The regulatory barriers to artificial intelligence in healthcare
How will these algorithms be certified? Will a model validated on American populations hold for European, Asian or African ones? The open questions.
Beyond the technical questions lies another obstacle: regulation. "We do not know what the certification procedures for this software will be," notes Alain Livartowski, data director at the Institut Curie in Paris.
Does an algorithm travel?
An algorithm validated on American populations — will it hold for European, Asian or African ones? The question is not academic. Disease prevalence, imaging practice and even equipment differ between regions, and a model trained on one distribution can quietly fail on another.
For African healthcare systems in particular, this matters: tools arriving from elsewhere were rarely validated on local populations.
The data problem
Training an algorithm to recognise medical images requires a considerable mass of health data. Assembling that data, with the consent and governance it requires, is the principal problem facing the companies developing these systems.