Machine learning-based models for outcome prediction in skull base and spinal chordomas: a systematic review and meta-analysis.

Abstract

Background: Chordomas are primary bone lesions originating from embryonic notochord remnants, most commonly developing along the skull base and spine. Managing chordomas is challenging due to the complex surgical approaches and significant resistance to chemotherapy and radiation. Consequently, the prognosis for chordoma treatment is unfavorable. We aimed to systematically assess the outcomes of machine learning (ML) models in predicting progression, recurrence, and survival in chordoma patients.

Methods: We conducted a systematic search on January 28, 2025, in PubMed, Embase, Scopus, and Web of Science. ML models that forecast skull base and spinal chordomas and report concordance index (C-index), area under the curve (AUC), accuracy (ACC), sensitivity, or specificity were included. A random-effects meta-analysis was performed using R with the "meta" and "mada" packages. The risk of bias (RoB) was assessed using the QUADAS-2 tool.

Results: Fifteen studies involving 3525 chordomas were included. The meta-analysis exhibited a pooled C-index of 0.81 (0.79-0.83), an AUC of 0.86 (95% CI: 0.83-0.9), and an ACC of 0.8 (95% CI: 0.75-0.85). The meta-analysis showed a pooled sensitivity of 0.74 (95% CI: 0.71-0.77), a specificity of 0.78 (95% CI: 0.74-0.81), and a diagnostic odds ratio (DOR) of 12.1 (95% CI: 7.1-20.6).

Conclusion: Our results indicated that ML models demonstrated robust predictive performance across various outcomes in chordomas, with pooled C-index, AUC, and ACC values ranging from 0.80 to 0.86 in relation to models that forecast progression, recurrence, or survival.

Clinical trial number: Not applicable.