AI-driven MIRA Outperforms Doctors in Emergency Diagnostics

An autonomous AI agent just outperformed board-certified physicians at diagnosing emergency room patients, while across the medical AI landscape, a drug designed entirely by algorithms entered its final stage of human testing.

Jakob Kather’s team at TU Dresden built MIRA, an AI agent that worked through 574 real emergency department cases inside a sandboxed electronic health record system. It took patient histories, ordered labs and imaging, interpreted results, and wrote treatment plans. When matched against four board-certified physicians on 311 of those cases, MIRA hit 87.8% diagnostic accuracy. The doctors managed 78.1%. The study, conducted with Heidelberg University Hospital and published in Nature on June 17, represents a measurable shift in how AI engages with clinical workflows.

MIRA stands for Medical Intelligence for Reasoning and Action, and it had to do more than answer trivia questions or respond to multiple-choice prompts. It navigated what researchers call a clinical action space: requesting the right test, reading it correctly, deciding what the results meant, and acting on them under the same kind of time pressure a doctor faces during a shift. The system operated within a simulated but realistic EHR environment, choosing from more than 85,000 diagnostic and treatment options to manage each case from start to finish.

MIRA functions as an autonomous agent embedded in a clinical workflow. It doesn’t just suggest a diagnosis after reading a paragraph of symptoms. It conducts patient interviews, requests specific lab work or imaging studies, waits for those results, interprets them in context, and then decides on next steps. That might mean ordering additional tests, prescribing medication, scheduling a procedure, or arranging hospital admission.

The Dresden team tested MIRA on real cases pulled from emergency department records. These weren’t simplified teaching examples. They included incomplete information, ambiguous presentations, and the kind of diagnostic uncertainty that makes emergency medicine difficult. MIRA had to work sequentially, building its understanding of each case the same way a resident would during overnight coverage.

The system showed clear strengths and specific weaknesses. On appendicitis cases, MIRA achieved 98.6% accuracy. On pneumonia, it dropped to 72.4%, falling below its own average and revealing genuine blind spots. The variation matters because it shows where the system performs reliably and where human oversight becomes critical.

MIRA stayed text-only throughout testing. It didn’t conduct physical examinations, assess tone of voice, or interpret visual cues. The cases came from complete, clean datasets rather than the chaos of a real overnight shift with an incomplete chart and a patient struggling to explain symptoms. Those limitations are significant, but they don’t diminish what the study demonstrated about AI’s capacity to handle complex, sequential clinical decision-making.

MIRA, an autonomous AI agent tested on real emergency department cases in a sandboxed electronic health record, outperformed its physician comparison group while carrying out a complete diagnostic workflow rather than answering isolated questions. Alongside rentosertib, an AI-designed drug entering Phase III trials, the result signals that AI is moving into the operational machinery of medicine. Yet both examples also expose the limits of current evidence: MIRA has not treated live patients, rentosertib still must prove safety and efficacy, and regulators have clearer rules for drugs than for autonomous diagnostic agents.

Most AI health breakthroughs this year came from benchmarks: tidy multiple-choice tests where models can look brilliant without ever touching a real workflow. A system might ace a standardized exam but fail completely when asked to navigate the messy reality of patient care. MIRA represents a step past that.

The agent sat inside an actual EHR structure and made the sequential calls that hospital shifts are built from. It’s still simulation. Nobody let MIRA touch a live patient. But simulation built from real ED cases makes a meaningfully different claim than a benchmark built from exam questions, and that’s why this result traveled further than most.

Experts who reviewed the study emphasized MIRA’s potential as a clinical decision support tool integrated into existing hospital systems. They also stressed that despite promising results, the system requires extensive real-world testing before clinical deployment. MIRA demonstrated physician-level expertise in controlled environments, but translating that performance to live patient care introduces variables the study couldn’t capture.

