Harvard Trial Finds AI Beats Doctors in Emergency Diagnosis and Treatment Planning

Context Emergency departments rely on rapid decisions made from limited data. Recent surveys show 16% of UK doctors use AI daily, yet concerns about error and liability persist. The Harvard study, published in *Science*, tested a large‑language model (OpenAI’s o1 reasoning) against practising clinicians using real patient records.

Key Facts - The trial enrolled 76 patients from a Boston hospital. Both the AI and a pair of doctors received identical electronic health records containing vitals, demographics and brief nurse notes. The AI identified the exact or near‑exact diagnosis in 67% of cases; doctors were correct in 50‑55% of cases. - When more detailed records were provided, the AI’s accuracy rose to 82%, while doctors reached 70‑79%; the gap was not statistically significant. - For longer‑term treatment planning, the AI was evaluated on five clinical scenarios alongside 46 doctors using standard resources such as web searches. The AI scored 89% for appropriate plans, compared with 34% for the physicians. - The AI correctly linked a pulmonary clot case to the patient’s lupus history, a connection missed by the doctors. - Lead author Arjun Manrai emphasized that the results do not imply AI will replace physicians, but signal a “profound change” in medical technology.

What It Means The findings demonstrate that AI can serve as a high‑precision second opinion, especially when clinicians must act on sparse textual data. In triage situations, the model’s superior diagnostic rate could reduce missed or delayed diagnoses, potentially improving patient outcomes. However, the study excluded non‑textual cues such as visual assessment and patient distress, limiting its scope to paper‑based decision making.

For practitioners, the practical takeaway is to view AI as an augmentative tool rather than a substitute. Integrating AI outputs into existing workflows may raise diagnostic confidence and streamline treatment planning, but clinicians must remain vigilant about AI errors and the current lack of formal accountability frameworks.

Looking ahead, regulators and hospitals will need to define liability standards while monitoring AI performance across diverse patient groups. The next phase of research will likely test AI alongside physical examinations and in real‑time emergency settings, revealing how the technology fits into the emerging “triadic care model” of doctor, patient, and machine.