2015 Marshall Urist Young Investigator Award: Prognostication in Patients With Long Bone Metastases: Does a Boosting Algorithm Improve Survival Estimates?

Background: Survival estimation guides surgical decision-making in metastatic bone disease. Traditionally, classic scoring systems, such as the Bauer score, provide survival estimates based on a summary score of prognostic factors. Identification of new factors might improve the accuracy of these models. Additionally, the use of different algorithms—nomograms or boosting algorithms—could further improve accuracy of prognostication relative to classic scoring systems. A nomogram is an extension of a classic scoring system and generates a more-individualized survival probability based on a patient’s set of characteristics using a figure. Boosting is a method that automatically trains to classify outcomes by applying classifiers (variables) in a sequential way and subsequently combines them. A boosting algorithm provides survival probabilities based on every possible combination of variables. Questions/purposes: We wished to (1) assess factors independently associated with decreased survival in patients with metastatic long bone fractures and (2) compare the accuracy of a classic scoring system, nomogram, and boosting algorithms in predicting 30-, 90-, and 365-day survival. Methods: We included all 927 patients in our retrospective study who underwent surgery for a metastatic long bone fracture at two institutions between January 1999 and December 2013. We included only the first procedure if patients underwent multiple surgical procedures or had more than one fracture. Median followup was 8 months (interquartile range, 3-25 months); 369 of 412 (90%) patients who where alive at 1 year were still in followup. Multivariable Cox regression analysis was used to identify clinical and laboratory factors independently associated with decreased survival. We created a classic scoring system, nomogram, and boosting algorithms based on identified variables. Accuracy of the algorithms was assessed using area under the curve analysis through fivefold cross validation. Results: The following factors were associated with a decreased likelihood of survival after surgical treatment of a metastatic long bone fracture, after controlling for relevant confounding variables: older age (hazard ratio [HR], 1.0; 95% CI, 1.0–1.0; p < 0.001), additional comorbidity (HR, 1.2; 95% CI, 1.0–1.4; p = 0.034), BMI less than 18.5 kg/m ² (HR, 2.0; 95% CI, 1.2–3.5; p = 0.011), tumor type with poor prognosis (HR, 1.8; 95% CI, 1.6–2.2; p < 0.001), multiple bone metastases (HR, 1.3; 95% CI, 1.1–1.6; p = 0.008), visceral metastases (HR, 1.6; 95% CI, 1.4–1.9; p < 0.001), and lower hemoglobin level (HR, 0.91; 95% CI, 0.87–0.96; p < 0.001). The survival estimates by the nomogram were moderately accurate for predicting 30-day (area under the curve [AUC], 0.72), 90-day (AUC, 0.75), and 365-day (AUC, 0.73) survival and remained stable after correcting for optimism through fivefold cross validation. Boosting algorithms were better predictors of survival on the training datasets, but decreased to a performance level comparable to the nomogram when applied on testing datasets for 30-day (AUC, 0.69), 90-day (AUC, 0.75), and 365-day (AUC, 0.72) survival prediction. Performance of the classic scoring system was lowest for all prediction periods. Conclusions: Comorbidity status and BMI are newly identified factors associated with decreased survival and should be taken into account when estimating survival. Performance of the boosting algorithms and nomogram were comparable on the testing datasets. However, the nomogram is easier to apply and therefore more useful to aid surgical decision making in clinical practice. Level of Evidence: Level III, prognostic study.