The system utilized more lab tests than physicians typically would but avoided over-ordering costly imaging studies. In safety evaluations, MIRA produced zero high-severity drug interactions and no medication mismatches. That safety profile matters as much as diagnostic accuracy when considering future applications.

Three weeks after the MIRA paper landed, Insilico Medicine announced it had begun a Phase III trial for rentosertib, an oral TNIK inhibitor for idiopathic pulmonary fibrosis. IPF is a lung disease that scars tissue progressively until breathing becomes impossible. Rentosertib targets TNIK, a kinase involved in fibrotic and inflammatory pathways.

What sets rentosertib apart is its origin. Both the TNIK target and the molecule itself were generated by Insilico’s Pharma.AI platform. A chemist didn’t find it by screening a compound library. AI identified the biological target through analysis of genomic and proteomic data, then designed the molecule to interact with that target. According to Insilico, rentosertib is the first drug where both target and molecule came from generative AI to reach Phase III trials.

The trial itself is specific and checkable: a randomized, double-blind, placebo-controlled study enrolling 320 patients across 47 centers in China, measuring the rate of lung function decline over 52 weeks. China’s drug regulator granted rentosertib Breakthrough Therapy Designation in May 2025 after Phase IIa data showed patients on a 60 mg daily dose experienced improved lung function compared to placebo.

Phase IIa results demonstrated a dose-dependent improvement in forced vital capacity, with patients receiving the 60 mg dose showing a +98.4 mL increase at 12 weeks. The safety profile was manageable, though some liver toxicity was observed. The study involved exclusively Asian participants, so further confirmation across diverse populations remains necessary.

Insilico closed a $293 million Hong Kong IPO on December 30, 2025, the largest biotech listing on that exchange last year. Rentosertib is the asset most of that funding is riding on. The company’s platform integrates AI-powered target discovery, generative chemistry, and clinical outcome prediction into a single workflow. The platform identified TNIK as a novel target within months, a timeline that would have been impossible using traditional drug discovery methods.

The hard part is no longer proving that medical AI can outperform humans in a sandbox; it is deciding who is responsible when it fails in a hospital.

MIRA demonstrates that autonomous systems can navigate complex clinical workflows, but its controlled setting leaves the central deployment questions unanswered. Until liability, monitoring, and escalation rules are clear, higher accuracy alone is not a sufficient case for clinical autonomy.

Put MIRA and rentosertib side by side and something interesting emerges. One is AI making diagnostic calls in something close to real time. The other is AI generating a molecule years before any trial begins. Different timescales, different failure modes, but the same underlying shift: AI systems aren’t confined to answering questions about medicine anymore. They’re inside the actual machinery of it, ordering tests and proposing molecules that then have to survive contact with real patients.

The regulatory apparatus isn’t remotely ready for the first of those two scenarios. If an AI agent orders the wrong test or misses a diagnosis inside a live hospital system, nobody has settled who’s liable: the hospital, the vendor, or the physician who signed off. The FDA has cleared plenty of narrow AI tools. Imaging software that flags a tumor. Algorithms that read an EKG. But nothing that autonomously drives a full diagnostic workup the way MIRA did in Dresden’s sandbox.

Kather’s team was careful to say what every serious researcher in this space says: this isn’t a replacement for doctors, and none of it has been tested on a live patient yet. That caution reflects the regulatory vacuum around autonomous diagnostic AI.

Insilico’s trial, by contrast, sits inside a regulatory pathway that already exists. A drug is a drug, however it was designed, and it still has to clear the same 52-week endpoint everyone else’s does. The molecule has to prove safety and efficacy in humans. The origin story doesn’t change the regulatory requirements.

That difference matters. Drug development has a rulebook that’s been refined over decades. AI-driven diagnostic systems that operate autonomously within clinical workflows don’t have that rulebook yet. States like California, Illinois, Nevada, and Texas have started enacting laws that regulate AI in healthcare, particularly around mental health applications, disclosure requirements, and prohibitions on misleading claims. California prohibits AI systems from implying licensed medical oversight without it, with direct enforcement by licensing boards. Illinois and Nevada restrict AI from making independent therapeutic decisions without licensed oversight. Texas requires disclosure of AI use in diagnosis and mandates review of AI-generated medical records.

