Human Versus Machine in Survival Prediction for Metastatic Spinal Cord Compression: A Retrospective Cohort Study

Accurate survival prediction in metastatic spinal cord compression (MSCC) is critical for guiding treatment decisions, yet remains challenging, particularly for intermediate survival durations. We compared the accuracy of oncologist judgment, surgeon-calculated Tokuhashi scores, and ChatGPT-assisted predictions in estimating survival outcomes in MSCC patients.

This retrospective study included 100 patients (n = 100) referred to the Centre for Spinal Studies and Surgery, Queen's Medical Centre, a tertiary spinal oncology center in Nottingham, United Kingdom, with radiologically confirmed MSCC. Anonymized clinical data were used to calculate surgeon Tokuhashi scores, document oncologist-estimated life expectancy, and generate ChatGPT-assisted survival predictions based on both literature review and Tokuhashi calculation. Predictions were compared against actual survival outcomes (<6 months, six to 12 months, >12 months). Machine learning analyses identified key predictors of survival.

Overall prediction accuracy was 53% for ChatGPT Tokuhashi-based predictions, 49% for surgeon Tokuhashi scores, 47% for oncologist judgment, and 36% for ChatGPT literature-based estimates. Recall for short survival (<6 months) was the highest with the surgeon (70%) and ChatGPT Tokuhashi (68%) methods, whereas intermediate survival (six to 12 months) remained difficult to predict across all modalities. For long-term survival (>12 months), oncologists performed better (74% recall). Functional status (Karnofsky score) and patient age emerged as the strongest survival predictors across logistic regression, random forest, decision tree, and XGBoost models, surpassing primary tumor type and metastasis burden.

Structured prognostic tools and AI-assisted scoring can complement clinical judgment in predicting short-term survival in MSCC. However, intermediate-term survival prediction remains a critical unmet need. Future prognostic strategies should prioritize dynamic functional metrics over static tumor classifications to improve personalized decision-making.