A data-driven system-theoretic Bayesian network framework for probabilistic safety assessment of passenger vessels

• STPA-based Bayesian Network mitigates subjective bias in maritime risk analysis • Novel dataset of 235 official reports enables data-driven causal modelling • Identifies Structural Failure & Defective Maintenance as dominant accident precursors • Reframes human error and COLREGs breaches as systemic, context-dependent • Framework enables proactive, data-driven passenger vessel safety governance Passenger vessel operations present a high-consequence environment where a paradox has emerged: incident frequency is decreasing, yet catastrophic severity is not. This trend exposes the inadequacy of existing risk models, which are typically localized, and reliant on subjective expert elicitation. This study develops a robust, data-driven risk assessment framework by synergizing System-Theoretic Process Analysis (STPA) with a Bayesian Network (BN), grounded in a novel database of 235 official European accident reports. STPA defines the BN’s causal topology, ensuring theoretical coherence and mitigating the epistemic uncertainty and bias of conventional expert-led modeling. Sensitivity analysis reveals the probabilistic primacy of latent systemic precursors, identifying Structural Failure and Defective Maintenance as dominant risk control points. The analysis moves beyond simplistic attributions of “human error”, revealing how operational failures like COLREGs infringements are symptoms of distinct causal pathways dependent on vessel type and operational conditions. The resulting model is a quantitative instrument that identifies the most probable pathways to catastrophe, offering an objective foundation for transitioning from reactive compliance to proactive, data-driven safety governance.