International Journal of Science, Engineering and TechnologyAuthors: Suryansh Anand, Mandeep, Dr. Vikalpa Tyagi
Abstract: High-volume clinical environments and remote patient monitoring pipelines are increasingly constrained by high cognitive loads, staffing shortages, and data saturation. While automated decision-making systems promise to optimize workflow efficiency, monolithic architectures introduce acute risks of automation bias, "explainability crises," and catastrophic undertriage of time-critical conditions. This paper introduces AegisTriage, a scalable, multi-agent autonomous artificial intelligence (AI) framework developed to dynamically assess patient acuity while integrating continuous, meaningful human oversight. Built using a distributed, tool-enabled Model Context Protocol (MCP), AegisTriage decomposes complex patient encounters into specialized sub-tasks managed by independent clinical agents: an Intake & History Agent, a Multimodal Context Synthesis Agent, and an Algorithmic Prioritization Agent. Rather than replacing the human clinician, the system operates as an intellectual filter that structures complex data, generates explicit clinical reasoning paths, and maintains human-in-the-loop (HITL) gates for high-acuity classifications. We evaluate the system using a retrospective cohort of simulated and real-world clinical vignettes against a human majority-vote gold standard. AegisTriage achieved a 95.8% sensitivity for emergency classifications and an 88.5% sensitivity for overall actionable alerts, operating at an average processing cost of $0.34 per triage encounter. Crucially, the integration of explainable clinical reasoning and structured override protocols significantly lowered automation bias in human operators during simulated stress testing. These findings present a scalable, legally compliant pathway for augmenting high-stakes clinical and remote triage without sacrificing professional accountability or patient safety.
