Machine Learning-Based Recurrence and Survival Prediction in a Large Institutional Upper Extremity Soft Tissue Sarcoma Cohort.

BACKGROUND AND OBJECTIVES: Upper extremity soft tissue sarcomas (UE STS) present unique management challenges; existing prognostic nomograms were developed from mixed anatomic cohorts. This study develops anatomically specific machine learning (ML) models to predict UE-STS outcomes and benchmarks them against established tools.

METHODS: Time-to-event models for local recurrence (LR), distant metastasis (DM), and overall survival (OS) were developed in 188 patients treated at a single sarcoma center, treating death as a competing risk. Five survival algorithms were compared within a nested cross-validation framework. Discrimination was benchmarked against the Sarculator, PERSARC, and MSKCC nomograms.

RESULTS: Five-year event rates were 24% (LR), 20% (DM), and 13% (death) over a median follow-up of 3.9 years. Best-performing models achieved competing-risks concordance indices of 0.79 (LR), 0.67 (DM), and 0.75 (OS). Parsimonious models performed comparably. Discrimination was similar to Sarculator and PERSARC and higher than the MSKCC nomogram for LR (C-index 0.90 vs. 0.68). Predicted-risk groups separated clearly on cumulative-incidence and survival curves (p < 0.001).

CONCLUSIONS: These results demonstrate the feasibility and potential value of anatomy-specific, time-to-event risk modeling in UE STS, achieving discrimination comparable to established nomograms despite a smaller cohort. Multicenter cohorts and multidimensional data may strengthen future efforts. External validation is warranted.