AI Episode 4: Is AI un-useful?
By now you should have heard that AI can find tumours in lung images better than radiologists can. You should have also heard, on this Blog or elsewhere, that AI that can help clinicians make better decisions have been around for about 50 years with very low adoption. In this fourth episode of our series on AI we see if the low adoption of AI in medical practice means, one way or another, that AI has no room in the clinic.
Let's first look at the reasons given for low adoption of AI, and some arguments for an against these reasons:
High cost of adoption and change management, integration with conflicting factors and understanding of patient complexity by AI system
Lack of transparency of the reasons given for a decision can erode trust in the AI
Resistance to change by clinicians
Lack of patient input
AI is too slow to update (e.g. in the time it takes to implement new evidence or new policies may leave the AI out of date even if only temporarily)
AI is too quick to accept single points of evidence (e.g. history at one hospital) without considering external factors such as outbreaks, drug availability etc.
Lack of reproducibility, especially of machine learning
None of these reasons can't be remedied in an ideal world. The problem is that progress is somewhat slow, software engineers typically are not aware of the enormous complexities of clinical workflows and the evidence base behind them and the ways in which clinicians make decisions.
At the current levels of maturity, AI is not ready to be integral to clinical workflows. While some AI can help with some tasks where sufficient evidence exists (e.g. finding cancer nodules in CT scans), it is nowhere near being able to make clinical decision or even recommendations. AI can already present relevant evidence to clinicians, but not yet to act or recommend an action.
As AI becomes more connected, more sophisticated and more integral it will be trusted more and given more control. Trust in AI, as in anything else, needs to be earned over long periods of times, even decades.