Applied Artificial Intelligence Across High-Stakes Domains: A Systematic Review of Methodological Foundations, Trustworthiness Pillars, and Cross-Domain Deployment Patterns

Abstract

Artificial intelligence (AI) and machine learning (ML) have moved decisively from controlled laboratory benchmarks into operational, consequential decision systems spanning finance, security, healthcare, sustainability, autonomous systems, and a widening range of human-centered services. This breadth has produced a fragmented literature in which closely related methodological advances are repeatedly rediscovered under domain-specific vocabularies, while genuinely cross-cutting concerns of trustworthiness are addressed unevenly. This paper presents a systematic review of 196 recent studies (2023-2026) that organizes the field along three orthogonal axes: methodological foundations (graph and relational learning, deep sequence and temporal models, multimodal and sensor fusion, large language models and agents, ensemble and classical learning, and reinforcement and optimization learning); trustworthiness pillars (privacy-preserving learning, fairness, explainability, and robustness and security); and seven application domains. We synthesize how methods migrate across domains, where trustworthiness requirements are satisfied or neglected, and how deployment patterns differ under the constraints of each setting. A quantitative coverage analysis shows that financial risk and RegTech (24%) and healthcare and biomedicine (22%) dominate the corpus, that robustness and security are addressed far more frequently than fairness, and that privacy-preserving learning and explainability are concentrated in regulated domains. We discuss recurrent design tensions-accuracy versus interpretability, utility versus privacy, and centralized performance versus federated governance-and identify open challenges, including the trustworthiness of autonomous multi-agent systems, evaluation under distribution shift, and the uneven treatment of fairness outside credit scoring. The review is intended as a structured map for practitioners selecting methods under real deployment constraints and for researchers locating durable gaps.