear to be statistically correlated and therefore these were not clustered but investigated separately.

We investigated the presence of visceral metastases (brain, liver, lung, other) as well as bone metastases at the time of presentation of the spinal metastasis. None of these or combinations of them were found to contribute statistically to the prediction model (P = .93). The number of vertebral metastases was also not found to be correlated with survival (P = .92). Therefore, these items were abandoned in the definitive prediction model.

If an additional systemic treatment (eg, a hormonal therapy) could be given it was accompanied with a statistically significant reduction of the hazards ratio. However, in the final construction of the model, this parameter did not appear to contribute anything and therefore was excluded.

The time between first presentation of malignancy and first presentation of spinal metastasis did appear to contribute to prognosis. The hazards ratio for each month was only 0.991 (P ≪ .0001), and in neither the forward nor the backward selection was a statistically significant contribution noted. This parameter was not included either.

The other mentioned items were not considered for further analysis because at univariate analysis statistical significance was not reached.

Finally, the prediction model consists of 12 parameters and is depicted in Table 4. In fact, it consists of only 5 variables because knowledge of 1 excludes several others. For example, if the KPS is 80% all the other KPS are known not to be present. Therefore, we should know: sex, location of the primary tumor (4 possible categories), intended curative treatment of the primary tumor, cervical location of the spinal metastasis, and KPS (5 possible categories).

Table 4. Definitive Model Representing the Hazard Ratios Estimating an Unfavorable Outcome
 Estimated coefficient in final proportional hazards modelHR95% CL for the HRP
  • HR indicates hazards ratio; 95% CL, 95% confidence limits; KPS, Karnofsky performance score.

  • *

    The HR was calculated compared with patients with a breast or a prostate carcinoma.

  • This is a true or false expression. It is true when considering the histology, location, and extension of the primary tumor (including metastases); the initial treatment is aimed at cure. Otherwise, when the initial treatment is palliative the expression is false. For example, an early breast carcinoma can be treated curatively by surgery in combination with radiation therapy, chemotherapy, or hormonal therapy, or combinations of these. The expression would be true. On the contrary, if distant metastases are present the treatment would be only palliative. Therefore, the expression would be false.

  • Cervical is defined as a symptomatic metastasis located in the cervical spine. Metastases may be present in other parts of the spine, but they are not symptomatic.

  • §

    Patients with a KPS of 90% were included for the calculation of the reference group for the HR.

Gender (female vs male)−0.4020.6690.469–0.954.0262
Lung carcinoma*0.8032.2321.333–3.737.0023
Kidney carcinoma*1.0292.7981.477–5.299.0016
Other carcinoma*0.7192.0521.423–2.958.0001
Curatively intended treatment of primary tumor−0.3940.6750.469–0.969.0333
Cervical location of spinal metastasis0.4891.6301.118–2.376.0111
KPS (10–20)§3.44931.4558.914–110.997.0001
KPS (30–40)§2.0407.6872.725–21.689.0001
KPS (50–70)§1.7355.6702.030–15.835.0009
KPS (80)§1.4924.4461.529–12.933.0062

The predictive ability of our final model is reflected in Rmath image = 0.36 (95% confidence interval [95% CI], 0.28–0.48) and the discriminative performance in the concordance index c = 0.72 (95% CI, 0.68–0.77). To assess the amount of optimism in these indices, we performed 150 bootstrap replications. The optimism in Rmath image appeared to be 0.07 (19%), in c only 0.02 (3%). This internal validation procedure also yielded an estimate for the needed linear shrinkage γ = 0.88.

Examples of the prediction model are shown in Table 5 and Figure 1.

Figure 1.

Predicted survival of (A) a patient with prostate carcinoma, with a Karnofsky performance score (KPS) of 80%, no cervical metastases, and receiving curative treatment of the primary lesion; (B) a patient with prostate carcinoma, a KPS of 50% to 70%, cervicalmetastases, and not receiving curative treatment of the primary lesion; (C) a male patient with renal cell carcinoma, a KPS of 80%, noncervical metastases, and receiving curative treatment of primary cancer; and (D) a patient with renal cell carcinoma, a KPS of 80%, cervical location of metastases, and not being receiving curative treatment of the primary lesion. The solid line depicts the estimated survival. Interrupted lines show the upper and lower borders of the 95% confidence limits.

Table 5. Prediction of Survival at 3 Months and 6 Months After the First Diagnosis of Spinal Metastasis
GenderPrimary tumor treated curativelyPrimary tumorKPS, rangeCervical locationSurvival at 3 MoSurvival at 6 Mo
  1. KPS indicates Karnofsky performance score; F, female; +, positive; −, negative; M, male.

