Medical AI Agent MIRA Outperforms Doctors in 87.8% of EHR Simulations: How Autonomous AI is Changing Diagnostics

Doctors spend hours on routine tasks: collecting medical history, ordering tests, synthesizing results from various systems. What if an AI agent could take on most of this work? A study published in Nature showed that the autonomous medical AI agent MIRA outperformed doctors in 87.8% of simulated clinical scenarios, working with Electronic Health Record (EHR) data and achieving 88.9% diagnostic accuracy.

In modern medicine, doctors are overwhelmed with administrative work: searching for information in EHRs, reconciling data, ordering tests. This consumes valuable time that could be dedicated to patients and increases the risk of errors due to human factors. Such routine not only reduces efficiency but also leads to specialist burnout. However, there is a way to alleviate this burden using AI agents.

The Gap Between Knowledge and Action: A Challenge in Modern Medicine

Large Language Models (LLMs) have long demonstrated impressive results in medical exams and are capable of answering complex clinical questions. They can cite thousands of articles, find non-obvious connections, and formulate hypotheses. However, a significant gap existed between this "knowledge" and its real-world application in a clinic's operational processes. Traditional medical AI tools often limited themselves to the role of search engines or text generators, lacking the ability to actively participate in the complex, multi-stage process of clinical decision-making.

True diagnosis and treatment planning are not just about information retrieval. They involve gathering detailed patient history, ordering necessary tests, synthesizing sometimes contradictory data, continuously updating hypotheses, and adjusting the course of treatment. All of this occurs within Electronic Health Record (EHR) systems, which require strict adherence to coding protocols and complex navigation. Doctors spent an enormous amount of time manually processing this data, which slowed down the process and increased the likelihood of errors.

The Path to an Autonomous AI Agent: Why Existing Solutions Didn't Work

Existing automation solutions in medicine were typically highly specialized. For example, one system might help with X-ray interpretation, another with blood test analysis, and a third with drug information retrieval. But none could integrate these functions into a single, continuous process that mimicked a doctor's thinking and actions. Each solution required manual data entry or switching between interfaces, negating some of the benefits of automation.

The medical community needed not just a tool, but a full-fledged AI agent capable of autonomously navigating the EHR environment, making data-driven decisions, ordering tests, and formulating diagnoses and treatment plans. This meant transitioning from a passive "assistant" to an active "participant" in the process, one that not only provides information but also acts like a doctor, albeit in a controlled environment.

How the MIRA AI Agent Was Designed

Researchers developed MIRA as an autonomous AI agent capable of operating within isolated EHR environments. The main idea was for MIRA not just to process data, but to "live" within the system, interacting with it autonomously. The agent was equipped with 11 specialized digital tools and had access to over 85,000 operational choices, allowing it to mimic a wide range of clinical actions.

MIRA's functionality included:

• Medical History Analysis. The agent autonomously reviewed all available data in the EHR.
• Diagnostic Test Ordering. MIRA could request physical examinations and order targeted lab tests based on the current clinical picture.
• Formulating Diagnoses and Treatment Plans. Based on the collected data, the agent proposed the most likely diagnoses and corresponding treatment plans.
• Generating Medication Prescriptions. In the simulated EHR environment, MIRA could write prescriptions, taking into account patient specifics.

A key aspect of the design was to ensure MIRA's autonomy in decision-making while retaining the ability for human oversight and correction in critical situations.

Implementation and Testing in a Simulated Environment

MIRA was not directly implemented into real clinical practice. Instead, to evaluate its effectiveness and safety, a controlled simulation was created using 574 real clinical cases from the MIMIC-IV database. This allowed for a comparison of the AI agent's performance with that of experienced doctors under identical conditions, eliminating the influence of external factors.

Testing phases included:

• Training and Calibration. MIRA was trained on a vast amount of medical data to accurately mimic a doctor's decision-making process.
• Comparative Analysis. In 311 cases, MIRA was directly compared against a group of certified doctors and mixed-level teams (residents + certified doctors).
• Safety Assessment. Independent medical experts reviewed MIRA's outcomes, prescribed medications, and decisions to ensure no risks to patients.

This approach allowed for a thorough analysis of the agent's performance before its potential application in real-world settings, ensuring maximum safety and reliability.

Study Results: MIRA Outperforms Doctors

The study yielded impressive results. MIRA achieved 88.9% diagnostic accuracy across all 574 MIMIC-IV cases. In direct comparison with the group of doctors, MIRA's accuracy was 87.8%, significantly higher than that of experienced human specialists under the same simulated conditions.

MIRA performed particularly well with diagnoses such as appendicitis and pancreatitis, achieving 100% completeness of detection for laparoscopic appendectomies. Importantly, the AI agent did not resort to excessive test ordering; its choices remained below historical baseline levels, indicating high efficiency and cost-effectiveness.

Safety assessments were also encouraging: an independent medical review of 56 patient-level outcomes and 468 prescriptions written by MIRA showed no high-intensity drug interactions, renal dosing incompatibilities, or discrepancies between medications and allergies. The agent also achieved a perfect completeness of detection score (1.00) in critical hospitalization decisions.

How to Implement This in Your Practice: Prospects for Medical Institutions

The MIRA study results demonstrate the enormous potential of AI agents to transform healthcare. If your clinic faces physician overload with routine tasks, lengthy information retrieval from EHRs, and the need to optimize the diagnostic process, then AI agents could be the solution. Here's where to start:

• Optimize Workflow. AI agents can take on routine EHR data processing tasks, freeing up doctors for more complex cases and direct patient interaction. For example, automating patient history collection, preliminary diagnosis preparation, and examination plans.
• Improve Diagnostic Accuracy. As the MIRA case showed, AI can surpass human diagnostic accuracy in specific areas, reducing errors and improving treatment outcomes.
• Reduce Costs. Efficient use of resources, such as tests and imaging, without excessive ordering, can lead to significant savings for both the clinic and the patient.
• Decision Support. AI can act as a powerful support tool for doctors, offering evidence-based diagnoses and treatment plans based on the analysis of thousands of cases and the latest medical research.

Despite the impressive results, the study authors emphasize that MIRA and similar AI agents do not replace expert human staff. They require constant human oversight and patient-level safeguards. However, their potential to transform healthcare is immense.

If this case sounds like what's happening in your company, our manager can help: he'll analyze your business and niche for free and point out where an AI agent would bring a real result in your case. Message the manager