Abstract
OBJECTIVES: To evaluate the performance of the Tokuhashi, Tomita, SORG machine learning (SORG ML), and OPTImodel algorithms as survival predictors for vertebral metastases in clinical practice. MATERIALS AND METHODS: A retrospective study (2013-2023) analyzed 573 patients from Cabueñes University Hospital (Asturias, Spain). Thirty-two demographic, epidemiological, clinical, and analytical variables were considered, including diagnosis chronology and survival. RESULTS: Among the 573 patients studied, 272 (47.4%) presented visceral metastases at the time of diagnosis. A total of 362 patients (63.2%) had associated comorbidities. The most frequent primary histological diagnoses in these patients were lung 147 (25.7%), prostate 146 (25.5%), breast 118 (20.6%), kidney 30 (5.2%), and colorectal 29 (5.1%). The median survival of the cohort was 185 days. The accuracy rates for the Tokuhashi, SORG ML, OPTImodel, and Tomita algorithms were 0.5509, 0.4812, 0.3404, and 0.3858, respectively. The models with the highest accuracy rates in specific time segments were Tokuhashi (77.5% for < 6 months) and OPTImodel (90.8% for more than 1 year). The areas under the curve (AUC) for survival intervals were as follows: Tokuhashi at 42 days (73.19%), 90 days (79.3%), and 365 days (82.73%); Tomita at 42 days (69.27%), 90 days (76.82%), and 365 days (78.79%); SORG ML at 42 days (52.77%), 90 days (51.69%), and 365 days (51.38%). CONCLUSIONS: All models showed relatively low accuracy. The newer models (OPTImodel, SORG ML) did not outperform the traditional Tomita and Tokuhashi in predicting survival for vertebral metastases patients.