Clinical Artificial Intelligence Agents in Nephrology: From Prediction to Action Through Workflow-Native Intelligence—A Roadmap for Workflow-Integrated Care

Abstract

Background: Artificial intelligence in nephrology has largely focused on predictive models for outcomes such as acute kidney injury (AKI), chronic kidney disease (CKD) progression, and transplant complications. Although these models demonstrate technical performance, their real-world clinical impact has remained limited because prediction alone rarely translates into coordinated clinical action. Clinical artificial intelligence agents represent workflow-native systems that operate in real time, interact bidirectionally with clinical environments, adapt to evolving patient and workflow states, and support coordinated clinical action rather than generating isolated predictions. This review proposes clinical artificial intelligence agents as a new paradigm for integrating artificial intelligence directly into nephrology workflows. Methods: We conducted a narrative synthesis of emerging literature on artificial intelligence systems, agentic artificial intelligence architectures, clinical decision support, and digital health infrastructures relevant to kidney care. Drawing from interdisciplinary sources in medicine, health informatics, and artificial intelligence research, we developed a conceptual framework describing the architecture, governance requirements, and evaluation principles of clinical artificial intelligence agents in nephrology. Results: Clinical artificial intelligence agents represent workflow-integrated systems capable of continuously perceiving patient data, reasoning under clinical constraints, planning tasks, and supporting coordinated clinical actions over time. We describe a layered architecture consisting of perception, cognition, planning and control, action, and learning components. Potential applications span the nephrology care continuum, including CKD management, AKI monitoring, dialysis and continuous renal replacement therapy (CRRT) optimization, kidney transplantation care coordination, glomerulonephritis management, and supervised patient-facing systems. Conclusions: Clinical artificial intelligence agents shift the role of artificial intelligence from isolated prediction toward longitudinal clinical orchestration. Future evaluation should prioritize workflow integration, time-to-action, clinician oversight, safety, and patient-centered outcomes rather than relying solely on traditional model performance metrics. This roadmap provides a conceptual foundation for the responsible development and clinical integration of agentic artificial intelligence systems in nephrology.

Ähnliche Arbeiten