A research project in the field of artificial intelligence (AI) and health care, led by a researcher from Instituto Superior Técnico Catarina Barata, was published in the journal Nature Medicine and reveals promising results related to the diagnosis of skin cancer.
Using the method tested in this study increased the rate of correct diagnoses made by dermatologists by 12%.
AI has demonstrated great accuracy in several areas of diagnostic imaging, leading to the development of decision support tools and creating possibilities for improving accessibility to health care.
The study published last Thursday focused on the use of an AI model called “reinforcement learning”, which was crossed with tables generated by experts, in which rewards and penalties were attributed for different diagnostic errors.
When compared to the traditional learning model, this reinforcement model produced notable improvements in sensitivity for two skin diseases: melanoma and basal cell carcinoma. Sensitivity for diagnosing melanoma rose from 61,4% to 79,5%, while for basal cell carcinoma it rose from 79,4% to 87,1%.
The team of researchers also discovered that this AI model reduces over-reliance on automated diagnosis while maintaining its accuracy. This data could be an important factor for change in the world of medical AI, since over-reliance on this type of diagnosis has generated concern. When tested on real dermatologists, this reinforcement learning model led to a 12% increase in AI accuracy.
Catarina Barata, a researcher at Técnico's Institute of Systems and Robotics, explains that this type of approach is increasingly essential. “The use of AI must be tailored to people and bring them benefit. This rewards model makes the process simpler and easier to understand for a non-expert, which is very much in line with the direction we want to take with regard to artificial intelligence.”
The researcher is also involved in the international project “Responsible AI”, which aims to implement a series of reflective practices and concerns with accessibility to this type of technology.
Comments