Recent research has introduced autonomous artificial intelligence agents, specifically MIRA and AMIE, designed to integrate into clinical workflows and manage complex patient cases.

最近的研究介紹了自主人工智慧代理，特別是 MIRA 與 AMIE，旨在整合至臨床工作流程並管理複雜的患者案例。

The development of MIRA (Medical Intelligence for Reasoning and Action) represents a shift from task-specific chat tools toward systems integrated within Electronic Health Records (EHR). Operating in a sandboxed environment and adhering to FHIR standards, MIRA utilizes a suite of tools to execute a comprehensive clinical sequence: history acquisition via a patient-simulation agent, the ordering and interpretation of diagnostic tests, and the formulation of treatment plans. In simulations utilizing the MIMIC-IV dataset, MIRA demonstrated diagnostic accuracy (87.8%) that exceeded that of both board-certified physicians (78.1%) and mixed-seniority cohorts (71.1%). Furthermore, MIRA exhibited superior alignment with clinical guidelines and higher recall in identifying necessary surgical procedures compared to human practitioners.

MIRA（醫療推理與行動智能）的開發代表了從特定任務聊天工具轉向整合在電子健康紀錄 (EHR) 中的系統。MIRA 在沙盒環境中運行並遵循 FHIR 標準，利用一套工具來執行全面的臨床流程：透過患者模擬代理獲取病史、開立與解讀診斷測試，以及制定治療計劃。在利用 MIMIC-IV 數據集的模擬中，MIRA 展現出的診斷準確度 (87.8%) 超過了執業醫師 (78.1%) 與不同資歷之醫師組別 (71.1%)。

Parallel advancements are evidenced by the evolution of the Articulate Medical Intelligence Explorer (AMIE). While previous iterations focused on diagnostic dialogue, the current system is optimized for multi-visit clinical management. By leveraging long-context retrieval and structured reasoning, AMIE aligns its outputs with authoritative drug formularies and clinical practice guidelines. In a blinded virtual Objective Structured Clinical Examination (OSCE) involving 100 scenarios, AMIE was found to be non-inferior to primary care physicians in management reasoning. Additionally, in the RxQA benchmark—a validated multiple-choice assessment of medication reasoning—AMIE outperformed human physicians on high-difficulty queries.

平行進步亦體現在 Articulate Medical Intelligence Explorer (AMIE) 的演進中。雖然之前的版本專注於診斷對話，但目前的系統已針對多次就診的臨床管理進行優化。透過利用長文本檢索與結構化推理，AMIE 將其輸出與權威藥典及臨床實踐指南對齊。在一項包含 100 個場景的雙盲虛擬客觀結構化臨床考試 (OSCE) 中，發現 AMIE 在管理推理方面不劣於基層醫療醫師。此外，在 RxQA 基準測試（一項經過驗證的藥物推理多選評估）中，AMIE 在高難度查詢上的表現優於人類醫師。

Despite these performance metrics, institutional integration remains contingent upon resolving specific limitations. MIRA's performance, while robust, did not achieve absolute reliability, and the patient-simulation agents used in testing may provide more structured responses than actual clinical encounters. Consequently, the researchers posit that the optimal deployment of such agents would be as collaborative 'copilots' rather than replacements, focusing on high-volume, documentation-intensive tasks under human supervision to mitigate risk and enhance resource stewardship.

儘管有這些性能指標，機構整合仍取決於特定局限性的解決。MIRA 的表現雖然強健，但尚未達到絕對可靠，且測試中使用的患者模擬代理所提供的回應可能比實際臨床就診更具結構性。因此，研究人員認為，此類代理的最佳部署方式應是作為協作的「副駕駛 (copilots)」而非替代品，專注於在人類監督下處理高容量、文書密集型任務，以降低風險並強化資源管理。

MIRA and AMIE have demonstrated the capacity to perform at or above physician levels in simulated diagnostic and management tasks, though prospective real-world validation is required.

MIRA 與 AMIE 已證明在模擬診斷與管理任務中，表現能達到或超過醫師水平，但仍需要前瞻性的現實世界驗證。