Prognostic scores for survival as decisional support for surgery in spinal metastases: a performance assessment systematic review.

PURPOSE

To review the evidence on the relative prognostic performance of the available prognostic scores for survival in spinal metastatic surgery in order to provide a recommendation for use in clinical practice.

METHODS

A systematic review of comparative external validation studies assessing the performance of prognostic scores for survival in independent cohorts was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines. Eligible studies were identified through Medline and Embase until May 2021. Studies were included when they compared at least four survival scoring systems in surgical or mixed cohorts across all primary tumor types. Predictive performance was assessed based on discrimination and calibration for 3-month, 1-year and overall survival, and generalizability was assessed based on the characteristics of the development cohort and external validation cohorts. Risk of bias and concern regarding applicability were assessed based on the 'Prediction model study Risk Of Bias Assessment Tool' (PROBAST).

RESULTS

Twelve studies fulfilled the inclusion criteria and covered 17 scoring systems across 5.130 patients. Several scores suffer from suboptimal development and validation. The SORG Nomogram, developed in a large surgical cohort, showed good discrimination on 3-month and 1-year survival, good calibration and was superior in direct comparison with low risk of bias and low concern regarding applicability. Machine learning algorithms are promising as they perform equally well in direct comparison. Tokuhashi, Tomita and other traditional risk scores showed suboptimal performance.

CONCLUSION

The SORG Nomogram and machine learning algorithms outline superior performance in survival prediction for surgery in spinal metastases. Further improvement by comparative validation in large multicenter, prospective cohorts can still be obtained. Given the heterogeneity of spinal metastases, superior methodology of development and validation is key in improving future machine learning systems.