AI Model Beats Doctors in Emergency Diagnosis Study
Harvard researchers found OpenAI's o1 model correctly diagnosed 67% of emergency cases versus roughly 52% for physicians, but experts say the technology is not yet ready for real-world clinical use.
May 3rd 2026 ยท Singapore
Singapore General Hospital is deploying artificial intelligence tools to address healthcare challenges in aging populations, including PENSIEVE-AI, a drawing-based diagnostic tool that achieved 93 percent accuracy in detecting cognitive impairment and dementia among nearly 1,800 seniors aged 65 and older. The tool requires less than five minutes to complete, costs significantly less than traditional testing methods, and could save up to SG$3,650 per senior while reducing clinician time by approximately 15 hours per patient. SGH has also partnered with Philips to establish the SGH and Philips MRI Training Center, inaugurated in November, which features AI-integrated workflow platforms that can reduce scan times by 15 to 50 percent and aims to strengthen MRI capabilities across Southeast Asia. Meanwhile, a new study published in Science by researchers from Harvard Medical School and Beth Israel Deaconess Medical Center found that OpenAI's o1 model outperformed human physicians in emergency room diagnoses. When presented with the same text-based information available in electronic medical records, the AI model provided the exact or very close diagnosis in 67 percent of triage cases, compared to 55 percent and 50 percent for two attending physicians. The researchers emphasized that this does not mean AI is ready to make real-life decisions, but rather indicates an urgent need for prospective trials to evaluate these technologies in real-world settings. Healthcare adoption of AI is accelerating despite implementation challenges. The American Medical Association reported that 66 percent of physicians were using AI in 2025, nearly doubling from 38 percent in 2023, with 57 percent reporting reduced burden through automation. However, experts caution that while AI pilots may demonstrate impressive results in controlled settings, they often fail in production environments when implementation does not align with operational requirements, underscoring the need for disciplined, operationally grounded deployments rather than novelty-focused initiatives.