F+Breast50–70%0.79 (0.71–0.88)0.64 (0.53–0.78)
FBreast50–70%0.71 (0.60–0.83)0.52 (0.39–0.68)
F+Breast50–70%+0.68 (0.56–0.83)0.48 (0.34–0.69)
FBreast50–70%+0.57 (0.43–0.76)0.34 (0.20–0.58)
F+Breast80%0.83 (0.75–0.92)0.70 (0.59–0.85)
FBreast80%0.76 (0.66–0.88)0.59 (0.46–0.77)
F+Breast80%+0.74 (0.62–0.89)0.56 (0.41–0.78)
FBreast80%+0.64 (0.50–0.83)0.43 (0.27–0.69)
M+Prostate50–70%0.71 (0.60–0.83)0.51 (0.38–0.69)
MProstate50–70%0.60 (0.49–0.73)0.37 (0.26–0.53)
M+Prostate50–70%+0.57 (0.42–0.77)0.34 (0.19–0.59)
MProstate50–70%+0.43 (0.29–0.64)0.20 (0.09–0.43)
M+Prostate80%0.76 (0.65–0.89)0.59 (0.45–0.79)
MProstate80%0.67 (0.54–0.81)0.46 (0.32–0.66)
M+Prostate80%+0.64 (0.48–0.85)0.43 (0.25–0.72)
MProstate80%+0.52 (0.35–0.75)0.28 (0.14–0.57)
M+Kidney80%0.46 (0.28–0.78)0.23 (0.09–0.61)
MKidney80%0.32 (0.14–0.71)0.11 (0.03–0.52)
M+Kidney80%+0.29 (0.11–0.74)0.09 (0.02–0.55)
MKidney80%+0.16 (0.04–0.67)0.03 (0.00–0.47)


The surgical management of metastatic disease of the spine continues to evolve. If the patient and the presenting pathology are feasible for surgery, 1 of the major criteria is the expected life expectancy of the patient. Generally, if the expected survival is > 3 months, patients are eligible for surgery.10 Of course, the predicted survival is not the only criterion with which to decide whether the patient is a surgical candidate. The wishes of the patient, comorbidity, and local characteristics are of course of equal importance.

In general, a search for other metastases is performed. The patient is assessed by an oncologist and a survival rate is estimated. In our experience, as has been confirmed by other studies, this estimation is not very accurate.21, 22 Therefore, a prediction model is required.

The simplicity of our model is appealing and unique. Only 5 parameters should be assessed, eliminating the need for an extensive search before treatment: sex, location of the primary tumor (4 possible categories), intended curative treatment of the primary tumor, cervical location of the spinal metastasis, and KPS. All data can simply be determined by the specialist. A more precise estimate of a survival beyond 10 months cannot be made; however, this is not relevant for decision-making. Only a life expectancy of at least 3 months is of value.

Some of our findings have been previously described by others. Others have also found a better prognosis for breast or prostate carcinoma.8, 15 Sex did not appear to influence the final prognosis.14 Prognosis was found to be highly correlated with the functional KPS.15 The ‘protective’ working of additional therapy to radiation therapy for spinal metastasis has been described.14

However, a striking feature is that the presence of other metastases did not appear to influence prognosis. This was statistically proven, but also biologically plausible. Multiple microscopic or macroscopic metastases will be present when 1 metastatic lesion presents itself. Certainly this is true for bone metastases. From this point of view, survival will not alter whatever the size of the lesion.

Because we wanted to predict survival at the moment the patient presents with a spinal metastasis requiring therapy (either surgery, radiation therapy, or a combination of both), we did not include the response to radiation therapy as has been described before.8 Furthermore, patients who were not referred for radiation therapy are not represented. We assume that these are not the patients who would be surgical candidates. Either they are asymptomatic or they are symptomatic but their clinical condition precludes referral for radiation therapy. All other patients are referred for radiation therapy. More recently, such patients also will be referred for surgical therapy. In our department, all surgical candidates will undergo postoperative radiation therapy. Therefore, we do not assume that the nonrepresented patients will have a significant impact on the predicted survival of the patients who are possible surgical candidates.

Finally, we emphasize that this approach only estimates the survival of any individual patient. The wide 95% CIs, particularly with poorer prognoses, indicate great caution in interpreting these results. We assume that the results of the ongoing study to test the external validity will make the attitude toward this approach less conservative.

The possibility of estimating the survival of each separate patient either by point estimate or 95% CIs has to our knowledge not been described to date. This possibility is of great importance to every spinal surgeon or radiation oncologist. Each of the variables within the prediction model should be known when the patient is referred. Without the expertise and personal judgment of a medical oncologist, a reliable estimation of the survival rate of the patient can be made. If surgery is technically and emotionally feasible, the expected survival is the most important variable. The predictive ability was estimated. Based on the Rmath image and c-index, the model does appear to predict survival reasonably well.23 Finally, the estimation of survival can be performed within a few minutes. We currently are working on a model that can be consulted on the Internet (available at URL:


We thank Mrs. Lynda Gerdsen, Drury DCR (R), for revising the text with regard to style and English grammar.