Abstract
BACKGROUND: Spindle cell variants of squamous cell carcinoma represent rare and aggressive malignancies with poorly understood prognostic factors and treatment outcomes. This study leverages machine learning approaches alongside traditional statistical methods to analyze a large cohort from the Surveillance, Epidemiology, and End Results (SEER) database. METHODS: We conducted a retrospective analysis of 1086 patients with spindle cell variants of squamous cell carcinoma from the SEER database. Traditional Cox regression and machine learning approaches, including random survival forests (RSF), gradient boosted survival (GBSurv), and DeepSurv models, were employed to identify prognostic factors and predict survival outcomes. Model performance was evaluated using concordance indices and decision curve analysis. RESULTS: Of the 1086 patients included in the analysis, patients were diagnosed with spindle cell variants of squamous cell carcinoma between 1992 and 2021. Median age was 70 years (IQR: 60-77). Primary tumor sites included larynx (21.4%), lung and bronchus (20.2%), and tongue (10.7%). 32.0% had localized disease, 25.4% had regional disease, and 17.2% had distant disease, with 25.4% having unknown stage. Treatment modalities included radiation therapy in 51.9% and chemotherapy in 30.7% of patients. In the multivariate Cox model, kidney and renal pelvis tumors showed the highest risk (HR: 6.28, 95% CI: 2.26-17.45, P < 0.001), followed by urinary bladder (HR: 2.72, 95% CI: 1.56-4.74, P < 0.001) and lung/bronchus sites (HR: 1.94, 95% CI: 1.51-2.50, P < 0.001). The RSF model demonstrated superior discriminative ability (C-index: 0.733, 95% CI: 0.680-0.784) compared to GBSurv (C-index: 0.294, 95% CI: 0.239-0.351) and DeepSurv (C-index: 0.314, 95% CI: 0.255-0.378) approaches. CONCLUSIONS: The findings suggest that anatomical site and disease stage significantly influence survival, while current treatment modalities show limited impact on outcomes. The superior performance of the RSF model indicates potential value in using machine learning for risk stratification in clinical practice.