The number of artificial intelligence-powered health tools on the market far outweighs the number of such tools actually adopted by hospitals. These tools' lack of transparency, sustainability and adaptability are key reasons for this disparity, according to a paper published Feb. 12 in The Lancet.
Many hospitals have an "algorithmic aversion" because their IT teams are often unable to view AI tools' input data and source code, according to the paper. There is also uncertainty about which datasets were used to train the tools' algorithms and whether they were tested using retrospective data, which may not coincide with clinical practice.
To help hospitals overcome their algorithmic aversion, the paper laid out three steps: provide transparency about the datasets used for AI tools' initial training, deconstruct neural networks so the features that drive an algorithm's performance are understandable for clinicians, and allow clinicians to retrain algorithms with local data.