5 flaws hindering the efficacy of AI-powered COVID-19 diagnostic tools

Machine learning-powered diagnostic tools for COVID-19 are inaccurate ways to identify the disease because of biases, design flaws, inadequate datasets and lack of reproducibility, according to a study published March 15 in Nature Machine Intelligence.

Advertisement

Researchers conducted a systematic review of 62 studies detailing machine learning models designed to diagnose COVID-19 from chest radiographs and computed tomography images published between Jan. 1, 2020, and Oct. 3, 2020.

Five design flaws many of the tools examined in the study had:

  1. Training on datasets that were too small to be accurate
  2. Training on public datasets that had evolved and merged over time
  3. Training on datasets that used images from children for their non-COVID-19 images and images from adults for their COVID-19 data
  4. Training and testing on the same data
  5. Lack of input from radiologists and clinicians

More articles on artificial intelligence: 
Cleveland Clinic, NFL Players Association launch AI venture for neurological diseases
5 recent studies exploring AI in healthcare
AMA’s 7 tips for responsible AI use in healthcare

Advertisement

Next Up in Health IT

Advertisement

Comments are closed.