Page 86 - Predicting survival in patients with spinal bone metastasesL

CHAPTER VI
The primary data sources were the patient’s clinical files, radiology reports and admission forms. Baseline variables obtained before start of treatment were gender, age, primary tumor, location and number of SBM, the presence of visceral and/or brain metastases, the presence of extraspinal bone metastases, pretreatment functioning according to the Karnofsky performance status14 (KPS) and neurological functioning according to the Frankel classification15. The Eastern Cooperative Oncology Group (ECOG) performance status16 was obtained based on the KPS values. Time to development of SBM was defined as the interval from primary cancer diagnosis to the diagnosis of SBM4. If required for a specific scoring system, primary tumors were classified as slow, moderate or rapid growing5,6,11. All variables were further subdivided according to the instructions of each predictive model, resulting in different classifications for the same variable.
During the creation of the current database, visceral metastases were scored as either being present or not present. The models of Tomita and Tokuhashi models further distinguish visceral metastases as either being removable or not removable. In our analysis, if visceral metastases were present, they were scored as ‘not removable’. Also, the model of Rades scores the time to development of motor deficits. Retrospectively, this variable could not accurately be obtained from the medical records used to create this database. Therefore, the maximum number of points in the model was reduced from 25 to 21 and the minimum from 6 to 5. The cutoff scores for the predictive groups were changed accordingly. The Bollen model was created based on a large percentage of patients also included in this study. Therefore, an additional analysis based on external data only was conducted for this model.
Statistical analysis
Survival time was calculated as the time between start of treatment for the spinal metastasis and date of death or latest follow-up. Survival curves were estimated by using the Kaplan-Meier method and were compared using log-rank tests. Follow-up was assessed by employing the reverse Kaplan-Meier method17. Hazard ratios were estimated by using Cox proportional hazard models. All multivariate models were corrected for differences in survival between the participating centers. Harrell’s c-statistic was used to assess the predictive power of each model. It estimates the probability of concordance between predicted and observed responses, with a value of 1.0 being perfect agreement18. For each model, survival rates of the
84