These state-level regulations create a patchwork of compliance challenges. Medical AI handling medium to high risk requires either internal self-regulation or external governmental regulation. Most regulation has been internal so far, leading to significant cost and variability among hospitals. Guidelines from organizations like the Joint Commission suggest ongoing quality monitoring and informed patient consent for AI use, but those compliance burdens might be prohibitively expensive for small and community hospitals.

That creates a real risk of unequal healthcare access. If regulatory burdens limit AI adoption to large, well-resourced hospital systems, it exacerbates disparities across geographies and patient populations. The UK’s National Commission into the Regulation of AI in Healthcare is collecting evidence to shape comprehensive regulations prioritizing patient safety and clear responsibility management. The commission includes experts across healthcare, technology, law, and patient advocacy to address these challenges.

MIRA’s performance demonstrates AI’s capacity to handle the kind of sequential, context-dependent reasoning that defines clinical work. The system didn’t just match physician performance. It exceeded it by nearly 10 percentage points in a head-to-head comparison. That gap is large enough to matter in real-world applications where diagnostic accuracy directly affects patient outcomes.

The system’s safety profile adds weight to the argument for AI-assisted clinical workflows. Zero high-severity drug interactions and no medication mismatches in safety evaluations suggest MIRA can navigate complex pharmaceutical decision trees without introducing new risks. The agent’s tendency to order more lab tests than physicians typically would raises questions about cost and efficiency, but avoiding over-ordering of expensive imaging studies balances that concern.

The variation in MIRA’s performance across different conditions points to a practical framework for deployment. The system excels at diagnosing appendicitis but struggles with pneumonia. That pattern suggests a path forward where AI handles specific diagnostic categories where it demonstrates consistent accuracy while human oversight remains primary for conditions where the system shows weaker performance.

Experts reviewing the study emphasized that MIRA represents a significant step toward a true clinical co-pilot. The system translates AI reasoning into structured clinical actions like ordering tests and prescriptions within hospital workflows. That integration matters more than raw diagnostic accuracy because it demonstrates AI’s ability to function within existing healthcare infrastructure.

The gap between where MIRA performs today and where it needs to be for live clinical deployment is substantial. Simulation using real cases isn’t the same as managing patients in real time. Physical examination findings, patient communication nuances, and the unpredictable chaos of emergency departments introduce variables that controlled testing can’t capture.

Future work needs to establish generalization across diverse patient populations, healthcare settings, and clinical contexts. The Dresden study used complete, clean datasets from a single institution. Deploying MIRA across multiple hospitals with different EHR systems, documentation practices, and patient demographics will reveal whether the system’s performance holds up.

Liability frameworks need to catch up to the technology. If an AI agent orders a test that causes patient harm or misses a diagnosis that a human physician would have caught, current regulatory structures don’t provide clear answers about responsibility. Hospitals considering AI-assisted clinical workflows need legal clarity before they can move forward.

Rentosertib’s Phase III trial will provide its own answers over the next year. The drug has to demonstrate safety and efficacy in 320 patients over 52 weeks. If it succeeds, it validates not just one molecule but an entire approach to drug discovery. If it fails, it raises questions about whether AI-designed drugs face different challenges than traditionally discovered compounds.

What’s clear is that AI has moved past the proof-of-concept stage in medicine. MIRA and rentosertib represent different applications of the same underlying capability: systems that can reason through complex problems, make decisions based on incomplete information, and generate solutions that perform at or above human levels. The regulatory, ethical, and practical challenges of deploying these systems at scale remain unsolved, but the technical foundation is demonstrably solid.

The next stage involves translating controlled success into real-world impact while building the oversight structures that make that translation safe. MIRA’s performance in simulation proves the concept. Now the work begins on everything that comes after proof.