Oxford team’s technology picked up danger signs with 86% accuracy in study of 72,000 patients in England