AI Outpaces Clinicians in Heart Scans, Diagnosis, and Parkinson's Screening — But Patients Aren't Convinced
AI outperforms doctors in cardiac prediction, complex diagnosis, and Parkinson's screening — while most Americans remain uneasy with the technology.
By Dr. Asher Knippel
This week delivered a cluster of landmark findings on AI in medicine: machines are beating doctors at reading hearts and diagnosing illness, a questionnaire tool may catch Parkinson's years early, and yet most Americans say they'd rather not have AI anywhere near their care.
AI reads heart scans 300× faster — and more accurately
A UCL-led study published in The Lancet Digital Health analyzed cardiac MRI data from 27,382 patients in the UK and United States and reached a striking conclusion: AI-derived measurements of ejection fraction — the share of blood pumped with each heartbeat — predicted five-year mortality significantly better than the same measurements made by expert clinicians.
Speed is part of the story: the AI completed each analysis in two to three seconds, against roughly 13 minutes for a specialist. But accuracy is the more important finding. The model flagged patients as high-risk even when their scores fell below the clinical threshold that would otherwise have cleared them for discharge. In practice, that means some patients who would have been sent home may actually need closer monitoring — and the AI is better at identifying who they are.
Microsoft's diagnostic AI beats generalist physicians two-to-one
A paper published in Science this week described Microsoft's Medical AI Diagnostic Orchestrator, MAI-DxO, achieving 83.2% accuracy on 304 complex cases drawn from New England Journal of Medicine case reports, compared with 36.1% for generalist physicians working with reference materials. On a separate emergency-department dataset the gap was similar: 80% for the AI versus 47% for physicians.
The system also ordered fewer diagnostic tests — and cheaper ones — to reach its conclusions. That combination of higher accuracy and lower cost is the scenario AI optimists have long promised, and it arrived with peer-reviewed backing. The harder question the paper raises is what this means for the training pipeline of generalist doctors: if a software system can outperform a physician who has access to reference materials, what does that imply for medical education and staffing?
Most Americans are still uncomfortable with AI-assisted care
Against that backdrop of technical progress, a Pew Research Center survey of 3,488 US adults released November 3 offers a pointed reminder that clinical performance and public acceptance are not the same thing. Sixty percent of respondents said they were uncomfortable with AI being used to help diagnose their conditions or recommend treatments. Only 5% said they were comfortable with AI making insurance claim decisions autonomously. Fifty-one percent expressed concern about the security of their personal health data.
The finding that stands out most: patients trust doctors most to use AI in their interests — more than hospitals, more than insurers, and more than the government. That matters for how AI tools get introduced in clinical settings. A technology that outperforms a physician in a controlled trial may still fail in deployment if patients disengage from care or refuse to act on AI-generated recommendations.
Biotechs push back on AI drug discovery hype
Not all players in health technology are rushing to embrace AI. A growing number of biotech startups are deliberately skipping AI-powered drug discovery platforms, opting instead for high-throughput wet-lab experimentation guided by PhD chemists rather than deep-learning models.
The concerns these companies cite are specific: training datasets that don't represent the full chemical and biological space of interest, neural networks whose decisions can't be interrogated or explained, and biological systems too complex for current architectures to model reliably. One profiled firm runs 50,000 compound tests per week using robotics — a scale that would be unthinkable without automation — but relies entirely on human scientists for hypothesis generation. The story is a corrective to the idea that AI adoption in drug development is uniform or inevitable.
A 30-question tool may catch Parkinson's years before symptoms appear
A team at Rensselaer Polytechnic Institute published a study in Nature Aging on November 4 describing Quest2Dx, a model that analyzes health questionnaire responses to identify people in the prodromal phase of Parkinson's disease — the period before motor symptoms appear.
The model looks for early warning signs that often precede a formal diagnosis by years or even decades: reduced sense of smell, disrupted REM sleep, and subtle changes in handwriting. A condensed 30-question version performed nearly as well as fuller questionnaire batteries, which matters enormously for real-world deployment. A short questionnaire can be administered in a primary care visit with minimal cost and no specialist equipment. If Quest2Dx is validated in prospective studies, it could give neurologists — and patients — a meaningful head start on a disease for which earlier intervention consistently produces better outcomes.
What connects these five stories is a gap that runs through all of them: AI systems are reaching and in some cases exceeding human performance in narrowly defined clinical tasks, but the path from benchmark to bedside remains longer and harder than any single study can resolve. Trust has to be built. Clinicians need to understand what the tools are actually doing. And the limits that biotech skeptics raise — black-box models, unrepresentative training data, biological complexity that defeats current architectures — are real even where the headline results look impressive. This week's research is evidence of rapid progress. It is not, yet, evidence that the transition is complete.