Trustworthy artificial intelligence in predictive medicine: cancer survival analysis using ethical-by-design and explainable artificial intelligence models.

Objectives: This study aims to develop and evaluate a trustworthy and ethical-by-design machine learning (ML) framework for predicting 5-year cancer survival across 10 Surveillance, Epidemiology, and End Results (SEER) cancer types. We assessed ML model(s) performance across localized, regional, and distant stages while examining ML fairness, ML explainability, and the added value of social determinants of health (SDOH) as features. The goal is to advance clinically interpretable and equity-centered survival prediction models suitable for translational use in health systems. Materials and Methods: Utilizing SEER Research Plus data, we built stage-stratified cohorts and defined a binary 5-year survival outcome. Four ML models were trained as follows: (1) base clinical models and (2) fairness-enhanced models incorporating SDOH features. Performance was evaluated using AUC and accuracy. Machine learning model(s) trustworthiness was assessed through SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations and fairness auditing via Equalized Odds Difference across different subgroups. Results: XGBoost showed the strongest discrimination across cancer types, and adding SDOH features generally maintained or modestly improved performance. Fairness evaluations revealed minimal to moderate bias across most SDOH strata, with larger disparities in specific cancer-stage cases. Machine learning explainability analysis demonstrated that clinical variables remained the primary predictors, while SDOH features added interpretable contextual risk without displacing core prognostic factors. Conclusions: This study demonstrates that trustworthy and ethical-by-design artificial intelligence for cancer survival prediction is achievable through a unified, explainable, and fairness-aware ML framework. Incorporating SDOH features enhances equity insights without diminishing predictive performance. The proposed framework offers a clinically meaningful and ethically aligned approach for precision medicine and population health applications.