DeepOSSA: a deep learning framework for personalized survival prediction in osteosarcoma patients

Osteosarcoma, the most common primary malignant bone tumor in children and adolescents, is marked by high invasiveness and heterogeneous outcomes due to inter-patient variability in tumor biology and treatment response. Accurate prognostic modeling is crucial for personalized risk stratification and clinical decision-making.

We developed and validated DeepOSSA, a deep learning framework for individualized survival prediction in osteosarcoma, leveraging high-dimensional, sparsely encoded clinical data from the Surveillance, Epidemiology, and End Results (SEER) database.

Clinical data from 895 osteosarcoma patients diagnosed between 2010 and 2018 were extracted. Continuous variables were optimally discretized using X-tile, and categorical covariates were one-hot encoded. We compared DeepOSSA against seven representative models spanning classical machine learning (support vector machine (SVM), random forest (RF), and gradient boosted decision trees (GBDT)), deep learning (multi-layer perceptron (MLP) and Transformer), and survival analysis (Cox proportional hazards (CoxPH) and DeepSurv). DeepOSSA integrates a Transformer-based representation tower for high-order sparse feature crossing with a Cox-style discrete-time survival head, enabling joint optimization of feature representation and time-to-event modeling. Performance was evaluated on a held-out test set using the area under the curve (AUC) and F1-score.

DeepOSSA outperformed all baselines in discrimination and balanced classification, achieving an AUC of $0.8058\pm 0.0081$ and an F1-score of $0.6506\pm 0.0060$ , representing average relative improvements of 13.11% (AUC) and 8.99% (F1), respectively. Among all evaluated models, DeepOSSA achieved the highest AUC (exceeding 0.8) and demonstrated competitive performance compared with previously published, externally validated Cox nomograms for osteosarcoma. In time-dependent evaluation, its predictive advantage over CoxPH increased during long-term follow-up, where proportional hazards assumptions are most violated. Ablation studies confirmed the necessity of each architectural component, and SHapley Additive exPlanations (SHAP) analysis demonstrated that the model's risk predictions are driven by clinically established prognostic factors. By capturing attention-based nonlinear interactions and temporal dynamics, DeepOSSA enables more precise patient stratification than traditional linear models.

DeepOSSA is a robust and interpretable framework that bridges Transformer-based representation learning with principled survival analysis. It effectively models complex, nonlinear relationships among sparse clinical features and survival outcomes, offering strong potential for translating heterogeneous clinical data into actionable insights for precision oncology.