Maritime accident severity modeling with operationally realistic multi-factor data-driven bayesian networks

• Trains and benchmarks TAN, ANB, and expert BNs for severity forecasting • Integrates two-layer geography: named sea regions plus operating locale • Uses multi-state met ocean factors: visibility, wind, sea state, storms, ice • Adds PSC inspection outcomes as a compliance signal in severity modelling • Provides scenario, sensitivity, SoI, and TRI analyses for actionable levers Maritime accident severity analysis supports safety management and emergency preparedness, yet operational realism remains limited in many quantitative models. Common limitations include coarse geographic representation, reduced-granularity met-ocean descriptors, and the omission of Port State Control (PSC) compliance signals. These constraints can mask actionable thresholds and hinder the identification of interpretable, decision-relevant severity pathways. To address this, the study develops Bayesian Network (BN) models to forecast accident severity using a globally distributed, multi-year (2015-2024) dataset of 1050 reports, with granular descriptors of route/region, operating locale, weather, sea-state, visibility, ship profile, human error, and accident type. It assembles a harmonized severity dataset unifying investigation narratives, PSC outcomes, ship registries, and met-ocean validation within a single standardized schema; embeds two-layer geography, multi-state environmental conditions, and compliance signals into one BN severity framework; and benchmarks three BN families, Tree-Augmented Naïve (TAN), Augmented Naïve Bayes (ANB), and a literature/expert topology, on the same dataset. Beyond structure comparison, a unified inference workflow combines Strength-of-Influence, state-wise tornado decomposition, True Risk Influence (TRI), and sequential and simultaneous scenario analysis. All three models were estimated via Expectation–Maximization (EM) and evaluated using stratified 10-fold ROC/AUC. The literature/expert BN achieved the strongest cross-validated discrimination across severity states and was selected for analysis, indicating that structure should be selected empirically rather than assumed a priori. Diagnostic, scenario, and sensitivity analysis show that severity escalates when open-sea operations on dense traffic corridors coincide with adverse environment (storm, poor visibility, rough sea, high winds), human error, and loss-of-control typologies (grounding/capsizing); severity attenuates under benign conditions and contact-type events. Strength-of-Influence results highlight the central roles of route/region→traffic density and weather/visibility→severity pathways. A targeted TRI assessment isolates Accident Type as the dominant lever shifting probability mass into the high-severity state. The findings enable consequence-aware routing and vessel traffic prioritization, type-conditioned emergency preparedness, and human-factor/equipment readiness.