In complex operational environments, hybrid decision-making frameworks offer a means to integrate human expertise - characterized by contextual sensitivity, adaptability, and experiential knowledge - with the objective, standardized precision of machine-based systems.MethodThis study develops a decision-support structure by comparing supervised Machine Learning (ML) models; random forest (RF), support vector machine (SVM) and a bidirectional encoder representation from transformer-based (KB/BERT) model - using hierarchical and flat approaches against a manual classification process, involving more than 200 train delay codes across 10 days. ML models are trained on same-day delay data and evaluated against the outcomes from a multi-actor decision process.ResultsHierarchical models outperform flat ones, achieving near-human assessors on basic level coding (Level 1 and 2), though with greater variability (mean F1-scores (50-91 per cent)), compared to manual classification (mean F1-scores (87-98 per cent)) at the most granular level (Level 3) of prediction. "Simpler" models also outperform the more complex KB/BERT.Practical ImplicationsWe discuss the functionality and accuracy of ML-based hybrid decision-support systems (HDSS), noting the need for trade-offs between precision and accuracy. ML models demonstrate potential to complement - not replace - human expertise, particularly with uncertainty estimation tools that mitigate classification risks and support decision-making. We conclude with implications for data representation in the design of HDSS within socio-techno-economic contexts.